Forem: Shishir Mishra

Why AI Projects Fail — 7 Patterns We See Repeatedly | KORIX

Shishir Mishra — Sun, 10 May 2026 02:15:05 +0000

Originally published at korixinc.com.

Why do AI projects fail? 87% of AI projects never reach production because of five recurring mistakes: unclear business objectives, poor data quality, no governance framework, wrong team structure, and scaling too fast. That number — from Gartner’s ongoing research into enterprise AI adoption — hasn’t improved much since 2020. If you’re planning an AI investment, understanding these failure modes is the single most important step you can take to protect your budget and timeline.

I’ve spent 19 years building software systems, and the last several focused specifically on AI implementation. I’ve seen projects fail for all five of these reasons — including some I was brought in to rescue. Here’s what actually goes wrong, and more importantly, how to prevent each one.

AI Project Failure Breakdown87% Never Reach Production13%
Failed / shelved / never deployed
Reached production successfully87%of AI projects never reach production (Gartner)

Failure Reason	Frequency	Prevention
Unclear business objectives	Most common	Define one measurable outcome before starting
Poor data quality	Very common	Budget 40-60% of project time for data preparation
No governance framework	Common	Design governance as a constraint from day one
Wrong team structure	Common	Assign operational owner, not just a champion
Scaling too fast	Common	Start with a focused pilot targeting one process

Reason 1 — The Business Problem Was Never Actually Defined

This is the most common and most expensive mistake. A leadership team reads about AI, gets excited, and tells their team: “We need to implement AI.” That’s not a business objective. That’s a technology preference.

Without a specific, measurable outcome, the project drifts. Engineers build technically impressive demos that solve no real problem. Stakeholders keep changing direction because there was never a fixed target. Three months and £80,000 later, the pilot gets quietly shelved.

The vague version

“We want to use AI to improve our operations.”

The specific version

“We want to reduce document processing time from 4 hours per batch to 90 minutes, with 95% accuracy, within 8 weeks.”

The specific version gives you three things the vague one doesn’t: a measurable target (time reduction), a quality bar (95% accuracy), and a deadline (8 weeks). You can evaluate whether the AI project succeeded or failed. With the vague version, you can’t — which is why those projects quietly die without anyone formally calling them a failure.

How to fix it

Start with a specific process you want to improve — not a technology you want to adopt
Define success metrics before writing a single line of code
Get sign-off from whoever controls the budget on what “done” looks like
If you can’t articulate the business value in one sentence, you’re not ready

Reason 2 — The Data Wasn’t Ready

AI is only as good as the data it learns from. This isn’t a cliche — it’s a budget reality. Most companies dramatically underestimate how much effort goes into data preparation. Industry benchmarks consistently show that 40–60% of total AI project time is spent on data cleaning, normalisation, and pipeline engineering. Not model building. Not fine-tuning. Data plumbing.

Here’s what “data not ready” actually looks like in practice:

Inconsistent formats: Customer records in three different systems with three different schema conventions
Missing values: 30% of critical fields are blank, making the dataset unreliable for training
No historical depth: You need 18 months of data for a forecasting model, but only have 4 months of clean records
Access restrictions: The data exists, but legal or compliance constraints prevent its use for training
Bias in the source: The training data reflects historical patterns you actually want to correct, not replicate

I worked on a document processing project where the client assumed their PDF archives were ready for extraction. They weren’t. Half the documents were scanned images with no OCR layer, and the other half had inconsistent layouts across three years of template changes. We spent four weeks on data preparation before the AI component even started. That’s normal — but it wasn’t in the original timeline.

How to fix it

Run a data audit before commissioning the AI project — not during it
Budget 40–60% of project time for data preparation
Check data volume, quality, format consistency, and access rights upfront
If the data isn’t there, the honest answer is to fix the data problem first and build AI later

40–60%of AI project time spent on data preparation, not model building

Reason 3 — Governance Was Bolted On After the Fact

This is the failure mode I care about most, because it’s the one with the longest tail of damage. Teams build a working AI model, start getting results, and then someone from compliance asks: “Wait — can we actually use customer data this way?”

The answer is usually no. And retrofitting governance into a system that wasn’t designed for it is brutally expensive.

In regulated industries — financial services under FCA rules, healthcare under HIPAA, legal services handling privileged data — governance isn’t optional. It’s not something you add in the last sprint. Compliance requirements must be design constraints from day one. That means audit trails, data lineage, explainability, access controls, and retention policies are part of the architecture, not a layer on top of it.

KORIX’s position on this

Governance-first AI is our core thesis. Every system we build starts with compliance mapping and data governance rules before any model architecture is chosen. It’s slower upfront. It saves months of rework later. We’ve written more about this approach in our services overview.

How to fix it

Involve compliance and legal from the project kickoff — not the final review
Map regulatory requirements to technical constraints before choosing your approach
Design audit trails and explainability into the system architecture
Document data flows, access permissions, and retention rules as part of the technical spec
If your AI partner doesn’t ask about governance in the first meeting, that’s a red flag

Reason 4 — The Team Structure Was Wrong

A common pattern: the company hires a machine learning engineer (or contracts an AI vendor) and expects them to deliver a production system. The ML engineer builds a model with 94% accuracy in a Jupyter notebook. Six months later, it’s still in a Jupyter notebook because nobody planned for integration, deployment, monitoring, or end-user workflows.

Successful AI projects need three perspectives from the start:

Domain expertUnderstands the business process, edge cases, and what “good” looks likeWithout them: technically sound model that solves the wrong problem
AI/ML engineerBuilds the model, handles data pipelines, manages training and evaluationWithout them: no working system at all
Operations personPlans deployment, integration, monitoring, and user adoptionWithout them: a brilliant model that never leaves the lab

Role	What They Contribute	What Happens Without Them
Domain expert	Understands the business process, edge cases, and what "good" looks like	Technically sound model that solves the wrong problem
AI/ML engineer	Builds the model, handles data pipelines, manages training and evaluation	No working system at all
Operations person	Plans deployment, integration, monitoring, and user adoption	A brilliant model that never leaves the lab

Pure technology teams build technically impressive systems that nobody uses. Pure business teams write requirements that aren’t technically feasible. You need both perspectives collaborating from the beginning, not handing off between phases.

Honest note about KORIX

I’m a solo founder, which means I handle domain analysis, AI engineering, and deployment myself. This works well for focused projects where one person can hold the full context — document processing, lead scoring, workflow automation. It means I’m selective about what I take on. I won’t accept a project that genuinely needs a 10-person team, because I can’t be a 10-person team. That honesty saves both of us time and money.

How to fix it

Ensure domain, engineering, and operations perspectives are present from kickoff
Don’t isolate the AI work from the people who will actually use the output
Plan for deployment and adoption from day one — not as a Phase 2 afterthought
If your vendor or team can’t explain how the model gets into production, ask harder questions

Reason 5 — They Tried to Scale Before Proving Value

The boardroom version: “Let’s build an enterprise-wide AI platform.” Six months and £300,000 later, the platform exists but nobody uses it because it wasn’t validated against a real business need first.

The pattern that works is boring and sequential:

Pick one process in one department with one measurable outcome
Build a focused pilot — not a platform — to prove it works
Validate the ROI with real numbers, not projections
Then invest in scaling what you’ve proven

This is exactly why we designed the 21-day AI pilot: a contained engagement with a defined scope, a working proof of concept, and measurable results. If the pilot succeeds, you have evidence to justify a larger investment. If it doesn’t, you’ve spent weeks and a modest budget instead of months and a fortune.

