Our reading: agents are earning trust where errors are reversible and detectable. Evidence, FDA, EMA and EU rules, risks, and three NKK models for bounded autonomy in plants, planning and GxP.
Written for: COOs and plant heads, supply chain leaders, quality and regulatory heads in life sciences, CIOs, OT security leads, heads of AI
12 chapters, 57 cited sources, 27 models and charts. The first 2 chapters are open; the rest unlock with the form.
of responses to one survey question reported AI tools in inspected cGMP or GDP use (no per-item n)
of agentic AI projects forecast to be cancelled by end-2027 (Gartner prediction, not an observed rate)
Our reading of the evidence is that autonomy in industrial and life-sciences operations is advancing in the order of reversibility, not capability: agents earn trust where a wrong action is cheap to undo and easy to notice. The support is indirect (regulator caution, WEF's wording on autonomy and FDA's own propose-then-verify use of agents), and no deployment data by action type exists. The binding constraint is whether anyone can show, for a specific action, what the agent may do, how an error would be caught and who is accountable.
The evidence base is thinner than the adoption narrative. Among the World Economic Forum's Lighthouse leader factories, AI agents enabled 5% of the top-five solutions in 2025, and WEF writes that fully autonomous agentic AI "remains theoretical in the production context" [1]. We found no audited, cross-company evidence of return on agentic AI in manufacturing, supply chain or life sciences.
Life sciences has the most explicit rulebook and the one enforcement signal we found: in April 2026 FDA applied existing CGMP rules to records drafted with AI agents, stating that failing to review AI-generated documents violates 21 CFR 211.22(c) [3]. The EU's draft Annex 22 still bars generative and adaptive models from critical GMP uses; EMA's October 2026 workshop report shows a risk-based path under discussion, with no revised text yet [4][5].
This report offers three NKK models and a readiness scorecard. It is a synthesis of public research; it does not contain primary survey data.
of responses to one survey question reported AI tools in inspected cGMP or GDP use (no per-item n)
of agentic AI projects forecast to be cancelled by end-2027 (Gartner prediction, not an observed rate)
Method and limits. Every figure here traces to a numbered source that was opened in October 2026. Four gaps matter. McKinsey's own pages could not be opened, so McKinsey figures appear only as quoted inside the WEF paper, which WEF produced with McKinsey [1]. The IEC 62443 standards are paywalled, so we use only the ISA overview [7]. The ISPE GAMP guide on AI is a paid publication [40], and the full OWASP Top 10 for Agentic Applications sits behind a gated download [30]; we read neither. NKK's models are opinion built on the cited prior art, and each states where it breaks.
Definitions diverge by body, and the divergence changes what a rule or a statistic means. FDA defines agentic AI as "advanced artificial intelligence systems designed to achieve specific goals by planning, reasoning, and executing multi-step actions" [8]. The joint guidance from CISA and eight partner agencies covers machine learning, LLM-based AI and AI agents under one umbrella, and applies to statistical modelling and logic-based automation too [9]. The EU's draft Annex 22 uses the AI Act definition, but its operative scope is static, deterministic models, which is not what most people mean by an agent [4]. Gartner estimates that only about 130 of the thousands of vendors marketing agentic AI are real, which it calls "agent washing" [6].
MIT Sloan Management Review and BCG add a practical warning: vendors are embedding agentic features in products, "causing organizations to implement agentic AI before they have developed a strategic management framework" [10]. In a plant, that means an agent can arrive through a software update rather than a project.
For control purposes we separate three things.
| Criterion | Analytical model | Copilot or assistant | Tool-using agent |
|---|---|---|---|
| Acts on its own? | No. Produces a score or class that a fixed process consumes | No. Drafts or answers; a person takes the action | Yes. Plans, calls tools and changes the state of a system |
| Manufacturing example | A vision classifier that flags defects (NKK archetype) | Code-generation help in an engineering tool [11] | A welding-programming agent that cut robot programming from 16 hours to 1 hour at one site [1] |
| Supply chain example | A demand or lead-time forecast (NKK archetype) | An LLM front end to an optimiser for what-if questions [12] | Seven agents that read news, map supplier exposure and recommend mitigations, tested on synthetic scenarios [13] |
| Life sciences example | EMA-qualified AIM-NASH, a locked model for scoring liver biopsies in trials [14] | Drafting a deviation report for QA review (NKK archetype) | An agentic deviation-management system, presented at EMA's workshop; some examples were pilots or theoretical frameworks [5] |
| Main control anchor | Annex 22 draft: static and deterministic [4]; FDA credibility framework [15] | Human review under existing rules, e.g. 21 CFR 211.22(c) [3] | CISA principles for OT [9]; unique accounts under Annex 11 draft [16] |
Every adoption figure answers a different question, which is why they cannot sit on one slide as a trend.
MIT SMR and BCG surveyed 2,102 respondents in spring 2025 and reported that agentic AI "has already reached 35% adoption, with another 44% of organizations planning to deploy it soon" [10]. That is self-reported, with a broad definition. Gartner's January 2025 poll of 3,412 webinar attendees found 19% with significant investment, 42% conservative, 8% none and 31% in wait-and-see mode, from a self-selected audience [6].
Narrow the question to production and the numbers fall. In the Lighthouse analysis, agents enabled 5% of top-five solutions in 2025 [1]. Gartner puts the share of enterprises using supply chain management software that had adopted agentic features at 5% in 2025, forecast to reach 60% by 2030 [17]. In a 2026 industry survey on digital and AI-enabled models in pharmaceutical development and manufacturing, 68% of respondents' companies had experience of AI tools, 29% reported AI tools in regulatory submissions and 6% in cGMP or GDP environments subject to inspection [2]. The survey had 26 responses in total, but the paper gives no per-item count, so the 6% may rest on only one or two respondents. The authors ask readers to treat the results as a "snapshot rather than as a statistically representative assessment" [2].
| Criterion | Who was asked, what was counted | Sample | How to read it |
|---|---|---|---|
| 35% adopting agentic AI [10] | Executives; self-reported adoption, broad definition, spring 2025 | 2,102 respondents, 21+ industries | Direction of travel, not production use |
| 19% significant investment [6] | Webinar attendees, January 2025 | 3,412, self-selected | Interest among the already engaged |
| 5% of top-five solutions [1] | Share of use cases at designated leader sites, updated classification | 223 sites in the network | Indicative share of headline solutions at leader sites; not a measure of all production use, since agent use outside the top five is not counted |
| 5% of SCM-software enterprises [17] | Enterprises using SCM software that had adopted agentic features, 2025 | Analyst estimate | Baseline for a 60% forecast by 2030 |
| 6% in cGMP or GDP use [2] | Responses reporting AI tools in inspected environments | 26 responses in total; no per-item n | Indicative only; the authors say not representative |
Gartner also forecasts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024 [6]. It is an analyst forecast, and it counts day-to-day work decisions, not regulated or safety-relevant ones. We would plan against the single-digit observed figures, not the forecast.
Every figure and regulatory statement in this piece traces to a numbered source below. Links open the original publisher in a new tab.
A vendor-neutral reference architecture for enterprise agents: NKK's nine-layer stack, a pattern-selection matrix, a blast-radius gate, and the evidence on reliability, state, control and cost.
Prompt injection remains a residual risk, so agent security is a privilege-and-containment problem. Evidence, controls and three NKK models: a Threat-to-Control Map, Autonomy Tiers and an Agent Identity Stack.
Official statistics, central-bank surveys and consultancy research on AI adoption, agents, spend, value and jobs, reconciled, with three NKK models for reading the numbers.
