The AI Business Case Has a Cost Line. McKinsey Just Showed It Moves.
Two McKinsey pieces from 14 October describe a pilot that costs little and a production bill that does not. Most business cases are written from the pilot.
By Mark Rhys Thomas
The Lead
The cost line in most AI business cases points one way. McKinsey published two pieces on 14 October that show it can point the other.
The first is "The AI value curve". It prices AI against the fully loaded cost of doing a task today, counting model and token cost plus the human review and rework still needed to get an acceptable result. Where AI comes out cheaper, the task is in the pool. McKinsey puts that pool at $2.52 trillion a year globally. Healthcare is $310 billion of it.
Then it raises the cost of AI by 30 percent. Of the $960 billion US pool, $320 billion drops out. Globally the pool falls from $2.52 trillion to $1.67 trillion, down 34 percent. McKinsey calls this an illustration, not a forecast, and says it is smaller than the swings AI pricing has already shown. It also says cheaper AI pulls work back into range just as fast. The finding is not that costs will rise. It is that the answer moves a long way on a modest change.
The second piece, "What does AI really cost?", explains why the bill might rise. In the first quarter of 2026, providers moved from per-seat licences to consumption pricing. One McKinsey expert says the same usage can cost two to three times as much overnight. Another describes organisations where usage and spend doubled between months two and four, because people drifted from one model to a more powerful one they did not need. A third cites a bank whose projected AI spend went from $3 million a year to more than $25 million once everyone had every model with no guardrails.
The same expert adds that moving from pilot to production nearly quadruples AI spend on average. That figure is attributed to "our research", with no source, and a version of it appeared in an earlier McKinsey piece. I treat it as one unverified claim, not as evidence.
Put the two together. The value curve says the case is sensitive to cost. The explainer says cost behaves badly once real people use the tool. That is the Pilot Trap seen from the finance side. A pilot is cheap because few people use it, few agents run, and nobody has drifted to the expensive model. A business case written from the pilot inherits that price. Production does not.
In the business cases I have seen, the cost line is a single per-task number that falls over time. The sensitivity table, where there is one, tests adoption and not price. Nobody owns the cost of a successful output. McKinsey's own metric is exactly that: model calls, human validation, storage and the failed runs that have to be repeated, set against the fully loaded cost of the person or tool doing the job today. Pharma already thinks this way about a batch, where a failed batch sits inside the cost of goods. An AI task that is rerun, re-reviewed and sent back through approval has a cost of goods too. That mapping is mine, not McKinsey's.
The rebuttal is fair, so take it. McKinsey sells the remedy. The value curve is its own proprietary tool, and the explainer ends on a recommendation for one control plane over AI spend. The anecdotes are anonymised client work from banking and engineering teams, not pharma. The four-times figure has no source. And the sensitivity runs both ways. The same explainer notes that per-token prices have fallen by almost 90 percent, a figure the three experts disagree on the usefulness of. A sceptic can say the bill may go down. It may. Then the case should survive that too, and in the cases I have seen the test is missing in both directions.
What survives the discount is a method, not a number. Take one live or planned AI use case. Write its cost per successful output, with human review inside the figure and not beside it. Re-run it at 30 percent higher cost and at 30 percent lower. Ask what the bill does when ten times as many people can reach the expensive model. If the answer flips from yes to no on a 30 percent move, that is the real margin of the decision, and it was not on the slide.
A pilot proves the technology works. Only the bill proves the business case does.
Also on the Line
COMPUTE
Open-weight models passed half of gateway tokens and still earn a fraction of spend. Vercel's September AI Gateway Production Index, covering August traffic, put open-weight models at 56 percent of tokens, up from 13 percent in April, but 14 percent of estimated spend. TLDR AI reported the next index this week at 71.5 percent of tokens and about 23 percent of spend. I have not read that one at source. Spend is estimated from list prices, the users are production teams on one vendor's gateway, and that vendor sells the routing. None of it is pharma. It matters to the lead because routing work to cheaper models is the cost lever McKinsey's experts keep naming.
DATA
The discount on a coding agent is your prompts. Meta's Muse Code, a terminal coding agent launched in beta on 5 August, has two tiers. The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens, and does not let Meta train on prompts. The contributor tier costs $0.10 and $0.20, in return for permission to train on prompts and completions. MacRumors headlined it as up to 20 times cheaper. A price cut paid for in data is the move a cost-pressured team will be offered, and a pharma team's prompts carry proprietary content. This is secondary reporting. Meta's terms were not read.