The scaling trap

Enterprise AI platform projects have a failure rate above 90%. Focused pilots with a single measurable objective succeed roughly 60–70% of the time. The difference isn’t technology — it’s scope discipline.

How to fix it

Resist the urge to build a platform — build a pilot instead
Define the smallest possible scope that still proves business value
Set a hard time limit (21 days is a good boundary for most pilots)
Only scale what’s been validated with real-world data and real users

Not sure if your organisation is ready for AI?Take our 2-minute assessment and get a personalised readiness score.
Take the Assessment →

How to De-Risk Your AI Investment

If you’re evaluating an AI project — whether you build it in-house, hire an agency, or work with a specialist like KORIX — use this checklist to protect your investment:

AI investment checklist

Business outcome defined: Can you state the target improvement in one sentence?
Data audit complete: Do you know the volume, quality, and accessibility of your data?
Governance mapped: Are compliance requirements documented as technical constraints?
Team roles clear: Do you have domain, engineering, and operations covered?
Scope contained: Are you starting with one process, one department, one measurable outcome?
Budget realistic: Have you allocated 40–60% of time for data preparation?
Success criteria agreed: Does everyone involved know what “done” looks like?
Iteration planned: Is there budget for at least 2–3 rounds of refinement after the initial build?

If you can check all eight boxes, your project has a significantly higher chance of reaching production. If you can’t, the gap tells you exactly where to focus before committing budget.

The best AI investment you can make is the one that starts with honest preparation. The worst is the one that starts with excitement and skips the hard questions.

The Bottom Line

87% of AI projects fail — but the reasons are predictable and preventable.

Define a specific business outcome. Audit your data. Build governance in from day one. Get the right team structure. Start with a pilot, not a platform. These five steps won’t guarantee success, but they’ll put you in the 13% that reaches production.

FAQ

Common questions about

AI project risks.

Have a question not listed here?

Ask us directly →

What percentage of AI projects fail?

87% of AI projects never reach production, according to Gartner’s ongoing research. The primary reasons are unclear objectives, poor data quality, missing governance, wrong team structure, and premature scaling.

How much should I budget for data preparation?

Budget 40–60% of total AI project time for data cleaning, normalisation, and pipeline engineering. This is not overhead — it’s the foundation. A model trained on messy data produces messy results. Read more about AI cost breakdowns.

What is the best way to start an AI project?

Start with a focused pilot targeting one process, one department, one measurable outcome. Our 21-day AI pilot validates whether AI can solve your specific problem before committing to a full build.

Why is AI governance important?

Governance ensures compliance with regulations (FCA, HIPAA, GDPR), provides audit trails and explainability, and prevents costly retrofitting. It must be a design constraint from day one, not an afterthought. Learn about our governance-first approach.

Should I build an AI platform or start with a pilot?

Always start with a pilot. Enterprise AI platform projects have a failure rate above 90%. Focused pilots with a single measurable objective succeed roughly 60–70% of the time. The difference is scope discipline, not technology.

Top 8 AI Workflow Automation Tools Compared (2026)

Shishir Mishra — Sun, 10 May 2026 02:00:45 +0000

Originally published at korixinc.com.

The 8 best AI workflow automation tools in 2026 are Zapier, Make, Airtable, n8n, Relevance AI, Lindy, Beam, and KORIX BYOS. Each breaks at a different point — Zapier on cost, Make on complexity, n8n on governance, Lindy and Beam on audit trails. Pick on feature count and you'll regret it within 12 months. Pick on the cliff edge nearest you and you'll buy 18 months instead of six.

I've spent 19 years building software systems, the last several focused specifically on AI implementation for service businesses with 20-150 staff. In that time, I've inherited two retail analytics rebuilds, watched a marketing automation system send hundreds of badly-targeted emails because no one had wired in a rollback, and seen an operations team manually reconcile thousands of documents per month because their AI extraction tool produced data their reporting workflow couldn't ingest.

Every one of those failures started with the same decision: someone picked an automation tool based on a feature comparison page rather than its production failure modes. This article gives you the comparison the vendors don't write — where each tool actually breaks, what that costs you in real money, and when the right answer is to stop renting and start building.

The Capability Cliff: The One Pattern Every Tool Hides

Every AI workflow automation tool on the market in 2026 has a capability cliff. The cliff is the specific operational threshold past which the tool's design assumptions stop holding — the point where its pricing model becomes punitive, its governance gets thin, or its visual builder starts fighting your actual logic.

The cliff is also why MIT NANDA's 2025 State of AI Report puts the production-success rate for AI workflow projects at just 5% — a number that hasn't moved meaningfully in three years. Atlassian's 2026 State of Product survey backs this up from the operational side: 46% of teams name integration as the single biggest barrier to scaling AI automation. Both findings point at the same root cause — teams pick tools by feature comparison and discover the cliff six to twelve months later.

Vendor comparison pages obscure this on purpose. Listicles compare features ("8,000+ integrations vs 1,800+ integrations") that don't matter for production decisions. The former Chief Decision Scientist at Google, Cassie Kozyrkov, frames it bluntly: "The bottleneck is not the AI technology. The bottleneck is knowing which problem to give it." The same is true for AI workflow tooling — the bottleneck is matching the tool's design assumptions to your operational reality. The questions that actually predict whether a tool will survive your next 18 months are different:

What's the cost shape at 10× current volume? Linear, sub-linear, or step-function?
What happens when a connected SaaS changes its API? Maintained integration, or "it broke, file a ticket"?
Who owns rollback when an AI agent misfires at 2 AM? The platform, your ops team, or nobody?
Can a compliance officer trace exactly why an automation made a specific decision six months ago?

Each of those questions has a tool-specific answer. The chart that visualises this — the cost-versus-complexity scatter where each tool sits at a different cliff edge — is the diagnostic I draw on whiteboards with operations leaders before any tool decision. We'll walk through each tool's specific cliff next.

1. Zapier — Breaks at Cost (~5,000 monthly runs)

Best for: Non-technical teams running 5-15 simple workflows with broad SaaS connector needs.

The cliff: Zapier charges per task, where every step inside a Zap counts. A 10-step workflow run 10,000 times produces 100,000 tasks. At standard 2026 pricing, that lands around $208 per month — and the cost shape is purely linear, with no native deduplication or batching that would dampen growth. Cost becomes the dominant constraint somewhere between 5,000 and 25,000 monthly runs depending on workflow length.

What works: Ramp-up speed is unmatched. Their integration catalogue is the largest in the industry — over 8,000 pre-built connectors — and the maintenance is genuinely active. AI features (Agents, MCP support, AI automation steps) shipped in late 2025 are credible.

When to leave: When monthly task count crosses roughly 25,000, or when you have more than three workflows that each fire 5,000+ times per month. The math gets uncomfortable fast.

2. Make — Breaks at Complex Data Shapes (~50 workflows)

Best for: Visual-builder teams running mid-volume operations with non-trivial data transformations.

The cliff: Make becomes cheaper than Zapier at around 750 monthly operations and stays cheaper through mid-volumes. The break point isn't price — it's workflow complexity. Once your scenarios need transforms across nested objects, real conditional branching with multiple outcomes, or shared sub-workflows that other scenarios reuse, Make's visual canvas becomes the bottleneck. You spend more time fighting the layout than shipping logic.

