The evidence
Essays
The argument, built piece by piece. 57 essays, grouped under three pillars.
Latest essay
Layer · Process · 6 min
The Layer You Cannot Buy
BCG has measured the gap between what CMOs say about AI and what they have built. In pharma that gap has a name, it sits between functions, and nobody owns it.
By Mark Rhys Thomas · October 13, 2026
Read the essay →
Layer
Compute
The strategic asset. Where you make AI, and on whose terms, decides whether you can afford to run it.
- 01Layer · The full system · Compute, Data, ProcessJuly 22, 2026 · 10 min
Running the system
The finale of the twelve-week series walks the whole production system, seam by seam, with the receipts the past fortnight has provided. The advantage is not the machines. It is the system that runs between them.
- 02Layer · ComputeJuly 14, 2026 · 10 min
The Org Chart Is the Strategy
What Anthropic's energy hires tell us about compute sovereignty, and about the difference between buying AI and running it.
- 03Cornerstone · The ArgumentJune 8, 2026 · 8 min
AI Is a Production System. Pharma Already Knows How to Run One.
Pharma keeps treating AI like a science experiment when it should treat it like a factory.
- 04Layer · ProcessAugust 25, 2026 · 8 min
Pharma Is Only Reading Half Its AI Ledger
Pharma can tell you what AI costs. It cannot tell you what AI returns. That gap is not a reporting problem, it is the mechanism behind every pilot that never scaled.
- 05Layer · ComputeJune 16, 2026 · 7 min
Compute is a production input
You do not own the substrate, and you cannot win the scale race. What is left is management.
- 06Layer · ComputeJune 21, 2026 · 7 min
The Machine on the Desk
The sufficiency thesis was a claim about workloads. The new desktop AI hardware makes it a claim about furniture.
- 07Series · Power & InfrastructureMay 9, 2026 · 8 min
Your AI Pilots Are Not Failing. Your Compute Contracts Are.
Power & Infrastructure — MIT named integration as the cause of the 95% failure rate. In regulated industries, the cause sits one layer below that, in the contract.
- 08Series · Physical Layer of AI - PostscriptJune 3, 2026 · 6 min
The routing layer arrives.
The Physical Layer of AI — Perplexity ships the first frontier-lab hybrid local-server inference orchestrator. The routing layer is real. The value-share question is exactly as open as it was the day before.
- 09Series · Physical Layer of AI - Part 6May 2, 2026 · 10 min
What pharma should be doing now
The Physical Layer of AI, Part 6 — five things that change in the operating model when compute, power, water, and regulation stop being separate problems.
- 10Series · Physical Layer of AI - Part 5May 1, 2026 · 9 min
Ireland is the canary
The Physical Layer of AI, Part 5 — the European test case has already been run, and the results are in.
- 11Series · Physical Layer of AI - Part 4Apr 30, 2026 · 9 min
The nuclear pivot
The Physical Layer of AI, Part 4 — for thirty years the argument for nuclear was climate. It took eighteen months of AI demand to actually finance it.
- 12Series · Physical Layer of AI - Part 3Apr 29, 2026 · 9 min
The fossil-fuel detour
The Physical Layer of AI, Part 3 — for three years the AI industry marketed itself on a clean-future story. What changed is the procurement.
- 13Series · Physical Layer of AI - Part 2Apr 28, 2026 · 10 min
The grid wakes up
The Physical Layer of AI, Part 2 — what happens at the other end of the cable when hyperscaler capex meets a grid built for a 2010 world.
- 14Series · Physical Layer of AI - Part 1Apr 27, 2026 · 10 min
Compute is not the new oil. It is harder to replace.
The Physical Layer of AI, Part 1 — the moment AI stops being a software story and becomes an industrial one.
- 15Decoding AIMay 7, 2026 · 7 min
The deal Ireland never got.
On Anthropic, SpaceX, and one sentence in the international expansion clause.
- 16Decoding AIMay 5, 2026 · 5 min
The Hyperscaler Consensus
Decoding AI — three announcements, one pattern, and the value-capture question pharma boards are leaving for someone else to answer.
Batch Record issues in this layer
The archiveThe weekly newsletter, not the essays. Numbered and dated.
Layer
Data
The production input. A raw material with a supply chain, not a strategy theme.
- 01Layer · The full system · Compute, Data, ProcessJuly 22, 2026 · 10 min
Running the system
The finale of the twelve-week series walks the whole production system, seam by seam, with the receipts the past fortnight has provided. The advantage is not the machines. It is the system that runs between them.
- 02Layer · DataJuly 10, 2026 · 9 min
The Price of Not Choosing
For three years, most organisations never chose an AI model. The subscription chose for them. That era is ending, and the adjustment it demands is not technical.
- 03Layer · DataJuly 8, 2026 · 10 min
Data is a production input
Every AI strategy talks about data as an asset to accumulate. Production logic treats it as an input to qualify. The difference decides who wins.
- 04Cornerstone · The ArgumentJune 8, 2026 · 8 min
AI Is a Production System. Pharma Already Knows How to Run One.
