Forem: Arnon Shimoni

The agentic runtime is a crystal ball

Arnon Shimoni — Wed, 27 Aug 2025 11:57:49 +0000

Two weeks ago, I was on a customer call that was supposed to be about margin alerts which is a feature we’re building. Standard stuff. They wanted to know when their agent costs spiked.

Then their engineer shared their screen and said something that's been rattling around in my brain ever since:

"Look at this. Every time our agent handles a refund request, it pulls the customer's entire purchase history, checks seven different policies, and generates a legal audit trail. But when it suggests a product recommendation? Just a simple embedding search. The refund costs us $3.40. The upsell costs $0.03. We charge the same for both..."

Why we built an agent monetization platform

Originally, we built Paid.ai to solve a simple problem: AI agents are black holes for money. You need to know what things cost before you go bankrupt. Track costs, set limits, optimize margins. Basic unit economics.

But this customer had figured out something else entirely which really made a difference - by instrumenting their costs, they'd accidentally instrumented their entire business intelligence. The cost data was just a proxy for something much more interesting: operational complexity.

High cost operations = high value interactions.
Low cost operations = commodity features.

Their most expensive customers were their most loyal. Their cheapest customers were tourists who didn’t stick around.

This matters more than we thought
Paid sits at a unique point in the stack. Owing to Paid's agentic signals architecture, we see every token, every API call, every tool invocation. We built this to track costs, but what we're actually tracking is the entire lifecycle of value creation.

When your agent handles a customer request, we see:

How many steps it took
Which tools it used
How many retries and fallbacks
The total resource consumption

But that's just infrastructure.

What we're really seeing is:

Problem complexity
Solution sophistication
Value being created
Trust being built (or destroyed)

We thought we were building a billing system for AI agents. Turns out we're building the business intelligence layer for agents.

The pattern emerged

After that call, I started looking at our own data differently. Started asking different questions.

The patterns are everywhere once you look. I took the ~250 calls I had recorded and browsed them again (with Claude’s help - projects are a LIFESAVER).

Pattern 1: Complexity correlates to commitment

Users whose agents run complex, multi-step operations stick around.

Sure, the operations are complex, but those complex operations mean they've integrated deeply into actual workflows.

They're not testing. They're depending.

Pattern 2: Cost spikes are often a type of PMF

When customers complain about costs, that's good. It means they're getting value.

When they don't use enough to even notice costs? They're already gone, they just haven't canceled yet.

Pattern 3: The Margin story can be a distraction

True, everyone's freaking out about negative margins. We do too.

But margins only matter if you're pricing on costs.

If you price on value, and you can see the value in the operations data, margins become a solved problem.

We’re sitting on a goldmine

This is the part where I'm supposed to be humble, but fuck it.

We're sitting on something special.

Every other tool in the agent stack sees only their slice:

Model providers see tokens in and out
Orchestration frameworks see flow control
Observability tools see errors and latency

We see the money. And money is the ultimate truth-teller. When someone routes a request through Paid, we don't just track what it costs. We see:

What they were willing to pay for
What complexity they were willing to tolerate
What value justified that cost

Yes, we built a cost tracker.

But costs get tracked through signals of operations.

And operations are just business logic made visible.

Did you know that you have this data too?

Every AI company is sitting on this data, but many don't know it or can’t access it.

They're so focused on the AI race - better models, lower latency, more features - that they're missing the business intelligence goldmine in their own logs.

Your agent's runtime is telling you:

Which features actually matter
Which customers will upgrade
Which use cases have product-market fit
What your business actually does (vs what you think it does)

But most teams are just looking at costs and trying to make them go down.

That's like having a gold mine and only caring about your electricity bill.

The agent economy needs more intelligence

We started Paid.ai because the agentic economy needed basic financial infrastructure. Track costs, manage margins, don't go broke.

But what we discovered is that the cost layer is actually the intelligence layer.

Every dollar spent is a decision made.

