The number at the top of the business case is usually a market forecast. In the material circulating in 2026, the size of “AI in pharma” in 2025 is $3.8 billion, $3.07 billion, and $1.94 billion. By 2030, 2035, and 2034 it will be $15.2 billion, $48.43 billion, and $16.49 billion.

Those are not three scenarios. They are three publishers measuring three different things, and they get quoted as one series.

That is the pattern in this layer, and it is worse than the pattern in the layers below it. In molecular design, virtual cells, biomarker discovery and clinical trial operations, the capability claims verify and the magnitude claims do not — and the magnitudes were at least describing science. Here, the magnitude is the argument. A board does not approve a validation budget without a forecast, an ROI model and a case study. When those three things are the softest evidence in the stack, the softness sits directly on the release valve for capital.

So this is what the strategic layer looks like when you take it apart: what verifies, what is drifted, and what has no source at all.

What releases capital, and what backs it

┌────────────────────────────────────────────────────────────────┐
│  EVIDENCE TIER INSIDE THE AI BUSINESS CASE                     │
├────────────────────────────────────────────────────────────────┤
│  Market forecast        three publishers, three definitions    │
│  Deal headlines         real; the money is "up to"             │
│  Capital markets        verifiable; rounds dated a year off    │
│  ROI case studies       no named company in either             │
│  Productivity gains     one commentary, one unreadable report  │
│  Budget allocation      no source                              │
└────────────────────────────────────────────────────────────────┘

The forecast problem: three publishers, three definitions

Every one of the three forecasts is internally sound. None is comparable to the others, and stacking them produces a growth curve that does not exist.

Publisher 2025 Horizon CAGR What it actually measures
BCC Research (AIT020A, Sep 2025) $3.8B est. ($3.0B = 2024 base) $15.2B by 2030 31.7% Core AI in pharma; software-dominant; North America 59.5% share (2024)
SNS Insider (#7678) $3.07B $48.43B by 2035 31.76% Extended — software platforms plus AI-as-a-service
Precedence Research (#1485) $1.94B $16.49B by 2034 27% Narrower definition; the publisher has since revised its own figure to $18.99B by 2035 at 25.62%

Two things follow. First, the $48.43B number and the $16.49B number are not the same market; one includes services revenue that the other excludes. Second, one publisher has already revised its own estimate downward on its own page, which is what a forecast looks like when it is a model and not a measurement.

There is also a live conflation risk worth knowing about: $3.8 billion in 2025 is simultaneously a figure for AI in pharma (BCC) and for AI in clinical trials (a different publisher, different segment). Identical number, two markets. That is how a figure acquires authority it did not earn — it gets repeated near itself.

And then there is the one that simply fails arithmetic. A widely circulated claim states global AI-pharma funding reached $50.9 billion in 2025, up from $272 million in 2019, with cumulative 2019–2025 investment exceeding $27.2 billion. A single year cannot exceed the cumulative total that contains it. No source for the $50.9B exists; the claim contradicts McKinsey’s AI drug-discovery funding figure for 2025 ($8.4B), PitchBook’s ($2.7B through Q3 2025) and GlobalData’s ($3B). Do not repeat it.

The ROI numbers are real — and the sample is not what you think

This layer has genuinely strong survey evidence, and it is worth using precisely, because the caveats change what it means.

PwC Strategy& modelled, across more than 200 use cases, that AI could double pharma’s operating profits by 2030 — an additional $254 billion in annual operating profit worldwide, distributed 39% operations, 26% R&D, 24% commercial, 11% enabling functions. Those figures verify exactly. They are also a model, and the distribution is of modelled indirect value, not observed savings.

Deloitte surveyed 1,854 executives (August–September 2025) with a set of findings that verify verbatim:

  • Most report satisfactory ROI within two to four years; only 6% report payback in under a year.
  • 10% currently realise significant ROI from agentic AI.
  • 15% report significant measurable ROI from generative AI; 38% expect it within a year.
  • 85% increased AI investment in the past twelve months; 91% plan to increase again.

