Pharma Does Not Get the Oversight Discount
McKinsey put human oversight at 70 to 75 percent of what it costs to run an AI agent. In every other sector that share falls as the models improve. In pharma it does not, because the requirement is written down.
By Mark Rhys Thomas
The Lead
Every pharma AI business case I have been shown this year was built on the price of tokens. Tokens keep getting cheaper, the argument runs, so the economics improve on their own. Be patient and the maths works.
Issue 04 said the first binding AI rule most pharma teams meet is written by a supplier rather than a regulator. This week the same suppliers published the figure showing that most of us are also metering the wrong thing.
McKinsey has now put numbers on what it actually costs to run an agent. For an agent handling a customer service task in banking, token costs are frequently 20 to 25 percent of the variable cost of a run. Human oversight is 70 to 75 percent. The people checking the output cost roughly three times what the model costs to produce it.
The firm is specific about who does that checking. Functional and risk experts. In customer onboarding, McKinsey expects 10 to 20 percent of agentic runs to be reviewed by those experts. Not the cheap reviewers. The expensive ones, the people you cannot hire quickly and cannot hire many of.
The absolute figures are worth holding as well. McKinsey's analysis puts a single-agent customer-facing workflow at some banks at $20,000 to $30,000 to run, and a multiagent team at $100,000 to $200,000. Multiply that by the number of workflows a large organisation would like to agentify, and the model licence stops being the interesting line in the budget.
Their recommendation follows directly from the split. Stop optimising model selection and token consumption, which are simply the most visible costs, and redesign the workflow to cut exception rates and simplify the reviews that require costly human intervention. Attack the oversight, not the invoice from the model provider.
Then comes the part no cross-industry report will translate for you.
In banking, that oversight ratio is a design choice. McKinsey says so plainly: for workflows that are highly deterministic, have limited variables, or have few compliance needs, and as AI agents become increasingly reliable, human oversight might be less necessary. The share falls as the technology improves. That is the implicit promise sitting inside every agent business case written this year.
Pharma does not get that promise. Look at where oversight actually sits in our workflows. MLR review. Medical information. Pharmacovigilance intake. Batch release. The human in those loops is not there because the model might be wrong. They are there because a procedure says a named, qualified person signs. That requirement does not soften as the model improves, because it was never a statement about the model in the first place.
So in pharma the 70 to 75 percent behaves as a floor rather than a ceiling.
One distinction matters here, because it is where the rebuttal will come from. McKinsey also shows volume working in your favour. Onboarding 2,500 customers a year costs $10,000 to $15,000, and doubling the customers takes it only to $15,000 to $20,000. That is fixed cost amortising, infrastructure and orchestration spread across more runs. Oversight is not fixed. It is variable and it attaches per run. Double the runs in a regulated workflow and you double the reviews, because the review requirement belongs to the run and not to the system. The fixed costs improve with scale. The regulated variable cost does not.
That is the Pilot Trap arriving in the accounts rather than in the technology. A pilot is cheap because a data scientist is watching it, and a data scientist is not a regulatory requirement. Production is expensive because a qualified person is, and by then there are thousands of runs. The pilot posts an attractive unit cost, the scaled version posts a terrible one, and everybody blames the integration.
Issue 04 said to meter consumption by use case now, while the numbers are still small enough to argue about. Refine that. Meter the oversight hours by use case, not the tokens. The tokens are the cheap quarter, and they are the only part anyone is currently counting.
A pharma AI business case with no line in it for the qualified person is not a business case. It is a token price with a hope attached.
Two Process issues in a row. That is not a rotation failure, it is where the evidence keeps landing.
Also on the Line
COMPUTE
The supplier opened a door this week. Issue 04 was about Anthropic's governance arriving as a setting in a cloud console. This week the same lab moved in the opposite direction. Anthropic reported using Claude to optimise more than thirty biomolecular models, the software that predicts how a protein folds and what binds to it, achieving a fourfold speed increase and a low memory mode that runs larger systems on a single GPU. It is open sourced. That capability sits outside the Covered Model perimeter entirely, because what you receive is code you run on your own hardware. Vendor research, so treat the multiple with care. The direction is the point.