Every API call is a value proposition.

Every operation is a business process made visible.

We built a cost tracker. It turned into a crystal ball.

And that customer who started this whole revelation? They've cut costs by 60% while increasing prices by 3x. Not because they optimized their infrastructure. Because they finally understood what their customers actually valued.

They're not paying for AI. They're paying for outcomes. The operations data shows you exactly which outcomes matter.

That's the real insight: In the agentic economy, your costs aren't your problem. They're your roadmap.

You need AI cost tracking, and you should start now

Arnon Shimoni — Mon, 11 Aug 2025 11:58:40 +0000

"We basically look at our monthly OpenAI bill and apply a distribution factor."

Sound familiar? You're not alone.

I analyzed hundreds of our customer conversations about AI agent monetization. The second most common pain point after pricing? Cost tracking. Or more accurately, the complete lack of it.

Way too many people have no idea what their AI is racking up in costs, and you see it on Reddit, X, and in communities everywhere.

According to CloudZero, the average monthly AI spend is jumping from $62,964 to $85,521 in 2025 - a 36% increase - but I actually think they’re underestimating it.

Why it gets harder at 10 customers

Imagine yourself in this person’s shoes who told us they track globally:

"We have separate API keys per customer so that we can track usage per customer. But it doesn't give you the dollar figure. It gives you tokens or usage. Then we just basically look at our monthly bill at OpenAI and apply a distribution factor."

This works when you have 5 customers. Maybe even 10. But here's what happens at scale:

At 5 customers: You're checking OpenAI dashboard daily. Life is good.
At 10 customers: You've got separate API keys. Maybe a spreadsheet. Still manageable.
At 50 customers: Welcome to the Spreadsheet Spiral of Death.

By the time you figure out a customer is unprofitable, they've already burned through months of margin.

The API key problems: What OpenAI and Anthropic won’t tell you about usage

These foundation model providers’s usage dashboard is built for developers, not businesses.

Can you figure out what was actually performed here or what the value delivered was in this OpenAI dashboard?

One founder told me:

"There are all these vendor tools, right? And certain discounts based off bundles. Pricing is very fluid, and so that cost may change. But right now it seems like it's a static view."

What you get:

Token counts by API key
Monthly aggregate spend
Basic usage graphs

Anthropic’s billing/costs dashboard is even more basic than OpenAI’s.

What you really need:

Cost per customer, per action, per outcome
Real-time margin alerts
Workflow-level profitability
Which features are margin killers

The gap between what you need and what you get? That's where companies die.

🛠️ Building a simple cost attribution

Before you spin up a massive data warehouse project - get practical. You need cost visibility TODAY, not in 3 months.

Here's the progression that works:

Level 1: Just ship it

Start tracking these 5 signals immediately - you can pull this off in an hour!

In Python it’ll look something like:

# Basic cost tracking - implement this in the next hour
def track_ai_operation(customer_id, operation_type, model, tokens):
    cost_map = {
        'gpt-4o': 0.03 / 1000,
        'gpt-4o-mini': 0.002 / 1000,
        'claude-4-sonnet': 0.003 / 1000,
        'embeddings': 0.0001 / 1000
    }

    cost = tokens * cost_map.get(model, 0)

    # Log to CSV, database, whatever - just START
    log_entry = {
        'timestamp': datetime.now(),
        'customer_id': customer_id,
        'operation': operation_type,
        'model': model,
        'tokens': tokens,
        'cost': cost
    }

    return log_entry

Pro tip from Paid: Even a CSV beats nothing. Perfect is the enemy of shipped.

Level 2: Attribution magic

You wrap every AI call:

In Javascript, it’ll look something like:

const trackCost = async (customerId, agentId, operation) => {
    const startTime = Date.now();
    const result = await operation();

    const costData = {
        customerId,
        agentId,
        operation: operation.name,
        model: result.model,
        inputTokens: result.usage.prompt_tokens,
        outputTokens: result.usage.completion_tokens,
        totalCost: calculateCost(result.usage, result.model),
        duration: Date.now() - startTime
    };

    await recordCost(costData);

    return result;
};

Level 3: Get a purpose-built system

"… this is crushing us. Our cost has gone up, and it's driving our margins down."