What works: Operations-per-month pricing is more honest than Zapier's task model for medium workflows, and the visual builder is genuinely better than competitors for showing data flow at a glance.

When to leave: When you have more than 50 active scenarios or your workflows require real algorithmic logic. Past that point, you're building software in a visual UI — which is harder than just writing the software.

3. Airtable + AI — Breaks at Ops Overhead

Best for: Teams that already live in Airtable, automating around their existing data rather than across systems.

The cliff: Airtable's automations were designed for in-base workflows, and they're excellent at that. The cliff appears when you start crossing bases or coordinating with external SaaS. Triggers fire silently, error handling is bare-bones, and once you have more than 5-10 cross-base automations the dependency graph becomes harder to debug than the workflows themselves are useful.

What works: AI Cobuilder shipped in late 2025 is genuinely useful for in-base AI logic — summarising records, classifying entries, generating outputs from structured data. If your workflow lives entirely inside Airtable, this is competitive.

When to leave: When you have more than 20 active automations, when you need observability (which automation ran when, what it did, why), or when external SaaS integration becomes the dominant pattern.

4. n8n (Self-Hosted) — Breaks at Governance

Best for: Technical teams running high-volume workflows with engineering ops capacity.

The cliff: n8n self-hosted is the cheapest option at scale, often by a 20× factor. The same 10-step workflow run 10,000 times costs around $208 on Zapier, around $20 on Make, and somewhere between $5-80 on a self-hosted VPS for n8n — depending on instance size. That cost differential makes n8n the obvious answer for high-volume use cases if you have engineering ops capacity.

What breaks: Governance. With self-hosted n8n, you own everything the SaaS platforms own for you — uptime monitoring, backups, security patching, audit trail design, role-based access control, model version pinning. n8n the product won't stop you from building all of that. But you have to build it. The 2025 Stack Overflow Developer Survey shows the trend clearly: 67% of developers handling automation infrastructure report self-hosting their workflow tools at scale, but the same survey notes that operational overhead is the leading reason teams revert to managed platforms within 18 months. Most teams that move to self-hosted n8n underestimate the ops overhead by a factor of two or three.

When to leave: When you're in a regulated industry that requires audit trails the platform doesn't provide out of the box, or when your engineering team can't fund the ongoing ops cost. The right answer here isn't to leave n8n — it's to add a governance layer your auditors will accept. Often that's where bespoke deployment overtakes the platform entirely.

5. Relevance AI — Breaks at Compliance + Scale

Best for: AI-native operations teams building agent-based workflows.

The cliff: Relevance is genuinely strong on agent orchestration. The break point is dual: compliance (SOC 2 in place; HIPAA, PCI, and GDPR audit-trail support thinner than enterprise buyers want) and pricing at scale (climbs sharply past 50 active agents). For a service business comfortably between SMB and enterprise, this can become an awkward middle.

What works: The agent-builder UX is excellent, and the platform handles tool calls and agent-to-agent handoffs better than most competitors. For prototyping and internal-use agents, Relevance ships faster than almost any other option.

When to leave: When you cross 50+ active agents or any compliance audit starts asking questions Relevance can't directly answer.

Top 8 AI Workflow Automation Tools Compared (2026) — at a glance.

6. Lindy.ai — Breaks at Audit Needs

Best for: Sales and customer success teams wanting fast-shipping AI agents.

The cliff: Beautiful product, weak governance. The audit trail isn't really there — you can see what an agent did, but not always why, and certainly not which model version was active when it did it. There's no real rollback semantics, and 18 months in you may not be able to answer the question every regulated industry asks: "show me the decision-trace for this output."

What works: Time-to-first-working-agent is among the fastest in the category. The UX is genuinely the best in the category for non-technical operators.

When to leave: First compliance review. If your industry has any external audit obligation — financial services, healthcare, regulated marketing — Lindy is a prototyping tool, not a production deployment target.

7. Beam AI — Breaks at Price

Best for: Fortune 500 enterprises with budget for white-glove deployment.

The cliff: Strong product, premium pricing model that's wrong for most service businesses. Custom enterprise contracts only — typical six-figure starting point with multi-year commitments. The product solves real enterprise problems but the contract structure prices out everyone below Fortune 500.

What works: If you're a regulated enterprise with a procurement department that can absorb a six-figure annual deployment, Beam ships excellent agent-native infrastructure with proper audit trails and compliance positioning.

When to leave: If you're not Fortune 500, you should leave on Day 1. The pricing isn't going to come down for you.

8. KORIX BYOS — No Cliff (Built Bespoke Into Your Stack)

The eighth option isn't a platform. It's the absence of one.

BYOS (Bring Your Own Stack) is what we deploy at KORIX: AI agents and workflows built directly into your existing software estate — HubSpot, Salesforce, Microsoft 365, Slack, Notion, whatever your team already uses — rather than on top of yet another SaaS platform. You own the code. There's no platform fee. No per-run cost after build. No vendor lock-in. No cliff.

What it costs: A KORIX 21-Day AI Pilot typically lands between $15,000-$40,000 for a single deployed agent or workflow with full ownership transfer. Compare that to five-year cumulative cost on a mid-volume Zapier or Make subscription, plus annual enterprise upgrade pressure, plus the implementation hours every platform demands anyway.

What you get: Code you own, sitting inside your existing stack. Agents auditable from day one because they were built with audit in mind. Production deployment by Day 22 — guaranteed, or you don't pay the second invoice.

When BYOS wins: See the threshold section below. It's not for everyone, and we're explicit about that. Read the full BYOS philosophy here.

The TCO Math (Real Numbers, 2026)

Cost-at-scale is the comparison the vendor pages won't show you. Here's the table for a typical 10-step workflow at three usage tiers:

Tool	100 runs/mo	1,000 runs/mo	10,000 runs/mo	Cost shape
Zapier	Free tier	~$30	~$208	Linear (per task)
Make	Free tier	~$9	~$20	Sub-linear at volume
n8n (Cloud)	~$20	~$20	~$50	Per execution (not per step)
n8n (self-hosted)	~$5 (VPS)	~$5-20	~$20-80	Infrastructure only + ops time
Airtable + AI	$24/seat	$54/seat	Custom	Per-seat + AI credits
Relevance AI	~$19	~$199	Custom enterprise	Step-function past mid-tier
Lindy.ai	$39+	$199+	Custom	Per-credit + tier upgrades
Beam AI	Custom	Custom	Custom (six-figure)	Enterprise contract only
KORIX BYOS	$15,000-40,000 one-time build, then $0 platform cost — you own the code	Capex, not opex

Pricing as of May 2026; verify on each vendor's site before committing. The numbers above assume a 10-step workflow at the indicated monthly run count.

The pattern: at low volume, Zapier and Make's free tiers are essentially free. At medium volume, Make and Cloud n8n become the value picks. At high volume, self-hosted n8n's TCO is unbeatable if you have ops capacity — but that's a real "if." Past 10,000 runs and 50 workflows, the cumulative five-year cost for any platform starts approaching the build cost of bespoke deployment. That's where BYOS becomes the rational choice.

The 7-Question Buyer Test (Ask Before You Sign)

Before committing to any AI workflow tool, walk through these seven questions. They're the diagnostic I run with operations leaders before any tool decision. Answer them honestly — the right tool falls out of the answers.

