Price AI for casework by the case

Operations leaders manage by the case

Every business case I wrote for a transformation programme came down to two numbers: what a case costs, and how many cases a handler closes in a month. I wrote a lot of them in legal operations and managed services, where my job was buying technology and outsourced services and then running them. Time to outcome sat beside them, because it is the number complainants and clients feel. Finance directors set headcount from all three.

Most vendors sell AI for casework by the seat or by usage, counted in tokens or pages. Both suit the vendor’s revenue forecast, and neither is a unit an operations team manages. I think buyers should pay per case, the unit the team already runs on. A per-case price also leaves the cost of an inefficient platform with the vendor, who is the one able to do something about it.

Seat and usage pricing leave the risk with the buyer

Seat pricing suits software whose value grows with the number of people using it. Operations leaders buy casework AI so that the same team can handle more cases. If the platform works, the seat count stays flat while volume climbs, and the vendor earns more from you when you hire, which is the cost you bought the product to avoid.

Managers also ration seats to the handlers with the heaviest caseloads. The supervisor and the quality team end up without a login, though they are the people who check what the platform produced and what the handler changed.

Usage pricing ties the bill to things the team does not control. A delayed-refund complaint can arrive as two screenshots and an email chain, while a housing disrepair claim brings years of repair logs and surveyor reports. A finance director setting next year’s budget cannot turn that spread into a number, so the contract ends up with a cap and a quarterly argument about the overage.

Metering also charges for the behaviour a careful team wants. A reviewer who checks the chronology against the source pages, or asks the file one more question before signing off, adds to the bill, and sooner or later someone asks the team to check less. Meanwhile the vendor earns more whenever its system sends more of the file to the model, so the contract gives it no reason to make the platform efficient.

What to settle before agreeing a rate

A per-case price turns the AI bill into arithmetic finance already does for headcount: forecast volume multiplied by a rate. If a vendor asks for a committed annual volume, set it from your lowest credible forecast and agree the rate for cases above it in the same negotiation. A surge in the autumn then moves the bill by an amount the finance director could have worked out when setting the budget.

The rate itself should come out of a trial on your own closed cases, picked to match the real caseload. Your team knows how those cases ended and which evidence mattered, so it can judge the platform’s output against the decision it reached without touching an open complaint or client matter. Measure the trial in the units you will manage afterwards: preparation time per case, and the number and kind of corrections your reviewers made. Those figures tell you whether the rate is fair, and finding out costs less before signing than in the second year of a contract.

Accountability and outcome-based fees

Managed services taught me that an organisation can outsource a process but keeps the accountability for it. A complainant or a court will ask who decided and on what evidence, and a supplier cannot answer that for the buyer. The decision belongs inside the organisation, with a person whose name is on it.

The price should follow the same principle. A vendor should earn its fee on each case it prepares, whatever the outcome. A price tied to upheld complaints or settlements gives the supplier a stake in the result, which an impartial scheme cannot accept and a firm acting for its client should not want.

At Ctrl AI we charge per case. Clients trial the platform on a representative sample of their closed cases before we configure it for their day-to-day work, and a handler in the client’s own team checks and confirms each assessment before the team acts on it.