Pharma keeps treating AI like a science experiment when it should treat it like a factory.
- 05Layer · DataJune 4, 2026 · 5 min
Data is a raw material, not a strategy
Layer · Data, stop having a data strategy and start qualifying your inputs the way a pharma plant qualifies an active ingredient.
- 06Layer · DataJune 4, 2026 · 5 min
The contamination problem
Layer · Data, a contaminated dataset cannot be told apart from a clean one by looking at it. The model just answers wrong, fluently, and the wrong answer becomes the next input.
- 07Layer · DataJune 4, 2026 · 5 min
Qualify your suppliers
Layer · Data, every dataset is delivered by a supplier. Most of them have never been qualified by anyone, and the model itself is the largest unqualified supplier of all.
- 08Layer · DataJune 4, 2026 · 5 min
Specifications for data
Layer · Data, good is not a feeling. It is a specification, written down before the batch arrives, and either the input meets it or it does not.
- 09Layer · DataJune 4, 2026 · 5 min
Your RAG is only as good as its corpus
Layer · Data, retrieval-augmented generation is a data problem in a tooling costume. Qualify the corpus, or tune the retriever forever.
- 10Decoding AIMay 2, 2026 · 6 min
AI adoption in pharma: a context problem, not a tooling problem
Decoding AI — a response to Ganes Kesari in MIT Sloan Management Review on why pharma AI programmes stall at the same point.
- 11Decoding AI — April 2026April 15, 2026 · 9 min
RAG AI Agents for Pharma Marketing Teams
Decoding AI — a plain-language guide to what RAG is, what it does, and how to deploy it safely.
- 12Decoding AI — April 2026April 8, 2026 · 3 min
RAG AI agents — what changes for a pharma marketing team
Decoding AI — generic AI tools are well-read strangers. RAG closes the gap.
- 13Layer · ProcessSeptember 29, 2026 · 6 min
Start a New Value Creation Plan
Bain tells private equity to start with the value creation plan. The plan is written before anyone can see the work AI changes, so it caps the answer.
- 14Layer · ProcessOctober 13, 2026 · 6 min
The Layer You Cannot Buy
BCG has measured the gap between what CMOs say about AI and what they have built. In pharma that gap has a name, it sits between functions, and nobody owns it.
Batch Record issues in this layer
The archiveThe weekly newsletter, not the essays. Numbered and dated.
Layer
Process
The human operating layer, where AI is run with discipline. Three Ps: Permission, Practice, Proof.
Process · Permission
- 01Layer · Process · PermissionJuly 20, 2026 · 9 min
The discovery bottleneck
What a non-pharma company gains, and gives up, by entering drug discovery with a model instead of a legacy.
- 02Layer · Process · PermissionJuly 15, 2026 · 9 min
The best lawyer in the world cannot be called to the bar
Andreessen says the intelligence is already in every pocket, but none of it can practise. That gate has a name in this framework — Permission, the first of the three Ps inside the Process layer.
- 03Standalone · the accountability gapJune 12, 2026 · 6 min
The constraint nobody will name
Twelve reports, same five things. The one constraint they all walk around is the one that matters.
- 04Series · Accountability in Regulated AIJune 5, 2026 · 6 min
The system is getting more efficient, and I need more budget.
The Named Owner / Judgement Muscle — cheap models will swallow most of the work. They will also send the bill up. The leaders who cannot explain why both are true at once are the ones who lose control of their spend.
- 05Series · Accountability in Regulated AIMay 19, 2026 · 10 min
You cannot acquire accountability.
Accountability in Regulated AI — the Publicis–LiveRamp deal is being read as a neutrality story. It is an accountability story, and pharma is about to learn the difference the expensive way.
- 06Series · Accountability in Regulated AIMay 13, 2026 · 9 min
The named owner.
Accountability in Regulated AI — most pharma AI deployments fail their first audit the moment someone asks who owns the decision. Not the function. The person.
- 07Series · Accountability in Regulated AIMay 22, 2026 · 9 min
The cost of holding it.
Accountability in Regulated AI — holding accountability properly is expensive, the cost never shows up in the line item that names it, and the organisations paying it cleanly are buying a position the rest of the market will eventually have to buy at a higher price.
- 08Series · Accountability in Regulated AIMay 28, 2026 · 10 min
The commercial design space.
Accountability in Regulated AI — for the rooms ICH Q8 does not reach. Medical affairs, market access, promotional review, KOL prioritisation, commercial analytics: where the design space has to be drawn from scratch.
- 09Series · Accountability in Regulated AIMay 25, 2026 · 10 min
The design space.
Accountability in Regulated AI — the operational answer. What the discipline actually looks like at 8:30 on a Tuesday with QA, Validation, IT, and a vendor on the call.
- 10Series · Accountability in Regulated AIMay 20, 2026 · 8 min
The judgement muscle.
Accountability in Regulated AI — name the owner, refuse to let the accountability migrate, and you have still left the hardest problem unsolved: the named owner who quietly becomes a rubber stamp.
- 11Decoding AIMay 6, 2026 · 4 min
FDA + EMA regulator AI moves
Decoding AI — two regulators, two different bets, and an uncomfortable implication for pharma boards still arguing about their own AI governance.
