The question a CFO asks

Before a CFO signs off on an AI build, two questions come first. What is the most this can cost? And when the work stops, what do we own?

The pricing model answers both, long before anyone writes code. It decides who pays when the work runs over and when you find out the total. The rate on the proposal matters less. A fair rate on the wrong model still gives you a number you cannot put in a plan.

Four models, by mechanism

Each model below gets the same four questions: who carries an overrun, when the total is known, what ends the work, and what you own afterwards. We left rates out on purpose. A rate prices an hour, and the question a CFO is asking is how many hours there will be.

Time and materials

You pay for the hours worked, at an agreed rate, usually invoiced monthly. If the work takes longer than planned, you pay for the extra hours, so the overrun is yours. You know the total when the last invoice arrives. The work ends when you stop it or the budget runs out. What you own depends on the contract. Check that the code, the prompts and the evals transfer to you, and that someone on your side can run them.

It suits work whose shape nobody knows yet. It also leaves your plan without a ceiling.

Monthly retainer

You pay a fixed amount each month for a team's time. The monthly figure is known. The total is not, because nothing in the contract says when the work is finished. An overrun shows up as more months at the same monthly figure. The work ends when you cancel, and the notice period decides how fast. You own whatever was delivered and written down along the way. A retainer pays the vendor for staying, so ask what you would keep if you cancelled tomorrow.

Fixed-bid project

The vendor quotes one price for a written scope. You know the total before work starts, and an overrun inside the scope is the vendor's cost. The trouble sits at the edge of the scope. Anything the specification missed comes back as a change request with its own price, so the real total is the bid plus the changes. The work ends at acceptance against the specification. You usually own the delivered system. Whether your team can change it afterwards is a separate question. Ask it before you sign.

A fixed bid works when the scope can be written down in full. For an AI system that is hard, because much of the work is finding out what the model gets wrong on your data, and nobody can write that into a specification in advance.

Per feature

You pay a published price for each feature. What counts as a feature differs by vendor. In our version, it is one AI action (extract, classify, draft, route or answer) on one step of a workflow, shipped with its eval, logs and docs. Before the first feature there is a setup, which measures the workflow, builds the harness, and connects the systems the features will use.

The total is a count times a price: setup, plus any extra systems, plus the number of features. You know it before work starts. Each feature has a fixed price, and you approve the count before work starts. If a step turns out to be bigger than one feature, we split it into more features first, and you approve the new count. The work ends when you stop buying features, and there is nothing to cancel. You own the code, the harness, the evals and the runbook. The catch is the setup: you pay for it before you see a feature.

The four models side by side

Model Total known before work starts Who carries the overrun How it ends What you own
Time and materials No. The last invoice tells you. You You stop it, or the budget runs out What the contract transfers
Monthly retainer Only the monthly figure You, as extra months You cancel, after the notice period What was delivered and written down
Fixed-bid project Yes, for the written scope The vendor inside the scope. You, for change requests. Acceptance against the specification The delivered system, as the contract defines it
Per feature Yes: setup, any extra systems and a count of features Each feature has a fixed price. The count, including any split of a big step, is agreed before work starts. You stop buying features The code, the harness, the evals and the runbook

Where per feature is the wrong model

Per feature works only when a feature has a known shape. In four situations it does not, and we would not quote a fixed price for any of them.

  • The work is research. If the question is whether a model can do the task at all, there is no unit to price. Pay for a fixed block of time and decide at the end of it.
  • Nobody can describe the workflow yet. A feature is one action on one step, so without a list of steps there is nothing to count.
  • We cannot get access to the systems the workflow runs in. Setup cannot connect them or measure a baseline, and any price would be a guess.
  • You need a new platform. A product built from nothing is a sequence of milestones, and features come after the platform exists.

How to count your own features

You can get close to the total before you talk to anyone. Take one workflow.

  1. Write down its steps in order, the way the people who run it would describe them.
  2. Mark each step that needs an AI action: extract, classify, draft, route or answer. Each marked step is one feature. A step that needs two actions is two features.
  3. Count the systems the marked steps read from or write to, such as the ERP, the CRM, a shared mailbox or a ticket queue.

Then put your two counts into this:

Total = $10,000 + (systems − 2) × $3,000 + features × $2,000

Setup includes 2 systems, so if your steps touch 2 or fewer, the middle term is zero.

Take supplier invoices that arrive by email and get keyed into the ERP by hand. The marked steps are:

  • Extract the invoice lines.
  • Classify each line as a match or a mismatch against the purchase order.
  • Draft the email to the supplier when a line does not match.

That is 3 features on 2 systems, the shared mailbox and the ERP. The total is $16,000. The first feature goes live about 4 weeks after kickoff: 3 weeks of setup, then 1 week for the feature.

Get the total for your workflow

If you have a count of steps and systems, bring it to a 30-minute scoping call. On the call we go through your steps and systems with you, and you leave with the total before any work starts. Book the scoping call.

The price list, and what does and does not count as a feature, are on the Priced AI Features page.