The Assistant Costs 35 Cents an Agent-Hour and Your Rate Card Says $24. July's Claude Enterprise Spend Caps Are How the Margin Survives.

You already tried this. In 2025 you put an assistant on the floor for wrap-up summaries and knowledge lookups, the team leads loved it, average handle time came down forty seconds, and then the month-three bill arrived at four times the pilot and your controller switched it off in an afternoon. Nobody could say which program had spent the money, which meant nobody could bill it to anyone, which meant it came straight out of margin.

That was a real problem and it had a real cause: there was no way to cap it, no way to see it by team, and no way to put it on a rate card. On 2 July 2026 that changed.

Why an outsourcer feels an uncapped AI bill harder than a law firm does

A law firm bills its client for the hour and the software is overhead. You sell a productive hour at a fixed rate agreed eighteen months ago in a statement of work with an annual productivity concession baked in — typically 2 to 4% off the rate every year, promised in exchange for the renewal. Your margin on that hour is single digits to low teens. There is no line on the invoice called "software we started using in March."

So an uncontrolled AI bill does not reduce your profit. It eliminates it on the specific program where it landed, and you find out at month-end. Worse, on a fixed-price or per-contact contract you have no pass-through at all. On a time-and-materials or gainshare deal you might, but only if you can produce the spend broken down by program — which the 2025 tooling could not do.

The way to think about this: AI in a contact center is not a project with a budget, it is a cost per agent-hour, and it belongs on the same page as headset amortisation, telephony minutes and desktop licensing.

What landed on 2 July 2026

The Claude Enterprise governance update on 2 July 2026 added the four things a finance-run business actually needs. Usage and cost analytics, so you can see what was spent and by whom. Spend limits at the organisation level and at the individual user level, so a runaway does not become a discovery at month-end. Alerts at 75% and 90% of the limit, so somebody hears about it while there is still time to act. And model defaults and entitlements, so you decide which people get the expensive model and which work runs on the cheaper one, rather than every team lead choosing for themselves.

Entitlements are the underrated one. In a 300-seat center you do not want every agent reaching for the heaviest model to summarise a four-minute billing call. You want the fast, cheap model as the default for wrap-up, and the heavier one available to the escalation desk and the complaints team, where the reasoning is genuinely worth paying for.

flowchart TD
  A["Monthly AI budget set from the rate card"] --> B["Retail care team cap"]
  A --> C["Healthcare intake team cap"]
  A --> D["Collections team cap"]
  B --> E["75% alert to the operations manager"]
  C --> E
  D --> E
  E --> F["Cheap model becomes default for wrap-up summaries"]
  F --> B
  E --> G["90% alert to the controller, escalation desk only"]

Setting the cap so it is a rate-card line, not a surprise

Work backwards from the bill rate, not forwards from a vendor quote. Take one program. Bill rate, target margin, and the share of that margin you are willing to spend on tooling. That number divided by productive hours is your allowance per agent-hour, and that is the number your caps get set to.

Then split the organisation cap across teams the same way your general ledger already splits everything else — by program, because that is how you win, lose, renew and price work. If the healthcare intake team burns its allowance on day 19, that is a fact about that program's economics, and it should show up in that program's monthly review rather than being averaged away across the site.

Set the 75% alert to go to the operations manager who owns the program and the 90% alert to the controller. Those are two different conversations. At 75% the question is "is this working hard enough to justify the run rate?" At 90% the question is "what stops today?"

The arithmetic: what you can actually afford per agent-hour

Illustrative assumptions: 300 seats across three programs, 160 productive hours per agent per month, blended bill rate $24 per hour, target margin 12%. You decide AI can consume up to 12% of that margin.

Now check it against real usage. If an agent takes 45 calls a day for 22 days, that is 990 calls a month. Suppose a wrap-up summary and a knowledge lookup on each of those costs three cents all in — which is a plausible illustration given frontier AI pricing fell roughly tenfold between 2025 and 2026, with capable models now around a couple of dollars for a very large volume of work. That is $29.70 per agent per month against an allowance of $56.00. You have headroom, and — this is the part your controller will care about — you know you have headroom before the invoice arrives rather than after.

