Most AI agent pricing is fake.

Not malicious fake. Just founder-delusion fake.

Someone builds a demo, imagines a clean workflow, picks a monthly number that feels professional, and sends a proposal.

Then reality shows up.

The client has three quiet days and one day where 240 weird edge cases hit at once. The handoff rules are messier than expected. The source data is filth. Humans override half the outputs in week one. The model cost is not the problem. The operational chaos is.

That is how builders end up with a “profitable” AI agent offer that quietly turns into unpaid support work.

So here is the rule:

if the workload is unpredictable, your pricing model needs to separate stable work from chaotic work.

Do not sell one flat number and hope the variance behaves.

Hope is not margin.

Why AI agent work is harder to price than normal automation#

Normal software pricing assumes one of two things:

  1. the usage pattern is reasonably stable
  2. the product is standardized enough that variability does not matter much

A lot of AI agent work has neither.

The variables that wreck pricing usually look like this:

  • inbound volume changes week to week
  • input quality is inconsistent
  • exception rate is unknown at the start
  • human review load changes with trust level
  • downstream tool failures create extra handling work
  • buyers change the rules once they see the system live

That means the real cost is not just tokens or API calls.

The real cost is a mix of:

  • setup and integration work
  • workflow design
  • monitoring and debugging
  • exception handling
  • human approvals
  • change requests disguised as “small tweaks”

If you price only the happy path, the unhappy path will eat you alive.

The three pricing mistakes that get people cooked#

1. Flat monthly unlimited pricing too early#

This is the classic self-own.

“We’ll run your lead triage agent for $1,500/month.”

Sounds simple. Buyers like simple.

But if you do this before you know the real workload, you are selling insurance without understanding the risk pool.

Maybe the client sends 120 clean leads a month. Maybe they send 1,400 chaotic form fills, forwarded emails, spam, partial records, and “can you just make it do this too?” requests.

Same price. Very different life.

2. Pricing off your internal cost only#

A lot of builders anchor on model cost because it feels measurable.

“This agent costs me $43 in API calls, so $500/month is fine.”

No.

Your margin is not determined by token cost. It is determined by the full support surface around the system.

Cheap inference with expensive exception handling is still an expensive service.

I already broke down what AI agents actually cost to run. The short version is that raw compute is only part of the picture. The business cost lives in operations.

3. Bundling exceptions into the base service#

This is where the silent loss happens.

If the agent does 80% of the work cleanly and the last 20% creates most of the effort, you do not have one service.

You have two services:

  1. high-confidence automated processing
  2. low-confidence exception handling

Those should not be priced the same way.

The model I would use instead#

Keep it boring. Boring pricing closes business and protects margin.

For unpredictable agent workloads, I would usually structure pricing in five layers:

  1. one-time setup fee
  2. base monthly platform fee
  3. included usage band
  4. overage pricing
  5. exception / human-review pricing

That gives you a clean way to price stability separately from chaos.

1. Charge a setup fee#

Do not hide setup inside month one.

Setup includes real work:

  • workflow mapping
  • integration and credential setup
  • prompt and rule design
  • validation logic
  • approval gates
  • dashboarding or notification wiring
  • test cases
  • launch support

That is not “free onboarding.” That is implementation.

Price it like implementation.

For small-business agent offers, setup might be a fixed fee in the low-to-mid four figures. For more complex systems, higher.

The exact number matters less than the principle:

the client should pay for the system becoming real.

2. Add a base monthly platform fee#

This is the charge for keeping the system alive and accountable.

It covers things like:

  • hosting and infrastructure
  • routine maintenance
  • monitoring
  • minor prompt/rule tuning
  • basic reporting
  • keeping the workflow from drifting into nonsense

This is not your volume charge. It is your “the machine exists and someone is responsible for it” fee.

Without this, every month starts to feel like a fresh negotiation.

3. Include a defined usage band#

This is where most people need more discipline.

Do not say “unlimited.” Say what is included.

Examples:

  • up to 300 inbound leads per month
  • up to 500 support tickets triaged per month
  • up to 100 call summaries processed per month
  • up to 1,000 documents classified per month

The point is not perfect precision. The point is creating a shared definition of normal.

If you do not define normal, every spike becomes your problem by default.

Usage bands make pricing legible to the buyer and survivable for you.

4. Add overage pricing before you need it#

This should be in the first proposal, not added later after you get hurt.

Overages can be priced:

  • per item
  • per batch
  • by tiered usage bands

Simple example:

  • includes first 500 tickets/month
  • 501-1,000 tickets: $1.25 each
  • 1,001+: custom review or next volume tier

Now if usage jumps, you do not have to improvise under pressure.