The caveat is material and almost never quoted: the sample is Europe and the Middle East only, and only organisations with live AI deployments. It is a good study of a specific population. It is not a global benchmark, and a pharma business case that cites “2–4 years to payback” as an industry-wide expectation is citing a regional self-selected cohort.

BCG’s contribution is the 10-20-70 split — 10% algorithms, 20% technology and data, 70% people and process. It verifies as cited. It is also a heuristic, not a measured share, and it is routinely quoted as though someone had measured it.

PitchBook does support the valuation claim: AI-native biotech companies carry a nearly 100% premium over biopharma valuations more broadly. That one is worth keeping, because it is a market observation rather than a model.

Two ROI case studies, zero named companies

Here is the part that should worry anyone assembling an investment case.

The first is a decision-rights case study, quoted in the material as a mid-cap European biopharma that rewrote supply-planner decision rights and achieved:

  • 41% reduction in deviation investigation time
  • 37% reduction in regulatory response authoring cycles
  • 18% improvement in right-first-time
  • 21% inventory reduction

There is no company. There is no analyst. There is no publisher. Search as long as you like: the case study does not resolve to a source. It is presented as a named case study with the name removed, which in practice means it cannot be checked, sampled, or rejected — it can only be repeated.

The second is a vendor ROI bundle attributed to AI-assisted medical/legal/regulatory (MLR) review integrated with existing Veeva infrastructure, claimed to deliver 40% faster review cycles, 50–70% shorter draft-to-distribution timelines, 58% time savings and over 50% cost reduction on regulatory dossier creation, and over $1M in ROI with 800% returns on onboarding. No vendor page carries these numbers, and several of them contradict the vendor’s own published figures — Veeva’s public claims are “up to 75% cycle-time reduction” and roughly 20% meeting-time reduction at two FTE. When a vendor’s third-party summary overstates the vendor’s own marketing, that is not a case study. It is an artefact.

Two related figures do trace, and they trace to a consultancy rather than to the industry. A ~5,000% first-year ROI on a European pharma marketing-mix-modelling project and a 600% ROI on “automating routine tasks” both come from Artefact. The 5,000% is what it says. The 600% is a distortion: its origin is one Belgian hospital medical-coding automation case, generalised in secondary write-ups from a single deployment to “routine tasks.” One case, relabelled as a category.

What actually verifies: the capital

If you strip out the forecast layer, the capital-markets evidence is the strongest material in this entire layer — and it points somewhere specific.

The deal sizes are real, and the wording matters. Eli Lilly and NVIDIA committed up to $1 billion over five years to a Bay Area co-innovation lab (January 2026). Merck committed up to $1 billion across a multi-year partnership with Google Cloud (22 April 2026), deploying Gemini Enterprise across R&D, manufacturing, commercial and corporate functions for 75,000 employees. Novo Nordisk signed three separate deals in 2026 — OpenAI in April, AWS as preferred cloud and AI partner in August, and Anthropic/Claude Science in September. “Up to” is doing real work in all of them: these are milestone-inclusive ceilings, not disbursements.

The structure is back-loaded, and that is the strategic signal. Across the AI-native partnership deals, upfront payments run a few percent of the milestone-inclusive total. Lilly’s March 2026 extension with Insilico Medicine is up to $2.75 billion against roughly $115M upfront. Risk stays with the platform until the platform can show something.

External innovation is the business, which is why the scouting layer is where AI compounds. More than 70% of new-molecular-entity revenues since 2018 came from externally sourced products. Life-sciences deal value reached $372 billion in 2025, up 47% year over year. 120 AI-based target and drug-discovery deals were signed in 2025 through October — 23% of the 513 such deals recorded since 2017. If seven-tenths of your revenue comes from outside, the capability that matters is finding, diligenced and pricing external assets earlier than the next company, and that is a decision-funnel problem rather than a chemistry problem.