We’ve seen customers go from 70% margins to 40% because ONE customer was "very chatty" with their voice agent.

This is where Paid’s AI cost tracking and margin management can help.

Some vendors like Paid.ai provide real-time visibility into AI agent costs by automatically tracking spend, profit margins, and usage across providers like OpenAI and Anthropic using SDKs and OpenTelemetry. It enables pricing accuracy and scalable profitability with alerts for cost spikes, per-agent margin insights, and unified dashboards for vendor expenses.

Don’t build this yourself - let someone else handle the details and stay up-to-date.

One of our customers had negative margins…

A founder walked into our office (figuratively - we met online) and told us “we're bleeding money and I don't really know why."

They have a clean UI (not shadcdn!), happy customers, $299/month pricing that looked profitable on paper. They were celebrating hitting 100 customers when their CTO dropped a "we're losing money on every single call"…

We helped dig in to their voice bills where most customers had totally normal usage - but then we found a really chatty customer.

Average customer: 2,000 voice tokens/month ($8 in costs)
Chatty customer: 187,000 voice tokens/month ($748 in costs)
The $299/month became -$449/month

The fix was to introduce complexity factors. More complex workflows cost more, and use the higher-value LLMs. Once a certain threshold has been reached, the models were downgraded.

💡 Your path forward

Every day that you operate without cost visibility is a day you're potentially losing money. But you don't need a perfect system to start.

Today: Implement basic signal tracking. Even a CSV is better than nothing.

In a few weeks: Add customer-level attribution. Know who's killing your margins.

In a couple of months: Graduate to real-time monitoring. Catch problems before they compound.

Stop guessing based on your vendor bills - start tracking, start managing, start profiting.

Your margins will thank you.

AI Billing Showdown: 6 Billing Platforms for AI Agents

Arnon Shimoni — Mon, 11 Aug 2025 09:12:19 +0000

Everyone says they're doing AI billing - but who's actually got the features to back it up?

Our data shows that roughly 75% of AI companies we spoke to struggle with billing their agents. The problem isn't just technical complexity. It's that legacy billing forces AI companies into Frankenstein solutions, stitching together usage tracking, margin monitoring, and outcome measurement across multiple platforms.

Customer-facing AI products need agent-native billing. Infrastructure tools need usage-based flexibility. The platforms that understand this distinction are winning.

The Agent-First Winner: Paid

Built for: AI agent companies charging per outcome, per workflow, or per "digital employee"

Paid is the only billing platform designed from the ground up for AI agents. While others retrofit subscription models with usage add-ons, Paid treats agents as the fundamental billing unit.

Paid’s command center lets you evaluate performance, ROI, human value equivalents, costs, and revenue of agents across different departments with ease.

Paid’s AI-focus and research identified four working models for AI monetization:

FTE replacement (price per agent as digital employee)
Consumption (price per action/token with margins)
Process automation (price per completed workflow)
Outcome and results-based (price per delivered outcome)

The standout feature is the signal-based architecture. Instead of forcing you to translate agent activities into "billing units," Paid tracks whatever signals matter to your business from meeting bookings, successful resolutions, completed analyses.

Integration is simple. Five lines of code to start tracking agent usage. The platform automatically handles margin monitoring (tracking what each agent workflow costs you) and generates customer-facing ROI reports that make renewals easier.

Vibe-code customer dashboards and deploy them within minutes, showing your agents’ work and value.

Developer experience: SDKs and integration to common agent frameworks. Built by engineers for engineers.

Paid’s agent cost tracking tracks cost performance across different agent types, customers, and even specific orders and integrates with common agent frameworks using OTEL.