One more sanity check. Forty seconds off average handle time on a 990-call month is eleven hours of capacity per agent. At $24 a productive hour that is $264 of value against $29.70 of cost. If your measured improvement is not several multiples of the spend, this use is not working on this program — and the cap told you in month one instead of month eight.

What a spend cap will not save you from

It will not tell you whether the money bought anything. A cap controls the bill, not the return. You still need the before-and-after on handle time, after-call work, quality score and repeat-contact rate, measured on the same program, ideally with one team held back as a comparison for a month. Nobody enjoys running a holdback team. Do it anyway; it is the only thing that survives contact with a sceptical client.

It will not fix a badly-drawn contract. If your statement of work has no clause about tooling changes, and the client believes any efficiency gain is theirs by right under the productivity concession, you can save forty seconds a call and hand every cent of it back at renewal. Get your commercial lead to read the change-of-method and gainshare wording first.

And a hard cap has a sharp edge. When the collections team hits its limit on 24 December, the assistant stops for the agents who are actually working the holiday queue. Decide in advance whether a cap is a hard stop or a warning with an override, who owns the override, and how fast they answer the phone.

Finally, budgeting AI does not exempt you from the rules governing it. Texas TRAIGA and California SB 53 both took effect on 1 January 2026, and the EU AI Act's transparency and high-risk obligations carry a 2 August 2026 date that reaches US operators whose systems affect people in the EU — relevant if you run English-language support for a European client. Federal preemption of state rules was still unsettled as of July 2026, so state law binds. A spend dashboard is not a compliance record.

Three numbers on a whiteboard on Monday

Write down your blended bill rate per productive hour, your target margin per hour, and the share of that margin you are willing to spend on tooling. Multiply the third by your monthly productive hours. That is your cap. Set it, split it by program, point the 75% alert at the operations managers and the 90% at finance.

Then put one line in your next monthly operations review, between telephony and facilities: AI cost per productive hour, by program, against allowance. Once it sits in that meeting every month, it is a managed cost like every other managed cost — and no controller ever switched off a managed cost in an afternoon.

Frequently asked questions

Should I pass AI cost through to the client or absorb it?

Depends entirely on the contract type. On time-and-materials or cost-plus, itemise it and pass it through — clients accept a named tool with a visible per-hour cost far more readily than a rate increase. On fixed-price or per-contact, you absorb it, which means the cap matters more and the measured handle-time improvement has to be real. On gainshare, put it in the cost base before the split is calculated, in writing, before the first quarter closes.

How do I stop agents from using the expensive model for everything?

You do not police it, you configure it. Model defaults and entitlements mean the cheaper model is simply what a care agent gets, and the heavier one is granted to the escalation desk, the complaints team and the quality analysts. It removes the discussion instead of turning your team leads into cost police, which is a job they will do badly and resent.

What is a sensible first cap if I genuinely have no idea?

Set it at one to two percent of program revenue and watch for a month with the alerts on. That is deliberately tight; the point of the first month is to learn the shape of the usage curve — which is almost always front-loaded Monday morning and heavy in the first hour of each shift — not to fund everything anyone wants. Raise it in month two with evidence.

Our client wants to see what we spend on AI. Do we show them?

On an open-book or cost-plus program you will have to, so build the reporting assuming they will look. On a fixed-price program, showing it invites a conversation about who owns the savings. Either way, decide it deliberately with your commercial lead rather than discovering it when a sourcing manager asks for the detail behind a line item mid-renewal.

A note from CallSphere

CallSphere builds AI voice and chat agents that answer business phone lines and web chat, book appointments and capture leads 24/7. If you are working out what AI should cost per hour on the floor, the same question applies to the calls arriving at your own front door — sales inquiries, RFP questions, candidate callbacks during a ramp. Those are a known volume with a known cost per answered call, which makes them one of the easier lines to budget.