What's your monthly run volume? Below 500: any free tier works. 500-5,000: Make is the sweet spot. 5,000-50,000: n8n (cloud or self-hosted, depending on ops capacity). 50,000+: bespoke or enterprise.
What connects to what? Pure-SaaS chains favour Zapier. Database-centric work favours Airtable or Make. Multi-system orchestration with custom logic favours n8n or bespoke. Inventory which integrations you actually need — not the catalogue size.
Is the data sensitive? If your workflow touches PHI, financial records, or any regulated data, the tools without audit trails (Lindy, basic Make) are out from Day 1. Don't paper over this.
Who owns the workflow when it breaks? If the answer is "the platform," you need SaaS support contracts. If "our ops team," you need observability tools. If "nobody," you have a problem regardless of which tool you pick.
How fast must it run? Real-time (sub-second response): bespoke or specialised. Near-real-time (under 60 seconds): most platforms. Batch (hourly/daily): the cheapest option wins.
What's the total cost at your projected scale? Calculate at 12 months and 36 months — not just the trial-month price. The cost shape (linear, sub-linear, step-function) matters more than the starting price.
Who owns the workflow code? If you stop paying the platform, do your automations come with you or evaporate? This is the question that gets ignored most often. The answer changes the entire conversation.

If your answers cluster around a single tool — pick that tool. If they're scattered across multiple tools, you're either at the cliff edge of one tool or you're a candidate for bespoke deployment. Which brings us to the threshold question.

When to Stop Shopping for Tools and Build Your Own

There's a specific threshold past which the math flips: licensing a workflow platform costs more over five years than building bespoke automation directly into your existing stack. The threshold has three triggers, and crossing any one of them is enough.

Trigger 1 — Volume: 5,000+ monthly workflow runs. At this volume, platform fees become a meaningful operational line item. The five-year cumulative cost of a mid-tier Zapier or Make subscription approaches the one-time build cost of bespoke deployment.

Trigger 2 — Workflow count: 50+ active workflows. Past this point, the platform's governance can't keep up — every compliance review surfaces the same gaps, and the cost of bolting governance onto a platform that wasn't designed for it exceeds the cost of building governance-first from the start.

Trigger 3 — Compliance: Any regulated audit. SOC 2, HIPAA, PCI-DSS, GDPR with audit-trail requirements — the moment a compliance officer enters the conversation, most no-code platforms become risk vectors rather than risk reducers. Governed AI isn't an extra layer you add to a platform; it's a design choice you make on Day 1, and bespoke deployment lets you make it.

BCG's 2026 AI Value Capture research on enterprise AI adoption found that organisations capturing the most value from AI are not the ones with the most sophisticated models — they are the ones with the clearest process definitions and the tightest feedback loops between AI output and human review. Translated to workflow tooling: the firms that win past the threshold aren't the ones with the most platform features, they're the ones who designed governance into the system from the start. Platforms make that hard. Bespoke makes it the default.

A different angle from Deloitte's 2026 State of Generative AI in the Enterprise: the survey found organisations that piloted AI on a specific, bounded workflow before broader rollout had significantly higher rates of successful scaled adoption — and the firms that scaled successfully were also the firms most likely to have moved away from generic platforms toward purpose-built integration. Both data points reinforce the same recommendation: a 21-day bounded pilot is the lowest-risk way to validate both the use case and the tooling architecture.

Andrew Ng, who has shipped more production AI than almost anyone alive, has been making this point publicly for years: the gap between "AI works in a notebook" and "AI works in production" is roughly 100× the engineering effort, and almost all of that effort is glue code, governance, and integration — the exact things vendor demos hide. His framing applies directly to workflow automation tooling: the platforms that demo well are not always the platforms that deploy well, and the ones that deploy well in regulated industries are almost never the ones that demoed best.

The bespoke threshold rule of thumb

If you cross any one of {volume > 5K runs/month, workflow count > 50 active flows, compliance audit pending}, the rational decision is to stop renting a platform and build your own stack. If you cross two of the three, you're already past the threshold — you just haven't done the math yet.

Three Real Failure Modes I've Seen

Pattern recognition from 19 years of inheriting and rebuilding AI projects. The names are disguised; the lessons are exact.

Failure 1: The marketing automation that couldn't roll back

A B2B services company had wired their AI lead-scoring model directly into outbound email automation through one of the no-code platforms. The model had been live for six weeks when it misclassified a batch of prospects — over-scoring them as warm leads — and the automation fired hundreds of inappropriate outreach emails over the course of a few hours. The platform had no rollback semantics for the agent's decisions and no audit trail granular enough to identify which specific model version had produced which output. The team ended up manually shutting down the entire automation and apologising to a list of prospects they'd burned. The fix was rebuilding the system with proper governance — including pinned model versions and a mandatory review checkpoint before any external action — but that work cost roughly twice what designing it correctly from the start would have cost.

Lesson: Speed-to-deploy without rollback isn't speed. It's debt that compounds the moment something goes wrong.

Failure 2: The platform that demoed beautifully and broke in production

A retail analytics rebuild I inherited a couple of years ago. The previous vendor had spent eight months on an AI platform building prediction models that looked excellent in demos. In production, the predictions were useless — because the models had never been connected to the client's actual inventory system. The data pipeline didn't exist. The platform was perfectly capable of supporting the integration, but supporting it and shipping it are different things, and "we'll connect that in phase 2" had quietly turned into "phase 2 isn't funded." The rebuild cost the client roughly three times what the original project should have cost if it had been designed as a deployment from the start.

Lesson: If your AI tool produces a strategy deck, an optimised prediction model, and a beautiful dashboard — but doesn't ship a working integration into the system that drives actual business decisions — you bought a strategy deck.

Failure 3: The document extraction that scaled into chaos

An operations-heavy services client manually processed several thousand documents per month. Multiple staff. Frequent data-entry errors compounding into reporting inaccuracies. They had tried two different no-code AI extraction tools before we got involved. Both had handled the basic extraction fine — but neither could structure the output in a way that fed cleanly into the client's downstream reporting workflow without extensive manual reformatting. The "automation" was producing data their actual systems couldn't ingest, which meant staff still had to touch every extraction manually to clean and reroute it. Net time savings: marginal. The bespoke rebuild wired extraction directly into the reporting pipeline, with structured outputs designed to match the downstream schema. Manual processing time dropped meaningfully and accuracy improved on its own.

Lesson: Workflow automation isn't extraction quality. It's whether the extraction output flows into the rest of your operation without a human reformatting it. Tools that don't model your downstream systems produce work, not automation.

The honest recommendation

If you're at the start of an AI workflow journey and running fewer than 500 monthly runs across simple workflows, start with Zapier or Make. Don't overthink it. The capability cliff is far away, the platforms work as advertised, and you'll learn what your real workflow needs are by running them.

If you're at 500-5,000 monthly runs, Make is probably the value pick — particularly if your workflows have any non-trivial data shape. n8n cloud is the runner-up if you want a simpler pricing model.

If you're past 5,000 monthly runs, or you have more than 50 active workflows, or any regulated audit is on the horizon — stop shopping for tools and have the bespoke conversation. The cost flips, the ownership question gets easier, and the governance you actually need stops being a feature you have to negotiate. Read the BYOS philosophy for the longer treatment of why service businesses past that threshold systematically benefit from the bespoke path. Or see how the 21-Day Pilot works for the operational side.