Best for: AI products where agents replace human work or deliver specific outcomes.

Pricing: Free for agent cost tracking. Currently in beta for billing features.

Metronome

Built for: High-scale usage-based billing, especially AI infrastructure.

Metronome powers billing for OpenAI’s metered plans, and Databricks. Their platform processes billions of usage events daily with Apache Kafka handling the streaming architecture.

Metronome's system is entirely architected around processing data in real-time, which matters when you're billing for millions of API call, token, or GPU-minute at massive scale.

What makes them different: They started with the most complex enterprise customers first, then scaled down. When OpenAI needed to change pricing, it previously took 6-8 weeks.

Developer experience: Robust APIs, extensive documentation, sandbox testing. Built by engineers for engineers.

Limitation: Not designed for AI - but for metering. You'll need to define your own usage metrics and build customer-facing value reporting.

Pricing: Custom, likely $10K+ annually for meaningful scale.

Orb

Built for: Product-led growth companies with mixed subscription + usage models

Orb has SQL-based metric definitions let engineering teams create custom usage tracking while giving product teams a modern UI to experiment with pricing.

Standout features include prepaid credits ledger for predictable customer budgeting, threshold billing to prevent usage abuse and real-time revenue reporting tied to product metrics.

Limitation: Like Metronome, Orb isn't AI-native. It's a flexible usage platform that works well for AI companies but doesn't have built-in agent workflow tracking or margin management.

Best for: AI companies with strong engineering teams who want to own their billing logic while getting infrastructure handled.

Pricing: Starts around $749/month, scales with usage volume.

Chargebee

Built for: Traditional SaaS companies adding AI features to existing subscription models

Chargebee has recently pivoted toward "Better Billing" for the AI era, partnering with companies like DeepL and Zapier. Their approach combines subscription management with usage-based add-ons.

What they do well: Mature platform with extensive integrations, solid dunning management, global tax compliance. If you're adding AI features to an existing SaaS product, Chargebee can handle hybrid pricing.

Limitation: Still fundamentally subscription-first. Complex AI usage patterns require workarounds. No native agent workflow tracking or outcome-based pricing models.

Best for: Established SaaS companies adding AI capabilities, not AI-first products.

Pricing: Starts free up to $250K billing, then 0.75% of revenue.

Stripe Billing

Built for: Startups with engineering resources who need reliable payment infrastructure

Stripe's biggest advantage is trust and scale. Their payments platform processes over $1.4 trillion annually with 99.999% uptime. Stripe Billing adds usage metering on top of this foundation.

For AI companies: Stripe supports metering per API call, token, or custom usage units. Their developer documentation is excellent, and integration with the broader Stripe ecosystem is seamless.

Limitation: Quoting and enterprise workflows require significant custom development. Complex usage models need engineering work. No built-in margin tracking or AI-specific reporting.

Best for: Early-stage AI companies with traditional subscription models plus some usage components.

Pricing: ~0.5-0.8% of invoiced volume plus Stripe payment fees - can go up to 2.9%

Togai (Zuora)

Built for: Enterprise companies with complex consumption models needing sophisticated rating engines

Togai handles up to 1 billion+ events per day and provides rating for tiered, inclusive allowances, and overage fees. Recently acquired by Zuora, it's positioned as the metering component of enterprise billing workflows.

Technical strength: Low-code pricing model builder, real-time usage ingestion, revenue simulation for forecasting.

Limitation: Togai bundles with Zuora Billing (or another invoicing system) to complete billing workflows. Enterprise complexity and pricing.

Best for: Large enterprises with existing Zuora implementations.

Pricing: Enterprise licensing, typically $100K+ annually combined with Zuora.

Choosing Your Next Billing Platform

If you're building customer-facing AI agents: Paid.ai is purpose-built for your use case. Agent workflows, outcome tracking, and margin management come standard.

