Drapeau’s AI Deep Dive: The Price of Acceptance in Regulated AI

Artificial intelligence (AI) vendors sell generation. GMP organizations must fund acceptance. AI can generate candidates quickly. Credibility still has to be built, and someone has to pay for it.

Published evidence shows generation getting much cheaper at fixed performance levels. The Stanford Institute for Human-Centered Artificial Intelligence 2025 AI Index reports a more than 280-fold reduction in the price of querying a model that scores at GPT-3.5 level on the Massive Multitask Language Understanding benchmark, from $20.00 to $0.07 per million tokens between November 2022 and October 2024.¹ Artificial Analysis reports that output throughput increased substantially across model classes in the second quarter of 2025. It also notes that reasoning and agentic workflows can generate enough additional tokens to offset those speed gains end to end.² None of those figures show whether a result can be used.

Six Gates, Six Kinds of Evidence

A pharmaceutical program passes through distinct stages, each producing a different kind of evidence.

• Discovery produces candidates, predictions, and hypotheses. A predicted candidate is not a validated result.
• Development produces experimental data and process understanding. Wet-lab confirmation happens here, and a model output cannot replace it.
• Validation and qualification show that a method, system, or process performs as intended for its defined use.
• Manufacturing produces batches and records under controlled procedures. A model output is not a batch record.
• Quality disposition releases, rejects, or escalates a batch, record, or decision. An authorized role accepts the output at this gate.
• Regulatory acceptance is the agency’s review of the evidence package.

A model can accelerate the first gate. It cannot carry a program past a later gate without evidence generated at that gate.

What Makes an Output Credible

Each gate rests on controls that let another person rely on a claim.

• Provenance records where the input data came from and how it was processed.
• Intended use defines the decisions and records the output may support. That definition sets the review depth.
• Acceptance criteria are fixed before the output exists, so the output cannot shape them.
• Version control records the model, configuration, prompts, sources, and procedures behind each output.
• Model-performance monitoring shows whether the system still performs as it did during qualification, and sets the thresholds that trigger review.
• Change control determines whether a model, source, or configuration change alters system behavior, and what testing follows.
• Human scientific review applies the judgment a regulated decision requires. The reviewer must be able to explain it.
• Independent verification checks consequential outputs with a person or method that did not produce them.

Each control adds cost and creates the evidence that lets a reviewer, auditor, or regulator trust the result.

McKinsey Identified the Cost Center

McKinsey’s August 2026 guide to agentic-workflow economics gives illustrative figures for a banking customer-service agent.³ Human oversight accounted for 70% to 75% of variable run cost, and tokens for 20% to 25%. In banking customer onboarding, McKinsey expects 10% to 20% of agentic runs to be reviewed by risk and functional experts.

McKinsey is also explicit that oversight can shrink. It stated that 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.” It also advised leaders to “simplify review processes that require costly human interventions.”³ GMP work tests poorly against those conditions. Outputs feed decisions that quality must accept, and the inputs are often variable and document-based. The compliance burden is the reason review exists. Simplification is possible, but only with evidence, as is discussed below.

In McKinsey’s banking example, oversight accounts for most of the variable cost. These examples are illustrative and not drawn from pharmaceutical manufacturing, so their percentages should not move into a GMP business case unchanged. The cost structure still matters: when human judgment is expensive, even selective review can dominate operating cost.

Inside a quality system, review frequency cannot be set by cost optimization alone. It has to reflect intended use, error consequence, control strength, system evidence, and the authority required to approve the action. GMP sets the terms under which acceptance can occur.

FDA Has Already Made the Acceptance Boundary Visible

FDA’s April 2026 warning letter to Purolea Cosmetics Lab described the firm’s use of AI agents to create drug-product specifications, procedures, and master production or control records.⁴ FDA stated that the firm needed to review AI-generated documents for accuracy and actual current GMP (cGMP) compliance. The letter also stated that if the firm resumed drug production and used AI for cGMP activities, an authorized human representative of the quality unit would have to review and clear the AI outputs or recommendations.