For the cost-decision angle specifically, our breakdown of what custom AI actually costs in 2026 walks through the £10K-£250K+ ranges by build complexity. For the failure-mode angle, five reasons most AI pilots fail covers the operational patterns this article touches on.

Either way: the right answer is rarely the platform with the most integrations or the cleanest UX. It's the one whose cliff you're least likely to hit. KORIX defines the capability cliff as the operational threshold at which a workflow tool's pricing model, governance design, or visual builder stops matching the team's actual production reality — and the test that matters is not which tool has the most features, but which tool's cliff is furthest from where you operate today.

The Bottom Line

Most workflow tools work fine until you hit the capability cliff. Choose on where each one breaks — not on feature lists.

Zapier breaks at cost around 5,000 runs/month. Make breaks at workflow complexity around 50 active flows. n8n self-hosted breaks at governance the moment a compliance officer enters the room. Lindy and Beam break on audit trails. The right answer for a 20-150 staff service business depends on which cliff you're closest to — and at a certain scale, the answer is to stop shopping for platforms and build your own bespoke stack instead. The TCO and ownership math both flip past that threshold.

FAQ

Common questions about

Operations.

Have a question not listed here?

Ask us directly →What's the cheapest AI workflow automation tool at scale?

Self-hosted n8n is the cheapest option once you cross roughly 5,000 workflow runs per month. It charges per execution (one workflow run regardless of step count), versus Zapier (per task — every step) and Make (per operation). At 10,000 runs of a 10-step workflow, Zapier costs around $208/month, Make around $20/month, and self-hosted n8n is whatever your VPS costs ($5-80/month). The catch: with self-hosted n8n you own all the operational responsibility — uptime, backups, security patches, governance — that the SaaS platforms handle for you. Cheapest by spreadsheet isn't always cheapest by total cost.

Is Zapier or Make better for non-technical teams?

Zapier wins on ramp-up speed and integration breadth (8,000+ pre-built connectors). Make wins on complex data transformations and visual workflow design — its scenario builder handles nested data structures and branching logic better. For a non-technical team running 5-15 simple two-step workflows, Zapier ships faster. For 15+ workflows with conditional logic or data reshaping, Make produces cleaner architecture. The crossover point is roughly 750 monthly operations: above that, Make becomes cheaper; below that, Zapier's UX is worth its premium.

When should I build my own AI workflow stack instead of using a platform?

Three triggers: (1) you cross 5,000 monthly workflow runs and platform fees become a meaningful line item; (2) you hit 50+ active workflows and the platform's governance can't keep up — every compliance review surfaces the same gaps; (3) a regulated audit (SOC 2, HIPAA, PCI, GDPR with audit trail) starts asking questions the platform can't answer. Past any one of those thresholds, the math flips: build cost becomes lower than five-year cumulative platform cost, and you stop renting access to your own automations.

Are AI agent platforms like Lindy and Beam ready for production?

For prototyping and internal-use agents — yes, both are excellent and ship fast. For customer-facing or regulated production — not quite. Lindy has elegant UX but lacks production-grade audit trails, model version pinning, and rollback semantics. Beam is more enterprise-ready but priced for Fortune 500 contracts (typically six-figure starting point). For a service business between those two extremes, the gap is real: you need governance Lindy doesn't ship and pricing Beam doesn't offer. That's the gap bespoke deployment fills.

What's the biggest mistake teams make when picking an AI workflow tool?

Picking on integration count instead of operational ownership. Every platform claims thousands of integrations — Zapier 8,000+, Make 1,800+, n8n 600+ native nodes. The number is meaningless. The questions that matter: which integrations does the platform actively maintain (not just list)? What's the failure rate when a connected SaaS changes its API? Who owns rollback when an automation misfires? Who can audit why an AI agent made a specific decision? Every team I've seen regret a tool choice in the last five years made the same mistake — they picked on the marketing comparison page, not the production failure modes.

AI Pricing Models: Per-Seat vs Per-Use vs Outcome (2026)

Shishir Mishra — Sun, 10 May 2026 02:00:43 +0000

Originally published at korixinc.com.

AI pricing in 2026 splits into six models: per-seat, per-token (usage), per-ticket, per-resolution (outcome-based), hybrid (base + overage), and bespoke (capex). Per-seat is collapsing — 21% to 15% of SaaS in 12 months. Hybrid is the new industry standard at 41% adoption. The cheapest model at trial is rarely the cheapest at 12-month or 36-month scale.

I've spent 19 years building software systems, the last several focused on AI implementation for service businesses with 20-150 staff. The single most expensive mistake buyers make in 2026 is choosing an AI tool on the trial-month price rather than the 36-month total cost of ownership. The pricing model determines the cost shape — linear, sub-linear, step-function, or capex — and the cost shape determines whether you're saving money or quietly hemorrhaging it.

This article walks through the six pricing models, where each one wins, what to ask vendors before signing, and how to calculate the all-in cost at scale. The data backing each section: Bessemer Venture Partners' 2026 AI Pricing Playbook tracks the shift across 200+ AI vendors and reports that hybrid pricing rose from 27% to 41% adoption in 12 months while pure per-seat fell from 21% to 15%. BCG's 2026 AI Value Capture research backs the alignment conclusion: vendors that align pricing to outcome capture more total revenue at higher buyer satisfaction than vendors that anchor on seats or tokens alone. MIT NANDA's 2025 State of AI Report separately found that only 5% of AI projects reach production — and pricing-model mismatch is one of the most common reasons projects stall before deployment, because the buyer's CFO never sees a path from "trial cost" to "scale cost" they can defend at the board.

The six models below are how the 2026 AI vendor market is structured. Each has a specific cost shape, lock-in dynamic, and incentive alignment. Get this wrong and the AI line item compounds quietly into a number nobody planned for.

The Six Pricing Models

Every AI vendor in 2026 picks one of six pricing approaches. The model determines the cost shape at scale, the lock-in dynamics, and the alignment between vendor incentives and buyer outcomes.

1. Per-Seat — $50-200/seat/month

How it works: Flat monthly fee per human user with access to the AI. Common for IDE-style developer assistants, design tools, and CRM-embedded AI.

Cost shape at scale: Linear with team size. Adding a 50th seat costs the same as adding a 5th seat.

When it wins: Internal tools where one human is genuinely the bottleneck. Developer assistants (GitHub Copilot, Cursor) and design tools (Figma AI) are the clearest fit. Per-seat is also right for products where usage is highly correlated with seat count anyway.

When it breaks: When AI agents do work that would otherwise require multiple humans. If one seat with an AI assistant handles 3× the workload of one seat without, per-seat pricing leaves 67% of the value on the table — and forward-thinking vendors are abandoning the model.

Trend: Pure per-seat fell from 21% to 15% of SaaS companies between 2025 and 2026, per Bessemer's tracking. Most vendors are moving to hybrid (per-seat plus usage overage) rather than killing per-seat entirely.

2. Per-Token (Usage-Based) — $0.0001 to $0.10 per 1,000 tokens

How it works: The buyer pays for raw computational input — every word the AI reads or writes generates a per-token charge. Standard for foundation-model APIs (OpenAI, Anthropic, Google) and API-first AI infrastructure products.

Cost shape at scale: Linear with usage. Predictable but punishing at high volume.

When it wins: API-level products where the buyer is building their own application on top. Predictable for prototyping and low-volume use cases.