If you're focused specifically on compute proxy: Metronome or Orb provide the scale you need around metering. Choose Metronome for enterprise complexity, Orb for faster implementation.

If you're enterprise with complex consumption: Togai + Zuora handles the most sophisticated rating scenarios but requires significant implementation.

AI monetization is still evolving rapidly. Pick a platform that can adapt as your understanding of value delivery changes.

What to Look For in AI Billing Platforms

Let's talk about what actually matters in AI billing. After working with dozens of AI companies wrestling with billing complexity, these are the features that separate the winners from the "we'll figure it out later" crowd.

Agent-Native Tracking

Traditional platforms track "users" or "API calls." AI platforms need to track agent behaviors, workflows, and outcomes. Look for systems that can capture:

Multi-step agent workflows as single billable units
Cross-agent collaboration (when Agent A hands off to Agent B)
Outcome completion vs. process initiation
Variable cost attribution per agent action

Real-Time Usage Processing

AI workloads spike unpredictably. Your billing platform needs to handle:

Millions of events per day without dropping data
Real-time cost visibility (customers hate surprise bills)
Threshold alerts before usage explodes
Event deduplication (agents sometimes retry)

Margin Visibility & Cost Attribution

This is where most platforms fail spectacularly. AI companies need to know:

What each agent workflow actually costs to run
Model switching impact on margins (GPT-4 vs Claude vs local models)
Token efficiency tracking across different use cases
Break-even points for different pricing tiers

Pricing Model Flexibility

AI monetization is evolving fast. Your platform should support:

Hybrid models (base subscription + outcome fees)
Dynamic pricing based on model performance
Prepaid credits with different expiration rules
Volume tiers that actually make sense for AI usage patterns

Pro tip: Avoid platforms that make you choose between "subscription" or "usage-based". Most successful AI companies use hybrid models.

The 5 stages of SaaS death

Arnon Shimoni — Mon, 11 Aug 2025 08:46:20 +0000

Last week, I met a founder who'd just lost a pretty big $2M deal to a 3-person startup with a half-finished product (I call these "3 Stanford grads in a trenchcoat").

He was annoyed they didn't even have a basic dashboard or settings page.

I think he wasn't angry about losing the deal, but mostly because he worked on the wrong thing.

I think most of us in product have been there, realizing we put engineering teams through the wringer for something that someone else did in a weekend.

The pattern I can't unsee

As part of the work I do at Paid, I've spoken to maybe 20 SaaS founders about AI, and lots of agent builders. Obviously these are VERY different companies, in different stages, and sometimes different industries.

With SaaS - it's like watching everyone go through the same cycle, but for business models.

Let me show you what I mean. Every SaaS executive I know is somewhere on this curve today, with the variable being how quickly they're going through it.

Stage 1: Denial

"Our customers need human oversight"

This is where it starts. Usually in a board meeting. Someone asks about AI strategy, and you hear yourself saying things like:

"Our customers value the human touch"
"AI can't handle our level of complexity"
"We're monitoring the space"
"It's mostly hype"

Three months ago, I was advising a company that runs sales-training software. Solid business, $10M ARR, growing 35% annually. The CEO told me their moat was "design expertise that AI couldn't replicate."

Last week, their biggest customer started building their own training content with Claude Opus 4.1 with projects... They didn't need design expertise for that...

Here's how you know you're in denial: You use terms like "AI washing", you think ChatGPT is just a toy and everything built on top of it is "just a wrapper".

You also believe your industry is somehow special, somehow immune. You're waiting for the hype to die down. While you're in denial, your customers aren't.

I got some data from a couple of SaaS companies last month - on average, most of their daily active users were copy-pasting data into ChatGPT, then pasting it back because they didn't like their AI interface..

Yeah, NOW your product has become expensive middleware for OpenAI.

Stage 2: Anger

"They don't even have SSO!"

You lose a deal to a young player - 3 Stanford grads in a trenchcoat who are just out of YC.