And the capital markets did price a pure-play. Insilico Medicine listed on the Hong Kong Stock Exchange Main Board in December 2025 under Chapter 8.05, raising HK$2.277 billion — the largest biotech IPO in Hong Kong that year — with Lilly, Tencent, Temasek, Schroders and UBS AM as cornerstone investors, and the public tranche oversubscribed roughly 1,427 times.

Two capital-markets claims in the same material are simply wrong, and both are the kind that would survive an internal review because they sound right. “Q2 2025 pharma venture funding hit about $21.4 billion, up roughly 20% year on year” is false: JPMorgan recorded $4.5 billion, the lowest quarterly figure in five years, with GlobalData showing a 34.8% quarter-on-quarter decline. And “global biotech funding reached $25.1 billion in 2024, up 5.5%” is a sector mislabel: those are digital health figures from Galen Growth. Biopharma venture funding in 2024 was approximately $26.0 billion. Same year, different industry, forty-one percent apart in the retelling.

What actually verifies: the operating-model change is narrow and named

The genuinely interesting strategic evidence is not the market forecasts. It is a short list of named companies that have moved AI into a decision path.

Roche reported at its 28 September 2026 Pharma Day that its Target Nexus tool contributed to 40% of pipeline decisions from Q4 2025 through Q2 2026, and that it expects to reach 80% of research portfolio decisions by end-2026. Roche’s own deck also states a Phase III success rate of >80% year-to-date 2026, against 65% in 2025, and plans to reallocate approximately CHF 2 billion into AI and productivity programmes (wording: “set to be reallocated,” alongside roughly CHF 1.6bn reinvested since end-2023).

Read that precisely. The 40% is described as AI/computational contribution broadly, not Target Nexus specifically; the 80% is a target; and the Phase III figure is company-reported with no methodology or trial list attached. Every one of those caveats is fair — and Roche is still the clearest instance of an AI system sitting inside a portfolio-decision process rather than beside it.

Sanofi’s Plai verifies verbatim against Sanofi’s own February 2026 material: an agent aggregating more than one billion data points, modelling R&D cost, clinical enrolment timelines and probability of success, with what-if scenarios incorporated into governance discussions and participation in governance meetings. “Formal seat” is Sanofi’s own metaphor, not a governance instrument, but the underlying claim — that the scenario model is in the room — is sourced.

Genentech exercised its first Validated Target Option milestone on an AI-discovered neuroscience target under its Recursion collaboration. The payment was $3 million.

That last figure is the most useful number in this section. A milestone payment is the market’s own price on an AI-derived discovery at that stage, and it is small. It is a better calibration anchor for a business case than any productivity percentage, because it is a transaction rather than an opinion.

The failure mode is drift, not fabrication

Across this whole layer, only two claims are outright fabricated. Everything else fails in a single recognisable way: attribution drift — the citation chain is intact, and the claim is no longer the claim.

  • Sector drift. Digital-health funding figures relabelled as biotech.
  • Vintage drift. Sanofi’s “$800 million reallocated on AI advice” is a 2023 event reported in November 2024, not a 2026 one. Isomorphic Labs’ $2.1 billion round is May 2026, not May 2025 — it is indeed the second-largest biotech round ever, just not in the year claimed. The Recursion–Exscientia merger is $688 million, agreed and closed in 2024, not “$700M in 2025.”
  • Actor drift. The first AI-designed drug submitted to FDA for bile duct cancer was submitted by Elevar, not Relay Therapeutics — and it was approved on 23 September 2026. The 47–49% (data operations and commercial) versus 17–30% (R&D and clinical) value split is ZS 2026 CDIO Research (n=115), not Define Ventures and not Deloitte. A “$200M GSK–Pathos upfront” does not exist; the $200M belongs to the AstraZeneca–Tempus–Pathos arrangement, and GSK’s comparable deal was a different company at $110M.
  • Descriptor drift. “LillyPod” is an AI supercomputer, not an AI-driven vaccine factory. GSK’s global data team is Onyx, not “dataOnyx.” Fosun’s PharmAID is described in its own source as China’s first self-developed AI decision intelligence platform in the industry — a narrower and more accurate claim than “China’s first AI decision agent in healthcare” — and its 50% improvement is content-generation accuracy against general-purpose LLMs, with no market-modelling or patent-analysis modules described.
  • Definition drift. Three forecasts, three scopes, one series.

