Check What Your AI Contract Counts

It’s renewal week, and three AI contracts are on the table. All three sell a seat, and each adds a different second meter on top: credits for advanced work, tokens, credits plus a price per resolved case. Your CFO asks what AI will cost next year, and there’s no single unit to answer in. The meeting is imagined, not one we’ve sat in, but the contracts are real, because Microsoft, Salesforce and Anthropic each describe their own version of this.

Before you renew, write down what each contract counts, which work the flat fee covers, your ceiling and what happens at any cap, and what one real task would cost. That’s four checks and one short record per contract, and it’s worth starting about 90 days before each renewal date, or before the notice deadline if your contract has one, because a meter can change, and at Anthropic that happens at renewal for its older plans. First, what the three vendors say.

Three vendors, three second meters

Microsoft’s September 25, 2026 post on Copilot pricing splits AI into two kinds of work. Everyday AI is “priced per user per month”, the way software has long been bought. Advanced AI, which Microsoft describes as “agents that run for hours” and “workflows without supervision”, runs on usage-based billing with Copilot Credits. Microsoft also takes a side on metering everyday work: “Some vendors put everyday AI work on a meter. We think that’s the wrong deal.”

Anthropic’s current Enterprise billing page describes a different design. The seat fee “doesn’t include any usage”, and every token (a small chunk of text the model reads or writes) the team consumes is billed separately at standard API rates. There’s no individual token allowance either. All usage across the organization is billed together, regardless of who consumed it.

Salesforce’s September 3, 2026 announcement sells three editions at $195, $395 and $550 per user per month, each with a block of Flex Credits (500,000, 1 million and 2.75 million, per the announcement). It also says its Agentforce Help Agent, included in the Advanced and Max editions, comes with outcome-based pricing, so “you only pay when issues are resolved.” Its pricing page lists Help Agent resolutions at $2.

A seat still does a lot of work. Microsoft says it’s setting its usage limits so that the vast majority of people never reach one, and for everyday use the seat may well stay the bill. What Microsoft keeps out of its seat is an agent working for hours. Gartner’s August 17, 2026 release says sending a task to an agentic reasoning model costs the provider “at least five times” as much to run as a basic chatbot exchange, which is a cost to the provider, not a price to you. Will Sommer, the Gartner analyst quoted in the release, describes an agent as one that “must constantly reason, negotiate, and question itself.” Microsoft’s post says its state-of-the-art AI experiences “cost more to run”.

Four checks

1. Name the meter

For each AI contract, write one line saying what the invoice counts: a person, a token or credit, or an outcome. Expect more than one: Microsoft’s license plus credits, Anthropic’s seat plus tokens, a Salesforce edition plus Flex Credits.

Do this first, because nothing else can be checked until you know what’s being counted.

2. Find where the flat fee stops

Once you know what’s counted, list the work your team really does with the tool and mark each task flat, flat with a limit, or metered. At Microsoft, everyday work sits under the license with fair-use limits: a heavy user who hits one gets a warning, then a choice between Microsoft’s Auto setting at no extra cost or moving to credits. On Anthropic’s current single Enterprise seat there’s no flat part, while on its older Standard and Premium seats the flat fee includes a usage allowance, so mark which tasks fit inside it.

At Salesforce the same employee can land on either side. Its pricing page lists a $125 per user per month add-on with “Unmetered Agentforce usage for employees” and a $5 per user per month license for “Metered Agentforce usage for employees” that requires Flex Credits.

3. Set the ceiling, and say what happens at it

Now that you know which work is metered, write down your own ceiling for each metered line, the most you’ll let it cost per billing period, then the cap if the contract has one, who gets the alert, and what the person sees when they’re blocked. The vendors differ here, and none of the pages we read shows the blocked screen, so where a vendor has a limit, test it with one account and a low limit.

On its current Enterprise seat, Anthropic lets owners set a limit for the whole organization and for each person, or set it to “unlimited”, which is still billed. On prepaid self-serve plans, usage stops for the whole organization when the credits run out. On invoiced (sales-assisted) plans, a limit stops usage until the next billing period or until an owner raises it.

For enterprise customers, Microsoft keeps usage-billed services off until an admin creates a spending policy, and nothing is billed before that. Admins set budgets at the tenant and group level, including per-user caps inside group policies, and pick what happens as someone approaches a limit: a warning, a way to request more, or no prompt at all. The post doesn’t say what happens when a budget is reached.

Salesforce’s pricing page says “There is no overage penalty.” If you exceed your entitlement, it says, you pay your contracted rate, billed monthly in arrears, and it points to alerts in its Digital Wallet tool to prevent overages. We found no hard stop described on that page.

So the pages describe a limit that stops the work at Anthropic, budgets and caps at Microsoft whose behavior at the limit the post leaves unstated, and alerts at Salesforce. Where a cap stops the work, it protects the budget at the cost of that work, and nothing we’ve read says which costs more. Microsoft argues that when every task has a meter, companies “may ration access”, and Anthropic’s admin page reports “member-days lost”, the days of blocked usage a limit has caused that month. The call is yours.