When it breaks: At high volume, per-token costs can dominate the AI line item. A customer support agent handling 100,000 conversations per month at GPT-4 pricing can land in the five-figure range monthly without optimisation.

3. Per-Ticket — $0.30-1.00 per inbound conversation

How it works: Fixed fee per inbound conversation regardless of resolution outcome. Common in customer support and chatbot deployments.

Cost shape at scale: Linear with conversation volume.

When it wins: When conversation volume is predictable and resolution rates are uniformly high. Buyer can budget cleanly because cost = volume × rate.

When it breaks: When the AI's resolution rate is variable. Buyer pays for every conversation including the ones the AI fails to handle, which is the worst alignment of the six models.

4. Per-Resolution (Outcome-Based) — $0.50-2.00 per resolved conversation

How it works: Vendor charges only when the AI delivers a specific business outcome — a resolved customer conversation, a qualified lead, a closed ticket. Failed attempts cost nothing.

Cost shape at scale: Linear with outcomes, but only with successful outcomes — sub-linear if the buyer drives resolution rate up over time.

Real numbers (May 2026): Intercom charges $0.99 per resolved conversation. HubSpot's Customer Agent dropped to $0.50 per resolved conversation in April 2026 (from $1.00). Salesforce's Agentforce uses similar outcome-based pricing for specific workflows.

When it wins: For customer-support, sales-qualification, and ticket-resolution use cases where the outcome is clearly defined and the resolution rate is the variable that matters. This is the most aligned model — vendor only gets paid when the AI actually solves the buyer's problem.

When it breaks: When the definition of "resolution" varies between vendors or can be gamed. Specify resolution criteria contractually before signing — including what happens for "false positive" resolutions where the AI claims success but the customer follows up.

Andrew Ng has been making the broader argument behind outcome-based pricing for years: AI value is best measured against business metrics that matter, not against engineering metrics that don't. Pricing models that anchor to outcomes force both vendor and buyer to keep that focus.

5. Hybrid (Base + Usage Overage) — variable

How it works: A base subscription fee covers a defined volume of usage; overage fees apply beyond the included tier. Now the dominant industry-standard model in 2026.

Cost shape at scale: Step function (constant within tier, jumps at tier boundary, then linear in overage).

Trend: Hybrid pricing rose from 27% to 41% of AI vendors between 2025 and 2026, per Bessemer. The reason: hybrid gives vendors a stable revenue floor while letting them capture upside as buyer usage scales.

When it wins: When buyer volume is somewhat predictable but with growth upside. The base covers committed usage; overage fees scale only when value-delivery scales.

When it breaks: When overage rates are punitively higher than included-tier rates. Some vendors charge 2-3× more per unit on overage than within the included tier — a hidden cost that compounds at growth.

6. Bespoke / Capex One-Time — fixed-scope build

How it works: The buyer pays a one-time build fee for a deployed AI system, with full code ownership transferred. No recurring platform fee, no per-seat, no per-token. Maintenance and feature additions are separately scoped.

Cost shape at scale: Capex (one-time). After the build, marginal cost is the underlying compute (LLM API calls, infrastructure) only.

Real numbers: A KORIX 21-Day AI Pilot typically lands between $15,000-$40,000 for a single deployed agent or workflow with full ownership transfer. Compare that to the five-year cumulative cost of any subscription model at meaningful volume.

When it wins: At scale (5,000+ workflow runs/month or 50+ active workflows), or for service businesses past the size threshold where licensing economics flip. Our breakdown of the capability cliff covers this in detail. Also wins for any regulated environment where the buyer needs to own every line of code.

When it breaks: For very-low-volume use cases where a SaaS subscription would cost less than the build fee for the first 12-18 months. BYOS deployment is the right answer past a specific threshold, not on Day 1 for a single small workflow.

Cost Shape: The Comparison That Matters

The trial-month price is misleading. The 12-month and 36-month total cost at scale is what determines whether you're saving money or hemorrhaging it. The table below shows projected cost for a representative customer-support workflow handling 5,000 conversations/month, scaling to 25,000/month over three years.

Model	Year 1 cost	Year 3 annual	3-yr total	Cost shape
Per-Seat (15 seats)	$18,000	$22,500	~$60,000	Linear (team size)
Per-Ticket ($0.50)	$30,000	$150,000	~$280,000	Linear (volume)
Per-Resolution ($0.99 @ 70%)	$41,580	$207,900	~$390,000	Linear with successful outcomes
Hybrid (base + overage)	$24,000	$72,000	~$170,000	Step function
Bespoke (KORIX BYOS )	$30,000 + ~$5K LLM	~$25K LLM only	~$95,000	Capex + marginal compute

Pricing as of May 2026; figures are illustrative based on stated vendor list pricing and BYOS one-time build cost. Actual costs vary by negotiated terms and underlying infrastructure choices.

The pattern: at low volume in Year 1, per-seat looks cheapest. By Year 3 at 25,000 conversations/month, bespoke is the dominant cost-winner. Cassie Kozyrkov, former Chief Decision Scientist at Google, frames this kind of decision bluntly: "The bottleneck is not the AI technology. The bottleneck is knowing which problem to give it." Pricing-model selection is just a specific case — pick on the right problem framing (cost at projected scale, not at trial), and the right model becomes obvious.

The Five Questions to Ask Every AI Vendor on Pricing

Before committing to any AI tool, walk every prospective vendor through these five questions. Their answers — or refusal to answer — predict whether you're being sold a sustainable solution or a one-year honeymoon followed by a cost surprise.

What's the all-in cost at 10× current volume? If the answer is "let's discuss enterprise pricing", you're getting a vague number that protects vendor optionality at your expense. Push for a written quote at projected Year-3 volume.
What's the overage rate? For hybrid and tiered models, the per-unit cost of going over the included tier is often 2-3× the in-tier rate. This is where surprise bills come from.
How is "outcome" defined for outcome-based pricing? Get the resolution criteria in writing. If the vendor's definition of "resolved conversation" includes cases where the customer follows up, you're paying for false positives.
What happens to my data and workflows if I leave? Lock-in cost is hidden until you try to leave. Ask: can I export my workflows? My data? My configuration? The full cost of vendor lock-in compounds over 3-5 years.
Can you show me a comparable buyer's actual 36-month TCO? Not a marketing case study. The actual quarterly invoice trajectory for a similar buyer over three years. Vendors who can produce this are confident in their pricing model. Vendors who can't are hoping you don't ask.

According to Atlassian's 2026 State of Product survey, 46% of operations teams cite integration as the single biggest barrier to scaling AI automation — but pricing surprise comes second. Buyers who calculate TCO at projected scale before signing avoid the surprise.

AI Pricing Models: Per-Seat vs Per-Use vs Outcome (2026) — at a glance.

Three Pricing Failures I've Seen

Pattern recognition from inheriting AI projects. Names disguised; numbers exact.

Failure 1: The per-token surprise

A B2B services firm built a customer-support AI on a foundation-model API at per-token pricing. Trial month cost: $300. Month 12 cost: $14,000 — driven by a feature launch that increased conversation volume 40× in two weeks. The team had no cost cap, no usage alerts, no fallback to a smaller model for low-stakes queries. The fix was a six-week re-architecture with model-tier routing (cheap model for FAQ, GPT-4 for complex cases) and per-conversation cost caps. The right design from Day 1 would have been hybrid pricing with a per-conversation budget rather than raw per-token.