This is the same failure class as the staleness documented in the clinical-trials layer, where a 2008 oncology benchmark was quoted as current with the citation chain fully intact. Fabrication is loud and rare. Drift is quiet and everywhere — and drift is harder to catch, because everything about the claim looks authoritative except its relationship to the present.

Two more claims survive verification only as attribution you cannot read at source. McKinsey’s published material does carry the finding that 75–85% of pharma workflows contain tasks that can be automated or augmented by AI agents. The companion figures — 25–40% of organisational capacity freed, +5–13 percentage points of revenue growth, +3.4–5.4 points of EBITDA, up to 95% of roles interacting with agents as teammates — rest on a single trade-press commentary (PharmExec, September 2026) attributing them to a McKinsey report that cannot be extracted from McKinsey’s own site. Gartner’s “by 2027, half of all business decisions will be augmented or automated by AI agents” is a real, dated press release — with the popular phrasing quietly dropping its “for decision intelligence” scope. And Kiin Bio’s “2–3 week workflow in under two hours” and the Tater agent’s “1–4 months of qPCR assay design in under two hours” are both real as vendor-reported case studies in a 2026 review paper, from a vendor called Happy Potato.

The business-case checklist

If you are assembling or reviewing a pharma AI investment case this quarter:

  1. Never stack forecasts across publishers. Pick one definition and defend it. If your model blends a core-software forecast with an AI-as-a-service forecast, you have invented a market.
  2. Name the case study or drop it. A mid-cap European biopharma with no name cannot be diligenced. If the source will not name the company, the number is not evidence — it is a plausible shape.
  3. Check the case study against the vendor’s own page. The MLR bundle quoted in circulation overstates the vendor’s own published figures. If the summary is better than the marketing, the summary is wrong.
  4. Quote the sample, not just the statistic. “2–4 years to payback” is Europe and the Middle East, live-AI organisations only. Say so, or do not use it.
  5. Price the milestone, not the percentage. Genentech’s first AI-discovered-target milestone was $3 million. A transaction beats a productivity estimate as a calibration anchor, every time.
  6. Separate “any improvement” from “at scale.” Deloitte: 45% report some measurable improvement, 13% at scale. McKinsey: 32% scaling, 5% realising significant financial value. Tufts CSDD: 10.7% fully implemented across clinical activities. Roughly one in ten is the honest base rate for the business case.
  7. Watch for drift, not lies. Check the year on every deal, the sector on every funding figure, and the actor on every first. Almost nothing here is fabricated. A great deal of it has moved.
  8. Label your own self-reports. Most of the numbers you inherit will be vendor claims or company keynotes — including Roche’s Phase III rate and Sanofi’s service-level gain. Using them is fine. Using them unlabelled is not.

The bottom line

The strategic layer of pharma AI is where the industry’s ambition is highest and its evidence is thinnest. The capital is real: two $1B hyperscaler commitments, a Hong Kong listing oversubscribed 1,427 times, $372 billion of life-sciences deal value. The operating-model change is real and narrow, and it is happening at companies you can name. The strategic logic — that more than seven-tenths of new-molecule revenue comes from outside, so the external-innovation funnel is the business — is the most defensible argument in the entire stack.

What is not real is the scaffolding around it. Three forecasts measuring three different things, quoted as one series. Two ROI case studies with no company attached. A funding figure that exceeds its own cumulative total. A merger that moved a year, a financing that moved a year, a filing credited to the wrong company. And a productivity block that — where it traces at all — traces to a single commentary citing a report nobody outside the consultancy has read.

Every prior layer of this inquiry reached the same conclusion from a different direction: the capability is genuine and the magnitude is marketing. This layer inverts the priority. Here the magnitude is the product, it is the thing the board is being asked to fund, and it is the least verifiable evidence in the stack.

The models are not what needs validating any more. The business case is.

Related: Five Agents, One Protocol, Zero Prospective Validation · 117 AI-Built Drug Programs, Zero Approvals · Four Points of Accuracy, Fifty Points of Marketing · Two of Eighty-Six: The Validation Deficit in AI Biomarker Discovery

Sources: BCC Research, Artificial Intelligence in Pharmaceutical Market (AIT020A, Sep 2025) · SNS Insider, AI in Pharmaceutical Market report #7678 · Precedence Research, AI in Pharmaceutical Market report #1485 · Strategy& (PwC), Reinventing pharma with artificial intelligence, 26 Mar 2024 · Deloitte, AI ROI: the paradox of rising investment and elusive returns (n=1,854, Aug–Sep 2025) · BCG, The Leader’s Guide to Transforming with AI · PitchBook, “AI biotechs fetch big premiums” · Fortune, “Sanofi CEO Paul Hudson…,” 27 Nov 2024 · Applied Clinical Trials, SCOPE COPD enrolment case study, 3 Jun 2026 · Merck newsroom, Merck–Google Cloud partnership, 22 Apr 2026 · Novo Nordisk/OpenAI, 14 Apr 2026; Novo Nordisk/AWS, 10 Aug 2026; Novo Nordisk/Anthropic, 16 Sep 2026 · Lilly investor release, Lilly–NVIDIA co-innovation lab, 12 Jan 2026 · Insilico Medicine prnewswire/company release, Lilly collaboration, 29 Mar 2026 · Insilico Medicine, HKEX listing announcement, 30 Dec 2025 · McKinsey, Life sciences dealmaking gains momentum as strategic pressures intensify · McKinsey, External innovation: biopharma dealmaking to boost R&D productivity · Nature/Biopharma Dealmakers (DealForma), 2025 AI deal analysis · JPMorgan, biopharma venture funding deck Q2 2025 · Galen Growth, 2024 digital health funding report · SVB, Healthcare Investments and Exits, 8 Jan 2025 · Crunchbase News, AI biotech and healthcare funding · TechCrunch, Xaira Series A, 24 Apr 2024 · Isomorphic Labs, Series B announcement, 12 May 2026 · Tempus IR, expanded AstraZeneca/Pathos agreements, 23 Apr 2025 · Recursion IR, Recursion–Exscientia definitive agreement, Aug 2024 · Roche Pharma Day materials, 28 Sep 2026 · Sanofi, AI in R&D: guiding Sanofi’s portfolio decisions, 24 Feb 2026 · Roche/Genentech–Recursion Validated Target Option milestone, 2026 · Define Ventures, Inside the C-Suite, Jul 2025 · ZS, Scaling AI in pharma and biotech: 2026 CDIO Research (n=115, Oct 2025) · Bain & Company AI adoption survey (n=408) · Gartner press release, top data and analytics predictions, 17 Jun 2025 · PharmExec, The Agentic AI Revolution in Biopharma, 10 Sep 2026 · Seal et al., AI Agents in Drug Discovery, Drug Discovery Today 2026 · Artefact, marketing-mix-modelling and automation ROI cases · Veeva product claims, MLR review · Tufts CSDD, Ther Innov Regul Sci 2025 · McKinsey, Scaling gen AI in the life sciences industry.

Research notes: [[AI-in-Pharma-Strategy-Layer-2026]]