4. Price one real task by measuring it

With a ceiling set, pick a task your team runs, not a demo, and run it a few times in a pilot, keeping the highest reading, since one run is a single sample. Put a spend limit just above what you expect on the pilot where the vendor offers one, and where it offers only alerts, set the alert low and watch it. Read what it consumed in the vendor’s own usage view: Salesforce’s Digital Wallet shows Flex Credit use in near real time, Anthropic’s admin page shows month-to-date spending for each member, and Microsoft’s post says its reporting is broken out by user, group and workload, with consumption insights for Cowork, one of its advanced AI tools, being rolled out to show which scenarios drive credit use. Multiply by today’s volume and by the volume you’d run if it works, then ask each vendor in writing to confirm the rate and your total in dollars per billing period.

The pages give you unit prices at best. Salesforce lists $500 per 100,000 Flex Credits, and four of its five examples work out to 20 credits per action, and the four-action voice example, at 120 credits, to 30, though the page says voice multipliers differ and points to a rate card. Its examples run from one to five actions per use case and from 20 to 120 credits, so the number to find is how many actions of which type your task takes. Anthropic points to an API rate card by model, and Microsoft’s post lists no dollar amount for a credit, so ask. Treat the answer as an estimate and the first real invoice as the first real number. The pilot and the highest-reading rule are our method, not any vendor’s.

What the record decides

Check four needs the unit from check one, the metered work and the ceiling, and its totals feed one rule. If a line carries several tasks, repeat check four for each one and add the totals. Then hold each line’s total at the higher volume against your ceiling. If it fits under the ceiling, renew as is. If it doesn’t, renegotiate the unit price or the allowance before the date, ask for a not-to-exceed limit in the contract where there’s no hard cap, or keep that work off the meter where the vendor sells a flat option, as Salesforce’s unmetered add-on does for employee use. Anthropic’s current seat has no flat part, so there the levers are the limit and the rate. If you buy a block of Flex Credits, size it with care, since unused credits don’t roll over to the next term.

If a vendor won’t confirm the rate in writing, or the total you worked out from your measurement, don’t grow the metered work with them until you’ve seen a real invoice. A vendor that bills at standard API rates, as Anthropic’s page describes, may decline to guess your token count, so there a written rate is enough and your own measurement is the quantity. The not-to-exceed clause and the don’t-grow rule are ours, not any vendor’s.

The CFO from the opening gets a ceiling to answer with, not a forecast: seats, plus the most you’ve allowed each metered line to add per billing period, with check four’s totals showing whether that ceiling is realistic. At Anthropic a limit that’s hit stops the work. Where a vendor offers only alerts, as Salesforce’s page does, the ceiling is a number you watch until a contract clause makes it a limit, so set the alert below it, name who acts when it fires, and ask for the clause.

Why 90 days before renewal

The meter can change, and renewal is when Anthropic says its older Standard and Premium plans do: their flat fee per seat includes a usage allowance, and they move to the usage-based single Enterprise seat at the next renewal. Microsoft says some capabilities will start on usage billing and move into the per-user license “as the economics allow”, and Salesforce says “Pricing and packaging are subject to change”. Neither page ties that to renewal. Salesforce also says unused Flex Credits “do not roll over into subsequent subscription terms”, so the end of a term is when an oversized block is lost.

Ninety days is our number, picked so there’s time to ask, compare and negotiate. None of the vendors gives one. If your contract needs notice to change or end the term, count the 90 days back from the notice deadline instead of the renewal date.

What the check can’t tell you

It tells you what to ask, not which meter is cheaper, and it doesn’t promise a lower bill. What it gives you is a known unit, a ceiling you set and a written total before you commit. It also covers only what a vendor meters: the time people spend on an agent is a cost no invoice shows, covered in An Agent Is Not an App, and a tool you build yourself brings its own monthly token bill, which Don’t Build What You Can’t Keep shows how to price. Token rates also move on their own: when a model maker cuts its price, The AI Price in Your Budget Is Really Two Prices covers what to test before you book the saving.

We found no source that tests whether the check lowers spend, and none that compares what the three charge for the same work. The one comparison we found, in Microsoft’s own post, is labeled internal analysis and covers Copilot, Claude and ChatGPT, and the saved text of the post gives no figures from it. Every vendor source here is a company describing its own pricing, Gartner’s “at least five times” is the provider’s cost, and Salesforce marks four of its five examples illustrative and says to contact a sales representative for detailed pricing. The ceiling trade has no verdict either: stopping the work and exposing the bill are both costs, and no source weighs them.

Before the next renewal

Pick the AI contract with the nearest renewal or notice deadline, since it has the least time left, and do check one: write down what it counts. If the answer is “people” for everything, ask the vendor whether any of the work your team does falls on a meter, because an agent that runs for hours may.

The meter is a vendor’s design choice, and these three chose three different second meters. It’s easier to read that choice off a billing page before you renew than off an invoice afterward.