This is when the anger kicks in...

I heard of a founder complaining that an AI competitor was "unsustainable", they only had a dozen features versus his hundreds. They didn't have compliance certifications. They were undercutting by charging 90% less.

He was right about everything except what mattered.

The anger phase sounds like:

"When their VC money runs out..."
"Wait until they hit enterprise requirements..."
"Our features took years to build..."
"Customers don't understand what they're giving up..."

Another CIO told me last week:

"These AI companies are cheating. They're not playing by the rules - they aren't making sure their stuff is compliant."

Yeah. It sucks that they're not playing your game at all. You're bigger and optimizing for software excellence and reducing risk. They're optimizing for outcome delivery.

Stage 3: Bargaining

"We'll add AI features to our enterprise tier"

This is where 80% of you are right now. You've accepted AI is real, but you think you can negotiate with it.

Lots of us have been there. The frantic product meetings driven by the board mandate on AI. The emergency AI taskforce (or "tiger team"). The Google Slides titled "AI Strategy v5_final"... It's literally "how to not die" in slides form.

I've seen this all the way back at the end of 2022:

November: "We're adding AI-powered insights!"
December: "We're building a copilot!"
January: "We're experimenting with usage-based pricing!"
February: "We're exploring a hybrid model!"
March: "We're refocusing on our core value prop!"

I know because I've helped three companies go through this exact progression. Same presentations, same pricing experiments, same surprise when their AWS bill exceeded their MRR.

You know you're bargaining when:

Your pricing page looks like a math textbook with lots of levers and options

You're running three different pricing models simultaneously

You use "Now with AI ✨"

You've added "credits" to your billing system

Your gross margins went negative but you call it "investment" or "cost of marketing"

Now you're just about paying customers to use your product.

Stage 4: Depression

"Maybe we should focus on enterprise"

Now your best engineers are updating their LinkedIn profiles. You start using phrases like "strategic pivot" and "return to fundamentals" and all sorts of McKinsey speak...

Even if your product was genuinely excellent, with a beautiful UI, thoughtful workflows, delightful to use - it just isn't solving the right problem anymore. It helped humans work better, when customers wanted to eliminate more of the work.

The depression markers:

You stop shipping features

Every conversation ends with "what about enterprise?"

You're considering acquisitions (buying or being bought - both...)

Your investors start sending "helpful" articles

You dream about the good old days of 2020 (right???)

Now you're actually ready to innovate. This is where real innovation happens.

When you stop trying to save your product and start trying to solve your customer's problem.

Stage 5: Acceptance

"What would we build if we started today?"

The companies that make it here have realized something pretty significant.

You can't bolt AI onto SaaS. You have to rebuild around outcomes.

Sounds like startup suicide? Here's what will happen if you go down this path:

Imagine a project management software that retails for $89/seat/month.

Now transition to selling completed projects for $2,500 each. Customers don't log in. They don't see dashboards with burndowns - instead they just get results.

Revenue is up 3x. Costs are down 60%. Customer satisfaction is the highest it's ever been.

You've reached acceptance when:

You stop defending your features

You price based on outcomes, not access

Your competitive set includes companies you've never heard of

You're building things that would have seemed insane two years ago

You're either terrified or exhilarated (both is a good marker)

These stages aren't always sequential. I've seen company leaders go from denial to acceptance in a single customer call. I've watched companies bargain for nearly a year. Some never leave anger...

The longer you stay in each stage, the more expensive it gets. Denial costs you market position. Anger costs you talent. Bargaining costs you capital. Depression costs you momentum.

Acceptance costs you your ego, but that's a price worth paying.

A few weeks ago, I was talking to a founder who'd just hit acceptance:

"We kept asking 'How do we compete with AI?' until we realized we were asking the wrong question. The right question was 'What would an AI-native solution look like?' Once we asked that, everything became clear."

His company now looks nothing like it did a year ago. They sunsetted tons of features. They reduced their codebase and replaced their entire pricing model.