Lesson: Per-token pricing is fine for prototyping. For production, wrap it in a per-conversation budget with tier-routing — or move to a hybrid plan that caps your downside.

Failure 2: The per-seat trap

A 60-person firm bought a per-seat AI assistant at $80/seat/month for 60 seats, projecting "everyone uses AI". Year-end usage data: 12 of 60 seats accounted for 95% of the value. The other 48 seats were near-dormant — sometimes used once a quarter. The firm was paying $46K/year for usage worth $9K/year if priced honestly. The fix was renegotiating to a tiered hybrid plan: $30/seat base + per-conversation overage for power users.

Lesson: Per-seat pricing assumes uniform usage. AI tools rarely have uniform usage. Track actual usage for 60 days before committing to per-seat at scale; renegotiate when usage is concentrated.

Failure 3: The bespoke project that should have been SaaS

A mid-market firm built a fully bespoke API service for a workflow that any of three SaaS platforms would have handled. Year 1 build cost: $45K. Year 1 ongoing infrastructure + maintenance: $20K. The same workflow on a hybrid SaaS plan would have cost roughly $9K/year. Two years later, the firm migrated to the SaaS option and ran the workflow at one-third the cost.

Lesson: Bespoke is the right answer when no platform fits or when scale flips the economics. It's the wrong answer when a platform would fit but the team has internal resistance to "using a tool". Validate against existing platforms before committing to bespoke.

Why Pricing Models Are Shifting in 2026 — and What It Means for Buyers

The 26-percentage-point swing from per-seat to hybrid in 12 months isn't random. Three forces are driving it.

Force 1 — Per-seat under-prices AI value. When one human with an AI assistant produces three humans' worth of output, the vendor charging per-seat is leaving roughly 67% of the buyer's value on the table. Forward-thinking vendors restructured to capture more of that upside, which they can do legitimately by sharing the productivity gain. Bessemer's research tracks the model shift across the 200 fastest-growing AI vendors — and the leaders are almost uniformly on hybrid or outcome-based pricing, not per-seat.

Force 2 — Outcome-based pricing increases buyer trust. When a vendor says "I only get paid when the AI actually solves your problem," the buyer's purchasing risk drops to near zero. Deloitte's 2026 State of Generative AI in the Enterprise found 72% of enterprise AI projects exceed their original budget by at least 30%, and outcome-based pricing is one of the few defenses against that overrun pattern — because the buyer literally pays nothing for the failure modes.

Force 3 — Hybrid is the rational compromise. Pure outcome-based pricing creates revenue volatility for vendors that VCs don't fund well. Hybrid (base + overage) gives vendors a stable revenue floor and gives buyers a predictable bill, with both sides participating in the upside as usage scales. That's why 41% of AI vendors landed there in 2026 — it satisfies both incentive structures.

For service businesses with 20-150 staff, the practical implication is simple: be skeptical of any AI vendor still pushing pure per-seat in 2026 as their only option. The vendors who haven't moved off per-seat by now either don't have the data to back the shift or are protecting old revenue at the expense of new value-alignment. Either way, they're not the vendors you want for your next three years of AI deployment.

The Honest Recommendation

For your first AI deployment at small volume, hybrid pricing on a major platform (Zapier, Make, Intercom, HubSpot) is almost always the right starting point. Predictable base cost, scales gracefully, no commitment beyond the monthly plan.

For customer-support and ticket-resolution use cases at moderate volume, per-resolution outcome-based pricing aligns vendor and buyer incentives best. Get the resolution definition in writing before signing.

For internal-tools use cases (developer assistants, design tools) where one human is genuinely the bottleneck, per-seat continues to work — but only if usage is uniform across seats.

At scale (5,000+ workflow runs/month or 50+ active workflows), or for any regulated environment, bespoke deployment via KORIX BYOS typically wins on 36-month TCO and on the ownership/governance dimensions that matter most. The 21-Day Pilot is the structured engagement we recommend for validating the use case and the deployment model in one bounded scope.

For the broader question of which AI workflow tool fits which use case at which scale, our breakdown of the 8 best AI workflow tools and where each one breaks covers the platform comparison side. For "what does AI implementation actually cost?", see our full cost breakdown.

KORIX defines AI pricing as the long-tail cost shape, not the short-tail trial price — a tool that costs $50/month at trial and $5,000/month at scale is more expensive than a tool that costs $200/month at trial and $300/month at scale. Most buyers in 2026 still optimise on the trial price and discover the scale cost months after lock-in. Don't.

The Bottom Line

Six AI pricing models exist in 2026. The cheapest at trial is rarely the cheapest at scale — pick on cost shape, not on starting price.

Per-seat pricing is collapsing (21% → 15% of SaaS in 12 months). Hybrid (base + usage overage) is the new industry standard at 41% adoption. Per-resolution outcome-based pricing aligns vendor and buyer incentives — Intercom charges $0.99 per resolved conversation; HubSpot dropped to $0.50 in April 2026. The cheapest model at the trial-month price is rarely the cheapest at 12-month or 36-month scale. Pick on the cost shape (linear, sub-linear, step-function) at projected volume — not the headline number on the pricing page.

FAQ

Common questions about

Costs & Pricing.

Have a question not listed here?

Ask us directly →What is the most common AI pricing model in 2026?

Hybrid pricing — a base subscription plus usage overage — is now the industry standard, adopted by 41% of AI vendors per Bessemer Venture Partners' 2026 AI Pricing Playbook (up from 27% in 2025). Pure per-seat pricing has fallen from 21% to 15% in the same period because it no longer reflects the value AI delivers when one seat can handle 10× the volume of work. Pure usage-based pricing (per token, per call) is most common for API products. Outcome-based pricing (per resolution) is rising fast for customer-support and sales agents.

How does outcome-based AI pricing work?

Outcome-based pricing charges only when the AI delivers a specific business outcome — typically a resolved customer conversation, a qualified lead, or a closed ticket. Intercom charges $0.99 per resolved conversation; HubSpot dropped its Customer Agent pricing to $0.50 per resolved conversation in April 2026. The buyer pays nothing for AI attempts that fail or get escalated to a human. This aligns vendor and buyer incentives because the vendor only gets paid when the AI actually solves the problem. The risk for buyers is that 'resolution' definitions vary between vendors and can be gamed if not contractually specified.

Is per-seat AI pricing dying?

Pure per-seat pricing is shrinking but not dying. Per-seat fell from 21% to 15% of SaaS companies in 12 months according to industry tracking. The reason: AI agents make seat-based pricing absurd. If one human seat with an AI assistant handles the workload of three humans, the vendor charging per-seat is leaving 67% of the value on the table. Hybrid pricing (base seat + usage overage) is replacing pure per-seat. For internal tools and IDE-style products where one human is genuinely the bottleneck (developer assistants, design tools), per-seat continues to make sense.

What's the difference between per-token and per-resolution pricing?

Per-token pricing charges for the raw computational input: every word the AI reads or writes generates a per-token charge (typically $0.0001 to $0.01 per 1,000 tokens depending on model). Per-resolution pricing charges only when the AI completes a specific business outcome. Per-token works for API-level products where the buyer integrates the AI into their own application. Per-resolution works for end-user products where the AI is the application. Per-token is cheaper at low volume; per-resolution is cheaper if the AI's resolution rate is high and the buyer would otherwise pay for many failed attempts.

How do I calculate true AI cost across pricing models?