PROCESS
Same request, different compliance verdict. A single-author paper posted to arXiv on 15 October sent identical requests to fourteen commercial model deployments, asking for preliminary assessments of medical device documents against regulatory requirements. All fourteen varied on at least one request. Twelve changed the compliance label itself. Lowering reasoning settings did not remove the variation. The author then built a pipeline that was deterministic at every stage and returned byte-identical outputs on glucose meter and infusion pump documents. I read the abstract, not the full paper. It is a preprint, and I have not seen peer review.
What Did Not Meet Spec
Thirty-one emails came through this week's pull across McKinsey, KPMG, EY and TLDR AI, plus three newsletters pasted in by hand and web hubs at McKinsey, Accenture and Deloitte checked directly. The two McKinsey cost pieces were read in full at source. Rejected by category: sponsored placements and job adverts in the TLDR editions; consumer hardware, robotics demos and benchmark papers; McKinsey digests on sectors with no AI content; a board digest and a webcast promotion with nothing on AI. Accenture's email channel was empty again, Deloitte's likewise, and neither hub could be dated.
Four named items.
McKinsey's "nearly quadruple". It is the most quotable number in the explainer and it suits this issue's lead. It carries no named source, and it has now appeared in two McKinsey pieces, which is one unverified claim repeated, not two confirmations. It is reported above as McKinsey's claim and not used as evidence.
The "almost 90 percent" fall in token prices. It cuts against the lead, since falling prices are the case for the optimistic cost line. McKinsey calls it a commonly cited industry figure and its own experts disagree on whether it tells a leader anything. I have not used it in either direction.
Brightbeam's "fewer than 10 percent of drugmakers are scaling AI in regulated manufacturing". Issue 08 promised to trace it. One search found no pharma originator. The nearest hit is a survey of manufacturers generally, finding that about 10 percent have scaled AI. That is a different population and I have not substituted it. The figure stays rejected.
Two figures that look like confirmation and are not. Brightbeam's relay that Chinese models handled more than five times the tokens of the top four US providers on OpenRouter comes from a different gateway than Vercel's open-weight share. The two are not merged here. And Brightbeam's wording that Meta's coding assistant trains on customer data "by default" is not what the MacRumors report says. It describes two tiers and does not say which is the default, so the Data item above states the tiers.
The Line
The AI case that survives a 30 percent price rise is the one worth scaling.
Sources
- 1. McKinsey, "The AI value curve: The economics of applying AI to existing work", mckinsey.com, 14 October 2026. Full article read. Source of the $2.52 trillion global pool, $310 billion for healthcare, the $960 billion US pool, the 30 percent cost scenario ($320 billion removed; global pool to $1.67 trillion, down 34 percent), the description of the scenario as "an illustration, not a forecast", the pharmacy technician result, and the recommendation to test at higher and lower cost. A proprietary McKinsey tool, with exhibits not reproduced in the page text. McKinsey sells AI cost management work.
- 2. McKinsey, "What does AI really cost?", McKinsey Explainers (video transcript, edited), mckinsey.com, 14 October 2026. Full text read. Experts: Lari Hämäläinen, Wasim Lala, Kaavini Takkar. Source of consumption pricing from the first quarter of 2026, the two-to-three-times cost effect, usage and spend doubling between months two and four, the bank moving from $3 million to more than $25 million, the "nearly quadruples" claim ("our research", unsourced), the almost 90 percent token price fall (described as commonly cited), cost per successful output, and the control-plane recommendation. Anecdotes are anonymised client work. The application to pharma cost of goods is my inference.
- 3. Vercel, AI Gateway Production Index (September edition), reported by IT Brief UK, 2 October 2026. Source of 56 percent of tokens and 14 percent of estimated spend for August, up from 13 percent in April. Secondary report, Vercel's own post not read. The October edition figures (71.5 percent of tokens, about 23 percent of spend) come via TLDR AI, 16 October 2026, and were not read at source. Vercel sells the gateway.
- 4. MacRumors, "Meta's New Mac Coding Agent Costs Up to 20 Times Less If You Let Meta Train on Your Data", 5 August 2026. Source of the two Muse Code tiers and their per-million-token prices. Meta's published terms not read.
- 5. Suzuki, R., "Making every stage of medical device document assessment with large language models deterministic", arXiv:2610.21300, submitted 15 October 2026. Abstract read; full paper not read. Source of fourteen deployments, twelve changing the compliance label, and the deterministic pipeline result. Not seen peer reviewed.
- 6. Brightbeam, Second Nature Intelligence, Week 42, 17 October 2026, and Week 41, 12 October 2026, via a copy pasted in by the author. Source of the leads for items rejected above. Not used as evidence.
- 7. Batch Record issue 08, 12 October 2026, for the Brightbeam trace promised there.
Applying McKinsey's cost findings to pharma business cases, and the cost-of-goods comparison, is my inference, not McKinsey's claim.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.