That statement applied to one firm in one enforcement setting. It does not establish a universal requirement for 100% human review of AI-supported GMP transactions. It does mark a boundary that weak business cases ignore: an AI-generated document must be reviewed for accuracy and cGMP compliance before anyone relies on it.

The letter does not define what that review must cover. Read in that light, the letter suggests that a completed output may still require a qualified person to confirm the source information, the supportability of the conclusion, compliance with applicable requirements, and that the action falls within authorized limits. If the surrounding control system cannot show those conditions, the cost of acceptance returns to human review.

What Acceptance Actually Includes

The phrase “human in the loop” covers several different activities. A person may confirm the records used, detect missing evidence, evaluate reasoning, correct or reject an output, escalate uncertain cases, and authorize the action. Each requires different qualifications. A document coordinator checking format does not do the work of a microbiologist assessing contamination risk. A serious cost model names the acceptance task and the role authorized to perform it (Table 1).

The Correct Denominator

Cost per generated output and minutes saved per task are easy to calculate, but they miss most of a regulated workflow’s economics. The more complete measure is the fully loaded cost per accepted outcome (Figure). It divides total workflow cost, including generation, qualification, review, correction, exceptions, monitoring, change control, failure management, and continuity, by the number of outcomes formally accepted for their intended use.

An outcome is acceptable when it meets the defined acceptance criteria. It becomes accepted when an authorized role completes the required review and records the disposition.

Rejected outputs still consume resources, and corrections carry rework. A workflow can produce cheap drafts and expensive outcomes, and it can cut preparation time while raising the total cost of closure.

Acceptance Yield Measures the Performance of the Acceptance Process

An earlier analysis by the author showed that qualified review can dominate the cost of an AI-supported GMP workflow.⁵ The next step is measuring the acceptance process itself. The 7 measures in Table 2 are proposed operational metrics, not regulatory requirements. Define each for the intended use and control design of the workflow.

A Hypothetical Worked Example

Consider a workflow piloted at 100 outputs and then expanded to 1000 (Table 3). The values are illustrative, not measured industry performance. They include generation and qualified review labor only. Qualified review costs $120 per hour. Routine review takes 10 minutes, material correction 40, escalation 120, and rejection handling 90.

Review Is a Control Decision

A review rate of 10%, 20%, or 100% is a claim about how reliable the remaining control system is. Reducing review can be defensible only with evidence that the intended use is narrow, inputs and configuration are controlled, independent checks catch material errors, exceptions route reliably, and monitoring detects deterioration. The organization also has to be able to return to fuller review when needed. Lighter human review requires stronger automated controls, better evidence capture, and more capable monitoring. Review labor may fall while control cost rises, so the business case should compare complete control strategies.

Section 211.22 of 21 CFR gives the quality control unit authority to approve or reject components, containers, labeling, and drug products, and requires it to review production records for errors and their full investigation.⁶ Qualified quality-unit capacity can become the limiting resource. It takes time to build, cannot be expanded overnight, and is shared with investigations, batch review, change control, and inspections. A model should test capacity at both expected and peak volume.

Exceptions and Change Drive the Long-Run Cost

Average handling time hides the difficult cases, and those cases set the cost. A workflow that succeeds in 95% of cases may still be uneconomic if the remaining 5% consume senior quality capacity, delay disposition, or trigger broad retrospective review.

A model update, prompt change, retrieval modification, new data source, or supplier release can change system behavior. The acceptance cost measured in a pilot therefore does not stay fixed. The January 2026 FDA and European Medicines Agency guiding principles call for risk-based oversight (Principle 2), documented data provenance (Principle 6), and scheduled monitoring with periodic re-evaluation across the AI life cycle (Principle 9).⁷ The National Institute of Standards and Technology reports that attendees cited rapid releases from vendors, and it lists scaling human-driven monitoring alongside rapid rollouts as a barrier to monitoring deployed AI systems.⁸

A business case should define which changes trigger reassessment, what evidence allows review to decrease again, and who can suspend the system. Before approving a regulated AI use case, finance, operations, technology, and quality should answer the same questions: the intended use, the acceptable outcome and the role that accepts it, the review depth assumed and its support, the fully loaded cost per accepted outcome, the cost after a model change, and whether the system state and human decision can be reconstructed later. A business case that cannot answer these has priced generation and left acceptance unfunded.