Compare on 12-month and 36-month total cost at projected scale, not on the trial-month price. The five inputs you need: (1) volume — how many users, conversations, tokens, or resolutions per month at year 1, year 2, year 3? (2) cost per unit at each tier; (3) overage rates if volume exceeds plan; (4) any hidden costs (data egress, integration fees, professional services); (5) lock-in implications (multi-year contract discounts vs flexibility cost). The headline price on the pricing page rarely matches the all-in 36-month TCO. For bespoke deployment models like KORIX BYOS, the calculation is simpler — capex one-time build, no recurring platform fee — which is why bespoke wins at scale.

Why I Refuse to Sell AI Platforms to My Clients

Shishir Mishra — Mon, 13 Apr 2026 13:19:05 +0000

And what I do instead — a philosophy called BYOS

By Shishir Mishra, Founder & System Architect (AI) at KORIX

I have been shipping software for 19 years. AI systems, web platforms, mobile apps — across fintech, healthcare, renewable energy, and SaaS. 150+ projects. 24 countries. And in the last two years, I have watched the AI consulting industry develop a pattern that I think is fundamentally broken.

Here is the pattern:

A business decides it wants AI.

A consulting firm sells them a platform.

The business pays seat licenses, training fees, and a 6-month implementation timeline.

The platform requires change management, data migration, and a dedicated internal team to manage it.

By month 6, the project is over budget. By month 12, the renewal arrives and the business realises it cannot leave.

The "AI transformation" quietly stalls.

This is not a failure of AI. It is a failure of the delivery model.

The 74% problem

Gartner published a number in 2024 that should have been a wake-up call: 74% of enterprise AI pilots never reach production. Three out of four.

When I dug into why, the reasons were depressingly consistent:

Platform adoption cost. Training 200 people on a new tool costs more than the tool itself. Change management kills more AI projects than bad algorithms.

Data migration risk. Moving customer data into a vendor's system creates compliance exposure that cannot be undone. Every GDPR officer I have worked with has flagged this as the #1 risk they lose sleep over.

Vendor lock-in. The business does not own the AI. It rents it. When the renewal comes, the vendor knows the business cannot leave — so the price goes up. Every year. Forever.

None of these problems are about the AI. They are about the business model wrapped around the AI.

The question nobody was asking

In 2025, I started asking a different question:

What if the AI came to the software, instead of the software coming to the AI?

Every business already runs on software. Salesforce for sales. Microsoft 365 for collaboration. SAP for operations. HubSpot for marketing. Slack for communication. Or some combination of custom-built systems that have been running since 2004 and somehow still keep the business alive.

What if, instead of selling a new platform on top of all that, we built the AI inside the systems the team already uses?

No new login. No new training. No data migration. No seat licenses. No vendor lock-in.

That question became a philosophy. I call it BYOS — Bring Your Own Software.

What BYOS looks like in practice

BYOS is not a framework, a library, or a product. It is a delivery model.

When a business comes to KORIX with an AI problem, we do not pitch a platform. We ask: what software does your team live in every day? Then we build the AI agent inside that software.

Here is a concrete example. A renewable energy company in the UK needed to score inbound leads faster. Their sales team was losing deals because quotes took 48 hours to turn around. A traditional approach would have been: buy a lead scoring platform, integrate it with their CRM, train the team, migrate historical data, and hope it works.

What we actually did: we built a lead scoring agent inside their existing Salesforce instance. The agent reads new leads as they arrive, enriches them from public data sources, scores them against the company's ideal customer profile, applies MCS compliance checks automatically, and routes the qualified ones to the right sales rep — with a plain-English explanation of why each lead scored the way it did.

Total deployment time: 21 days. Total new software purchased: zero. Total training required: zero — the sales team just sees better-qualified leads appearing in the CRM they already use every day.

The result: 3.2x conversion lift. Zero compliance violations. And the company owns the source code, the trained models, and the documentation. If KORIX disappeared tomorrow, the agent would keep running.

The delivery model: 21 days, fixed fee, you own everything

Every BYOS engagement at KORIX follows the same structure:

Days 1-3: Discovery. We pick the most painful workflow with a clear input, a clear output, and a measurable outcome. We scope the agent and write a one-page brief.

Days 4-18: Build. We build the agent inside the client's existing stack. Governance — confidence thresholds, audit trails, rollback policies — is built in from day one, not bolted on later. The client sees daily progress.

Days 19-21: Handover. We transfer the source code, train the internal team, and step back. On Day 22, the client's team operates the agent independently.

The fee is fixed — typically $15,000 to $40,000 depending on complexity. Payment is split 50/50: half at kick-off, half at handover. And here is the part that makes traditional consultancies uncomfortable: if the agent is not in production by Day 21, the client does not pay the second milestone.

No fine print. No "well actually." No sunk cost.

Why I ate my own dog food

Last week, I built seven AI agents for KORIX itself using the same BYOS approach. Not for a client — for my own business.

A GSC Health Agent that monitors Google Search Console daily and alerts on coverage drops.
A PageSpeed Agent that tests five key pages every morning for Core Web Vitals regressions.
A Sitemap Audit Agent that validates sitemap structure daily.
A Backup Verifier that confirms backups ran and escalates if any are stale.
An IndexNow Agent that pushes new content to Bing weekly.
An Auto-Index Agent that submits to Google on every publish.
A Lead Monitor that alerts me if any inbound lead sits unprocessed for more than two hours.

Total time: six days. Total new software purchased: zero. Each agent runs inside systems I already use.

If I can build this for myself in less than a week, imagine what I can build for a business with a real budget and a real problem in 21 days.

The objection I hear most

"But if you do not sell a platform, how do you make recurring revenue?"

I do not. And that is the point.

KORIX sells craftsmanship, not subscriptions. The client pays once for a working agent and owns everything afterward. If they want a second agent or a third agent later, that is a separate engagement. We would rather earn the next project than rent the first one.

This makes KORIX structurally different from most AI consulting firms. We have no incentive to make the engagement longer, more complex, or more dependent on us. Our incentive is the opposite: ship fast, hand over cleanly, and earn the next referral.

Who BYOS is for

BYOS is not for everyone. If you are building a consumer AI product from scratch, you probably need a platform. If you need a chatbot on your website, there are good off-the-shelf tools for that.

BYOS is for businesses that:

Have already invested in a software stack and do not want to rip it out.
Have been burned by a previous "AI transformation" that never reached production.
Operate in regulated industries where data residency and auditability are non-negotiable.
Want to own the AI, not rent it.

If you have ever said "we already have too many tools," BYOS was built for you.

The AI consulting industry will change

I believe the platform-first model has about 18-24 months left as the default. The economics do not work for the buyer, and buyers are getting smarter. The next generation of AI consulting will look more like BYOS: custom agents built inside existing systems, fixed-fee engagements, full ownership transfer.

KORIX is not the only company that will figure this out. But we are one of the first to name it, build a methodology around it, and ship it in production.

If you want to see the full philosophy, it is at korixinc.com/byos. If you want to see the agents we build, it is at korixinc.com/agents. And if you want to try it, book a free 30-minute Fit Check — we will tell you honestly whether an agent solves your problem, or whether it does not.

Shishir Mishra is the founder & System Architect(AI) at KORIX, a systems-first AI adoption agency based in Ahmedabad, India. He has spent 19 years shipping production software across 24 countries. Every KORIX pilot is led personally by Shishir, from kick-off to handover.