They're growing faster than ever.

Where we're going

I used to think AI would make SaaS better. More efficient, more powerful, more valuable.

I was wrong.

AI and agentic AI doesn't improve SaaS. It replaces the need for it. Not the functionality—the entire concept of software as an intermediary between intent and outcome.

The question used to be: "How can we help humans do this better?" The question now is: "Why do humans need to do this at all?"

If you're in SaaS - you're probably in Stage 3 right now, bargaining with reality. That's okay. Most of us will go through it - you're not alone.

But don't stay there too long. Finish your cycle and start building again. Once you accept what's happening, building becomes fun again.

You'll stop protecting the past and start inventing the future again.

Even if that future doesn't include your current product.

Sorry.

If you're ready to make the move from SaaS to AI native pricing and business - give us a shout at paid.ai

How AI Agent Companies Go From $0 to Profitable

Arnon Shimoni — Wed, 18 Jun 2025 09:30:37 +0000

I've spoken to about 160 agent builders in the last half year, and my estimate says 75% of them are leaving money on the table. They tell me so, and they tell me they're worried about it.

Last week, on a discovery call with a company building a support agent for automotive use, the founder told me "Our agent handles 50,000 customer support tickets monthly, but we're losing $3 on every interaction".

What's missing in my opinion is a bit of education about what you can and can't do for agentic monetization.

So after helping 40+ AI agent companies implement profitable pricing models and having spoken to at least 100 more, I'm going to share the pattern that I see: You've built something incredible, but you're struggling to turn it into sustainable revenue.

The problem: Normal SaaS pricing doesn't work for AI agents

This one is hard to unpack but let me explain:

LLM costs are unpredictable: One customer saw costs spike from $500 to $8,000 in a single month due to INPUT token lengths changing
Value varies wildly: The same agent might save one customer $100/month and another $10,000/month
Usage patterns are non-linear: Unlike SaaS, AI agent usage doesn't scale with team size. There is no seat to price for!

Because we're all used to SaaS, tons of founders default to per-seat pricing because it's familiar. But when one user can deploy 50 agents doing the work of an entire department, seat-based pricing leaves lots to be desired...

The 4 AI agent pricing models that actually generate profit

After extensive testing with real customers, these four models consistently outperform traditional pricing:

1. Agent-based pricing: The FTE replacement model ($2K-$20K/month)

When it works: The agent replaces a specific job function or part of it
Real example: One of our legal tech customers charges $8,000/month for an AI contract reviewer that replaces a $120,000/year paralegal.

The math

Human paralegal: $10,000/month fully loaded
AI agent: $8,000/month all-in
Customer saves: $2,000/month
Your margin: ~65% after infrastructure costs

My take: You can position against headcount budgets, not software budgets. HR budgets are 10x larger than IT budgets and it may work in your favour.

What to track today: Start documenting how many human hours your agent replaces. You'll need this data to justify your pricing.

2. Action-based pricing: The consumption model ($0.10-$5.00/action)

When it works: Competing with BPOs or call centers
Real example: A voice AI company charges $0.12/minute for calls, undercutting call centers by 70%.

Actual breakdown from a customer support agent company we work with:

Voice infrastructure: $0.03/minute
LLM costs: $0.02/minute
Other APIs: $0.01/minute
Total cost: $0.06/minute
They charge: $0.12/minute
Margin: 50%

The problem with this model is it can be "crunched" with margin. When things get commoditized, you will race to the bottom on pricing.

What to track today: Calculate your true cost per action including ALL components: LLM, APIs, infrastructure, even that Google Maps geocoding service you may be using.

3. Workflow-based pricing: The process automation model ($50-$500/workflow)

When it works: Multi-step processes with clear deliverables
Real example: An SDR agent that charges:

Lead research: $2/lead
Email personalization: $1/email
Meeting booking: $8/meeting

Even a simple workflow could involve:

3 data enrichment API calls
2 LLM summarization steps
1 quality verification check resulting in a total cost of $0.47/lead

By bundling into workflows and charging $2/lead, one of our customers achieved 76% margins while still being cheaper than human SDRs.