What a Defensible System Requires

Pharma should neither reject AI nor scale it blindly. In my experience, the organizations that benefit most are those that build the strongest path from output to evidence. That path has concrete parts. Define the intended use before deployment, and set acceptance criteria before the first output is reviewed. Record provenance, model version, configuration, prompts, and sources for every output, so any decision can be reconstructed. Set monitoring thresholds that trigger review. Put model and source changes under change control with defined testing. Keep independent verification for consequential outputs. Fund qualified acceptance capacity in the business case.

These systems cost more to launch than a demonstration. They are also the ones that hold up when an inspector, auditor, or regulator asks how a result was produced and who accepted it.

The Two Economic Mistakes

The first mistake is measuring production while ignoring acceptance. McKinsey’s example made the oversight cost visible. GMP determines how much oversight, evidence, and authority the workflow requires. Tokens have a market price. Qualified acceptance has an evidence requirement. The business case must price an outcome the organization is prepared to use, defend, and act on.

The second mistake is assuming acceptance demand stays flat while generation gets cheaper. That assumption has to be stated and tested. Cheaper generation encourages more outputs, more use cases, and more AI-supported decisions, so demand for qualified acceptance can outrun the capacity to provide it. At that point unit economics become system economics, and regulated AI meets its own version of Jevons paradox.⁹ Jevons argued in The Coal Question (1865) that increased efficiency in coal use would not reduce demand. It would spur greater use and further deplete supplies.

References

1. Stanford Institute for Human-Centered Artificial Intelligence. Artificial intelligence index report 2025, chapter 1: research and development. Stanford HAI. Accessed November 2, 2026. https://hai.stanford.edu/assets/files/hai_ai-index-report-2025_chapter1_final.pdf
2. Artificial Analysis. State of AI: Q2 2025 highlights report. Accessed November 2, 2026. https://artificialanalysiscdn.com/public-reports/state-of-ai-2025-q2-highlights-artificial-analysis.pdf
3. Asif C, Kiewell D, Hämäläinen L, Kolaja T, Olanrewaju T. Where AI agents pay off: a practical guide to the economics of agentic workflows. McKinsey & Company. August 2026. Accessed November 2, 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/where-AI-agents-pay-off-a-practical-guide-to-the-economics-of-agentic-workflows
4. FDA. Purolea Cosmetics Lab: warning letter 320-26-58. FDA.gov. April 2, 2026. Accessed November 2, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/purolea-cosmetics-lab-722591-04022026
5. Drapeau B. McKinsey priced the oversight. GMP sets its terms. LinkedIn. October 27, 2026. Accessed November 2, 2026. https://www.linkedin.com/pulse/mckinsey-priced-oversight-gmp-sets-its-terms-brian-drapeau-82mee/
6. 21 CFR §211.22. Responsibilities of quality control unit. eCFR.gov. Accessed November 2, 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-C/part-211/subpart-B/section-211.22
7. FDA, European Medicines Agency. Guiding principles of good AI practice in drug development. FDA.gov. January 2026. Accessed November 2, 2026. https://www.fda.gov/media/189581/download
8. Rao A, Keller D, Kalra N, et al. Challenges to the monitoring of deployed AI systems. NIST AI 800-804. National Institute of Standards and Technology; 2026. doi:10.6028/NIST.AI.800-4
9. Jevons WS. The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal Mines. Macmillan; 1865. https://oll.libertyfund.org/titles/jevons-the-coal-question.