You already agreed on the rules.

That matters because buyers hate surprise pricing, but they also understand paying for extra volume if the logic was clear up front.

5. Price exceptions and human review separately#

This is the part almost everyone underprices.

If an agent needs approval, escalation, manual correction, or policy review, that is not normal automated throughput.

That is premium operational handling.

Price it separately.

Examples:

  • includes 25 exception reviews per month
  • extra exception bundle: $250 per 25
  • human QA on outbound replies: additional monthly fee
  • high-risk categories always routed to manual review under separate SLA

This protects you from the worst version of agent work: the client who says, “Great, now can your system also handle all the weird stuff?”

Yes. For money.

A practical pricing formula#

If you want a dead-simple framework, use this:

Monthly price = platform fee + included volume band + expected exception load + overage rules

And separately:

Initial project price = setup fee

You can think about it like this:

  • setup fee = make the system real
  • platform fee = keep the system running
  • included volume = normal work
  • overages = more work than normal
  • exceptions = harder work than normal

That is the whole game.

A concrete example: lead triage agent#

Say you are selling an AI lead triage system for a service business.

The system:

  • ingests leads from forms, email, and DMs
  • classifies lead type
  • enriches basic context
  • drafts a response
  • routes hot leads fast
  • flags low-confidence cases for review

A sane pricing structure might look like this:

Setup#

$3,000 one-time

Includes:

  • workflow audit and mapping
  • CRM + inbox integration
  • classification logic
  • response drafting rules
  • routing and escalation rules
  • test set and launch tuning

Monthly#

$900/month platform fee

Includes:

  • system monitoring
  • basic maintenance
  • small monthly rule updates
  • reporting
  • first 400 leads processed

Overage#

  • 401-800 leads: $1.50/lead
  • 801+: move to next plan or custom review

Exceptions#

  • first 30 low-confidence reviews included
  • extra review bundle: $300 per 25 cases

That structure is not sexy. It is good.

The client can understand it. You can survive it. Both sides know what happens when reality gets weird.

Price by risk, not just by volume#

Two clients can have the same volume and wildly different profitability.

Client A sends 500 clean, structured inputs with obvious outcomes. Client B sends 500 messy, ambiguous, politically-loaded garbage fire cases.

If you price both the same, you are subsidizing complexity.

So when scoping an offer, ask:

  • how clean is the input?
  • how expensive is a wrong action?
  • how many cases need human approval?
  • how often do the rules change?
  • how messy is the tool stack?
  • how fast does the buyer expect exceptions handled?

That is operational risk. Operational risk deserves pricing.

Use a 30-day repricing clause early on#

When you do not yet know the true workload, just say that.

Not in a nervous way. In a competent way.

Something like:

Pricing includes a 30-day calibration period. If actual volume or exception rates differ materially from the scoped assumptions, pricing will be adjusted to the next agreed tier.

That one sentence can save a lot of pain.

It tells the buyer:

  • this is based on current information
  • we will use real usage, not vibes
  • there is a fair mechanism if reality changes

Smart buyers do not hate this. Smart buyers prefer it to hidden resentment and sloppy service.

What not to do#

A few things I would avoid:

Do not sell unlimited support under an agent fee#

That is how a productized service turns into you babysitting a workflow all month.

Do not promise fixed outcomes when the inputs are unstable#

Promise the system, the process, the SLA, the review gates, and the reporting.

Do not promise magic.

Do not let the client redefine scope through Slack messages#

If they want the agent to start doing adjacent work, that is a scope change, not a fun little optimization.

Do not bury the human layer#

If humans are reviewing outputs, say so. Price it. Explain it.

You are not selling failure. You are selling controlled reliability.

The real mental model#

Here is the cleanest way to think about it:

You are not pricing a model. You are pricing a managed decision workflow.

That workflow has:

  • automation
  • monitoring
  • escalation
  • exceptions
  • accountability

The more unpredictable the environment, the more your pricing needs to reflect workflow management instead of raw software access.

That is also why early on, a lot of good AI-agent businesses look more like productized services than pure SaaS.

And that is fine.

If you want the early-stage version of this playbook, read Sell the Audit Before the Agent and The First 5 AI Agent Offers I’d Sell Before Building a SaaS.

The bottom line#

If the workload is unpredictable, do not hide the variability inside one monthly number.

Break the offer into:

  • setup
  • base platform fee
  • included normal volume
  • overages
  • exception handling

That is how you protect margin without making the offer annoying to buy.

Start boring. Define normal. Price chaos separately.

That is how you avoid building yourself a job with better branding.

If you want help turning a messy workflow into a narrow AI agent offer that can actually be priced and sold, work with Erik MacKinnon.