DATA
Scale lost to laboratory data. Periodic Labs released Neon, trained on real experimental data rather than general internet text. It beat GPT-6 Astra and Claude Fable 5.1 on FrontierXRD, a hard benchmark for scientific analysis, at a lower cost per analysis, and it is already deployed in labs analysing superconductor and magnet experiments. A smaller model with better inputs beat two frontier models. Issue 04 asked you to label predicted data before it spreads. This is the other half of that: pharma holds the measured data and will not route it anywhere a model could learn from it. That is a governance decision being taken by default rather than on the merits.
PROCESS
Somebody started building the certificate. The Artificial Intelligence Underwriting Company, founded by an early Anthropic hire and a former chief operating officer of METR, launched a third-party audit and certification layer for AI agents. Agents run the tests, AI analyses the results, humans verify the final audit. This is the independent assurance a regulated buyer has so far had no way to obtain, which is why every agent conversation with a quality function stalls in the same place. Whether it survives contact with an inspector is unknown. That it is being built at all says where the accountability question has moved.
What Did Not Meet Spec
Thirty-one newsletter emails came through this week's pull, across five consultancies and two TLDR editions, with three research hubs checked directly. Six emails were read in full and eleven claims were checked against primary sources. Rejected by category: every sponsored placement across ten TLDR issues, more than a dozen of them, each written to read as editorial; the general edition's standing diet of rocketry, robotaxis, consumer hardware and developer career essays; and model tooling with no strategic signal.
Two named rejections.
The 6 percent, and the 37 percent I have already used. Issue 04 carried McKinsey's finding that 37 percent of companies report any contribution to earnings from AI. The CEO Shortlist of 18 September repeats that 37 percent in its editorial text, then states in an event blurb four paragraphs below that only 6 percent of companies see significant EBIT impact, while separately describing 6 percent as the share of firms that redesign workflows rather than adding AI to existing ones. Those cannot all be the same measurement. I used the 37 percent a week ago in good faith. I am not using either figure again until the survey text settles which is which, and issue 04's citation should be read with that caveat attached.
Accenture's $27 trillion. Accenture's claim that $27 trillion is moving to companies with “Value Leader DNA”. No publication date on the hub, no published method, and the firm sells the conclusion that reinvention pays. A number that large needs to arrive with its working.
The Line
The model got cheaper. The signature did not.
Sources
- 1. McKinsey QuantumBlack, “Where AI agents pay off: A practical guide to the economics of agentic workflows”, 24 August 2026. Chandana Asif, Dieter Kiewell, Tunde Olanrewaju, Lari Hämeläinen, Tom Kolaja. Full article read. Source of the 20 to 25 percent token share, the 70 to 75 percent oversight share, the 10 to 20 percent expected review rate in customer onboarding, the $20,000 to $30,000 single-agent and $100,000 to $200,000 multiagent run costs, the 2,500 customer amortisation example, and the statement that oversight may become less necessary in deterministic or low-compliance workflows. Figures are described as McKinsey's own analysis; method not published.
- 2. McKinsey, The CEO Shortlist, “From productivity to returns”, 18 September 2026, curated by Alex Panas and Becca Coggins. The email that surfaced the above and that carries the 37 percent and 6 percent figures rejected in this issue.
- 3. Anthropic, “How Claude is uplifting biomolecular modeling”, research post, 18 September 2026, reported via TLDR AI the same day. Vendor research on its own model. The fourfold figure is Anthropic's own and is not independently verified.
- 4. Periodic Labs, “Nature is our learning environment”, 16 September 2026, reported via TLDR AI. Vendor announcement. The FrontierXRD comparison against GPT-6 Astra and Claude Fable 5.1 is the company's own and is not independently verified.
- 5. The Artificial Intelligence Underwriting Company, reported by TechCrunch, 15 September 2026, via TLDR AI. Press reporting, not company material, which makes this the most independent item in Also on the Line.
- 6. Accenture, “Growth in the age of AI”, accenture.com insights hub, accessed 21 September 2026. Undated on the hub. Rejected in this issue.
The pharma application of source 1 is mine, not McKinsey's. Their measurements are taken from banking workflows. The claim that the oversight share behaves as a floor in regulated pharma workflows is my inference from how MLR, medical information, pharmacovigilance and batch release actually work, and it should be read as inference rather than as a finding.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.