Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence Service Pricing: What Operators Should Test First

How service companies can test whether more capable agent work changes their quotes: define the unit, measure review effort, and set rules before any price moves.

Be Human typographic cover with the title: Superintelligence Service Pricing: What Operators Should Test First
Conceptual illustration created with AI · Be Human

Superintelligence service pricing is the question of whether a service company should change how it quotes work once agents can handle more of the work itself. Nobody can yet say with confidence which direction the cost moves, so the sensible approach is to test the quoting model on real tasks before touching a price. Operators should measure the effort each unit of work takes today, measure it again with agent support, and decide in advance which evidence would justify a change.

Key takeaways

  • Quote against a defined unit of completed work, not against hours that someone happens to spend, and check whether that unit still holds when an agent does part of the task.
  • Run a bounded, labeled test on a few recurring task types before any quote changes, and record effort, rework and review time for each one.
  • Count human review and approval as cost. In many services, checking the output is the part that does not shrink.
  • Write the quoting rule, the evidence that would trigger a change and the person who signs it off before the test starts.
  • Expect results to vary by task type and by client. Treat one good run as a hypothesis, not a new price list.

Why a capability change reaches your quote at all

A quote is a set of assumptions about effort. A fixed fee for a monthly reconciliation assumes a certain number of person-hours, a certain amount of back-and-forth with the client and a certain level of exception handling. When a tool changes how that work gets done, those assumptions can go stale in either direction. Preparation might take a fraction of the time it took before, while review of exceptions might take longer, because the reviewer now has to check work they did not produce.

The mistake to avoid is assuming the direction in advance. Some operators will expect lower costs and cut prices to win work, and then discover that scope expanded or that review consumed the savings. Others will assume nothing changes and keep quoting the same way while their margins shift underneath them. Both positions skip the step that produces usable information, which is measuring the work as it is actually performed.

Define the unit of work before you test anything

The unit is the thing the client actually buys: a closed ticket, a reconciled batch, a drafted proposal that passes review, a resolved support case, a processed application. A good unit has a clear completion standard, so that two people looking at the same output would agree whether it counts. If the unit is vague, every measurement that follows will be vague too.

Once the unit is clear, look at how your current price relates to it. Some firms price per unit, some per month with a volume band, some per project. The structure matters because it determines who benefits when effort falls. The comparison of usage, seat and retainer structures in AI Agent Pricing Models Explained: Usage, Seat or Retainer is a useful reference when you decide which structure fits a given service. The FinOps Foundation's guidance on unit economics makes the same general point: cost and revenue are easiest to reason about when they are tied to a unit that the business can count.

A test plan for a bounded trial

A useful test is small, labeled and reversible. It does not reorganize the team or change client terms. It produces a set of comparable records. The following sequence works for most service firms:

  • Choose three to five recurring task types that have clear completion standards and enough volume to show a pattern.
  • For each task type, record the current effort in person-minutes by step, along with the number of revisions clients request.
  • Run the agent alongside normal delivery in a shadow mode, where a named person reviews every output before anything reaches a client.
  • Log agent run time, tool calls that failed, review minutes, corrections made and any output that had to be redone.
  • Compare unit effort with the assumption built into the current quote, one task type at a time.
  • Note every case where review took longer than the original work would have taken, and find out why.
  • Decide, for each task type, whether to keep the current quote, adjust one rule, or hold and run the test again with a different scope.

The last step matters most. A trial that ends without a decision has produced data but no operating change, and it tends to be forgotten.

What to measure, and what to ignore

The measures that matter are the ones tied to the unit. Review minutes per unit, correction count, escalation count, time from request to accepted output, and the severity of errors that reached review are the core set. Add the usage cost of the agent per unit if your agent is billed by usage, so that the trend is visible alongside the human effort.

What to ignore is the general capability score of a model. Public benchmarks describe performance on tasks that may not resemble yours. Our piece on Superintelligence Benchmarks: Which Ones Predict Business Results explains why a benchmark rarely predicts a specific business outcome. Anthropic's engineering notes on evaluating agents describe the same discipline: define realistic tasks, define how success is graded, and run the same tasks repeatedly so that changes can be attributed to something real.

Illustrative scenario: a monthly close service

The following scenario is illustrative and does not describe a real client. A small accounting services firm sells a fixed monthly close for each client. The firm wonders whether an agent that drafts account reconciliations and flags unusual entries should change that fixed fee.

The test covers three client groups. For clients with simple, stable accounts, preparation time falls clearly, and reviewers catch few problems. For clients with many intercompany entries, the agent drafts quickly but produces more flagged items, and reviewers spend longer checking them. For a third group, the agent's output is rarely usable and the work reverts to the old process.

The firm does not conclude that the agent makes the service cheaper. It decides that the standard close for simple accounts can move to a lower fixed fee tier with a volume allowance, that complex accounts keep the current fee with a written clause covering exception volume, and that the third group stays on the current process until the agent handles their input format. Each decision is recorded against the test evidence, and the partner responsible for pricing signs each change.

Deciding when a quote changes

A quote change should be a decision with a written trigger. Before changing any price, check the following:

  • The unit of work and its completion standard are written down and unchanged during the test.
  • The effort comparison covers enough completed units to show a pattern, not one or two unusual cases.
  • Review and approval time are included in the effort figure.
  • The change applies to a named task type or client group, not to the whole service.
  • Client terms explain what is included, what triggers an additional charge and how exceptions are handled.
  • A named person has approval authority for pricing changes, and the agent does not change prices itself.
  • A review date is set, after which the rule is checked against new evidence.

The last two points deserve particular attention. Pricing rules should be governed by people, and the boundaries of what an agent may do should be explicit. Human Approval Thresholds for Stronger AI Agents: A Practical Guide covers how to set those boundaries as capability grows. The NIST AI Risk Management Framework is a practical reference for mapping and measuring the risks of AI use in a way that a leadership team can review.

Be Human's managed approach is built around this discipline. Each engagement starts by defining one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, which gives a pricing test a stable baseline. Managed operations include recurring evaluations, permission review and controlled changes to the agent, so that effort data is collected the same way from one period to the next and changes to the agent are requested and approved rather than made informally. Be Human does not set your prices. It supplies the operating evidence your pricing decision depends on.

Where to start

Start with one service line and one recurring task type where the completion standard is already clear to everyone who works on it. Measure its current effort for two weeks before the agent touches it, then run the shadow trial and compare. Keep the quote unchanged until the comparison is complete and the approver has read it. If you are not sure which task type deserves the first trial, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Will more capable agents lower my service prices? No one can promise that, and the answer depends on your task mix, your review requirements and how clients respond to changes. A reasonable approach is to test the effort on real units of work, separate the task types that show a clear change from those that do not, and adjust prices only where the evidence supports it. Treat any claim of a fixed reduction with caution.

How long should a pricing test run? Run it long enough to see a pattern across several completed units of each task type, and not so long that the process itself changes during the test. The right window depends on your volume, so define it by the number of completed units rather than by a calendar period. Write down the stopping rule before you start.

Should I tell clients that we use AI agents in their work? Disclosure depends on your contracts, the kind of work you perform and the rules that apply in the markets where you operate. Review your client terms and get advice from counsel before changing how you describe your service. Whatever you decide, make sure the people who sign client-facing commitments know exactly where agents are involved.

Do I need a data scientist to run this test? You do not need one to run a useful test. You need a clear unit of work, a simple log that records effort, review and corrections, and a named person who reads the results and makes the decision. Specialist help becomes more useful when the test shows a pattern that needs deeper analysis.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →