One of the software invoices in your firm’s accounts payable now reads like a utility bill: units consumed, a rate per unit, and a total nobody could have predicted at signing.
That bill is older than it looks. Computing started on it: the time-sharing bureaus of the 1960s charged for processor time and connect hours, and the customer learned at month’s end what the answer had cost. Enterprise software later settled into seat pricing, one flat price per person per month, and the invoice stayed boring for decades. The AI vendors have swung it back: in an April 2026 note drawn from meetings with about 40 software companies, Goldman Sachs described the shift to units of work. Salesforce sells “agentic work units,” Workday sells credits tied to “units of work,” and OpenAI’s Sam Altman describes selling tokens like electricity. The tools your firm buys are once again priced on the work they do.
Stephan Liozu, a pricing strategist writing in IndustryWeek in July, told those vendors the model will fail: industrial CFOs won’t sign an open-ended metered bill for a technology whose output they don’t yet fully trust. He summed up consumption pricing in one line: it “is not a business model. It is a confession that you do not know what your product is worth.”
Watch what your clients do to their AI vendors’ invoices. Yours are next.
The fight should sound familiar
A meter charges for inputs consumed and leaves the buyer holding the uncertainty. That describes a token contract, and it describes an hourly engagement letter.
Winston Weinberg runs Harvey, which sells AI to law firms, so he profits from the argument that follows. He told the Financial Times in May that the hour was never what the client was buying: “If I pay a law firm $100,000 to do something, I actually paid for the partner to give me advice on what to do. It wasn’t billed that way though: it was billed as if a bunch of junior associates did all these research tasks at $1,000 an hour.”
A partner will object that hours aren’t tokens, and the objection is half right: compute is a commodity with a posted price, and judgment isn’t. But neither meter prices the thing the buyer came for. The client paying $1,000 an hour wanted the answer, the enterprise burning tokens wanted the finished task, and both were billed for the machinery in between. The hour was packaging, and AI has been pulling that package apart.
The parallel breaks in one place, and it breaks in the firm’s favor. An AI vendor that commits to a fixed price is betting its gross margin against compute costs it doesn’t control, so almost none of them will. A services firm runs the opposite cost structure, with compute a small line next to professional fees. What the vendor can’t afford to offer, a firm can.
In-house teams learned it first
In the same interview, Weinberg described the refusal forming inside legal departments: “If you’re an in-house team and there’s something that you can do with AI, you’re not going to accept crazy bills from a law firm to do that particular task.”
EY put a size on that refusal in its April–May 2026 US AI Pulse Survey: 82% of senior leaders whose organization is investing in AI say the traditional per-seat software pricing model will become less relevant in their industry within five years, and 95% say their relationship with traditional software vendors will change. The seat was the last flat price in enterprise software, and the buyers have stopped defending it.
The AI subscription and the outside-counsel fees sit in the same quarterly review now, in front of the same CFO, and a concession won from the software vendor in one meeting becomes the opening ask in the next. A firm billing hours has no better answer to that ask than its own vendors had.
Some sellers have already stopped defending the meter. Norm Ai, another legal AI company, raised $120 million in July in a round led by Khosla Ventures, and launched Norm Law, an affiliated AI-native firm built to sell outcome-priced legal work to exactly the in-house teams Weinberg is describing. Norm Law’s own pitch makes the case against the hourly invoice in the client’s voice: with traditional outside counsel, “you pay for the time it takes to understand your business, and then pay again when that understanding walks out the door.”
What a fixed price risks now
The objection a partner will raise isn’t about compute. Firms avoided fixed fees for decades because scope overruns ate the margin: the engagement that sprawled past its estimate, the rework nobody budgeted, the client who kept calling. AI compresses exactly the hours that used to blow those budgets, which shrinks the exposure without deleting it.
What removes the rest is a defined offering underneath the price: a named scope, a stated result, and a boundary a client request either fits inside or doesn’t. The packaging has to come before the price can: deciding what the firm sells and writing down where the work starts and stops, so the number attaches to the result the client came for rather than the inputs consumed along the way. That design work is the part most firms haven’t done, and it can’t be bought from a vendor.
The common response in the field is to buy the tools and leave the invoice alone. An hourly invoice from a faster firm shows fewer hours on it, and the client pockets the difference. A firm that does the packaging keeps the efficiency gains, and the client gets the predictability they’re now demanding from every other seller.
The CFO who beat an AI vendor’s meter down to a fixed contract this year is the same person who opens your next engagement letter, and it reads the way the vendor’s first offer did: units consumed, at a rate.