My tip: Bundle workflows into packages. "500 leads/month for $750" converts better than "$1.50 per lead" even though it's the same price.

What to track today: Map out EVERY step in your agent's workflow. You'd be surprised how many hidden costs lurk in "simple" processes.

4. Outcome-based pricing: The results model ($500-$5,000/outcome)

This is my favourite and it's the holy grail, but it can be tricky to get to it!
When it works: Clear, measurable business results
Real example: A recruiting AI that charges:

$500 per qualified candidate
$1,000 per scheduled interview
$5,000 per accepted offer
$0 for everything else

As the agent success rate improves (say, from 10% to 25%), the revenue increases and the cost can remain rather flat.

Yes, you need rock-solid attribution. We have one customer that spent 3 months building attribution before switching to this model.

Attribution tracking with Langfuse (if you're already using it):

# Track outcome attribution in your observability
generation = langfuse.generation(
    name="interview_scheduled",
    trace_id=trace.id,
    metadata={
        "outcome_type": "interview_scheduled",
        "candidate_id": candidate_id,
        "attribution_confidence": 0.95,  # How sure are you?
        "billable": True,
        "amount": 1000
    }
)

Tracking is one thing. Actually billing for it, handling disputes, and managing different pricing tiers? That's where things get messy - and I suggest you look away from the usual suspects for handling it :)

What to track today: Start measuring success rates and outcomes NOW, even if you're not charging for them yet. This data is so important...

Recommended decision framework for agentic monetization models

Which model should you choose?
Here's my framework:

What budget are you targeting?

→ Headcount budget (10x larger)?
Use Agent-based pricing

→ Outsourcing/BPO budget?
Start with Action-based, plan your escape

→ ROI/Innovation budget?
Use Outcome-based if you can prove attribution

→ Operational efficiency budget?
Use Workflow-based pricing

What should you do today?

Margins will kill you if you don't know what they are. You can't price well if you don't know what your inputs are.

I recommend adding a cost calculation now! If you're using something like n8n, add a cost calculator node after each service:

// Cost aggregator
const costs = {
  openai: $('OpenAI').first().json.usage.total_tokens * 0.00003,
  whisper: $('Whisper').first().json.duration * 0.006,
  elevenlabs: $('ElevenLabs').first().json.characters * 0.00018,
  // Add all your services here
};

const totalCost = Object.values(costs).reduce((a, b) => a + b, 0);

// Log it somewhere (or don't, I'm not your boss)
console.log(`Workflow cost: ${totalCost.toFixed(4)}`);

If you're using Langfuse, you can track costs per trace:

# Add cost calculation to your traces
from langfuse import Langfuse

langfuse = Langfuse()

# When creating a trace
trace = langfuse.trace(
    name="customer_support_workflow",
    metadata={
        "customer_id": customer_id,
        "token_count": completion.usage.total_tokens,
        "estimated_cost": completion.usage.total_tokens * 0.00003
    }
)

The above will work for understanding costs, but they won't help you with billing, margin analysis, or scaling.

What you should do next week

After you've started tracking your costs, I highly recommend you talk to 10 customers about value. Don't ask "What would you pay?", instead understand:

What were they doing before?
How much time/money did that cost?
What specific outcomes matter to them?

After that, you can start to figure out which models (as I outlined above) could match.
Test the water with a few friendly customers (simulate it!), and see if one of them leaves you more profitable.

I'm very happy to discuss: What's your biggest pricing challenge?
I'm curious about your experience. Are you currently:

Still figuring out your costs?
Struggling to communicate value?
Worried about customer pushback?
Already profitable and scaling?

Drop a comment below. I respond to every question and often write follow-up posts based on common challenges.