Process · Practice
- 01Layer · Process · PracticeJuly 15, 2026 · 10 min
The cheapest thing in your AI programme is the intelligence
Capability is abundant and near-identical across competitors. The routines wrapped around it are not. Practice — the second P inside the Process layer — is the compounding advantage that shows up 6 to 12 months after adoption.
- 02Layer · Process · PracticeJuly 15, 2026 · 9 min
The 350 engineers Ford could not prompt
Ford rehired 350 veteran engineers after its AI quality system fell short. The story is not about the AI. It is about Practice — the discipline of running AI inside the work, with named people who correct it and corrections that feed back.
- 03Cornerstone · The ArgumentJune 8, 2026 · 8 min
AI Is a Production System. Pharma Already Knows How to Run One.
Pharma keeps treating AI like a science experiment when it should treat it like a factory.
- 04Decoding AIMay 7, 2026 · 5 min
AI Is a Power Tool. Treat It Like One.
The nine-second database wipe was not an AI failure. It was an operator failure. Pharma leaders should pay attention.
- 05Decoding AI — April 2026April 27, 2026 · 4 min
Pharma doesn't trust one expert opinion. Why are we trusting one AI model?
Decoding AI — Karpathy's LLM Council pattern, and why multi-model deliberation maps onto how pharma already governs hard decisions.
- 06Decoding AI — April 2026April 28, 2026 · 5 min
The Council pattern is easy to sketch. Making it pharma-grade is the harder problem.
Decoding AI — a seven-layer architecture for a multi-model Council that Quality and Compliance will actually sign off on.
- 07Decoding AI — April 2026April 29, 2026 · 5 min
A worked example: running a medical writing QC pass through a Council.
Decoding AI — what a four-model Council actually produces on a late-stage CSR efficacy section, and where the human reviewer still earns their keep.
- 08Decoding AI — April 2026April 22, 2026 · 8 min
Agentic AI in Pharma
Decoding AI — when the AI stops answering questions and starts taking actions.
Process · Proof
- 01Layer · ProcessAugust 25, 2026 · 8 min
Pharma Is Only Reading Half Its AI Ledger
Pharma can tell you what AI costs. It cannot tell you what AI returns. That gap is not a reporting problem, it is the mechanism behind every pilot that never scaled.
- 02Layer · ProcessOctober 13, 2026 · 6 min
The Layer You Cannot Buy
BCG has measured the gap between what CMOs say about AI and what they have built. In pharma that gap has a name, it sits between functions, and nobody owns it.
- 03Layer · Process · ProofSeptember 23, 2026 · 8 min
Anthropic's Independent Evaluator Is Owned by Its Commercial Partner
Quis custodiet ipsos custodes, who will guard the guards themselves? Pharma answered that question decades ago and wrote the answer into law. AI governance has not.
- 04The EvidenceSeptember 2, 2026 · 6 min
Emergence Capital's $150 Million Bet Isn't About AI
A venture firm just paid $150 million to confirm what a pharmaceutical content operation proved years before AI-native services had a name.
- 05Layer · Process · ProofAugust 13, 2026 · 8 min
The Watermark Is Not a Control
Anthropic now marks everything Claude writes. Here is what that signal can carry, and what it cannot.
- 06Layer · Process · ProofAugust 7, 2026 · 8 min
Evidence Is Not Oversight
Brightbeam AI has published a draft specification for the part of the agent stack nobody has standardised. Worth crediting, and worth reading the caveats it publishes about itself.
- 07Layer · The full system · Compute, Data, ProcessJuly 22, 2026 · 10 min
Running the system
The finale of the twelve-week series walks the whole production system, seam by seam, with the receipts the past fortnight has provided. The advantage is not the machines. It is the system that runs between them.
- 08Layer · Process · ProofJuly 17, 2026 · 9 min
The record is the second half of the product
Proof is the third P in the Process layer — the evidence, produced by the system as it runs, that lets an absent sceptical reader trust the work was done to standard. This week the FDA, Verisk and McKinsey all repriced it.
- 09Layer · Process · ProofJune 4, 2026 · 6 min
The evidence trail nobody is building
Layer · Process · Proof, if it is not documented, it did not happen. The companies running AI in production right now mostly cannot prove what their systems did last March.
- 10Layer · Process · ProofJune 5, 2026 · 5 min
What the inspector will ask
Layer · Process · Proof. The questions were always coming. Rehearse them now, while the answers are still yours to change.
- 11Layer · Process · ProofJune 5, 2026 · 6 min
ALCOA for AI
Layer · Process · Proof. Pharma already has the rulebook. The standard exists, fully formed. Deploying AI earned no one an exemption from it.
- 12Layer · Process · ProofJune 5, 2026 · 5 min
Reproducibility is something you design in
Layer · Process · Proof. A result you cannot reproduce is not a result. It is an opinion the machine happened to hold that day.
- 13Layer · Process · ProofJune 5, 2026 · 5 min
Audit after the fact is too late
Layer · Process · Proof. There is no such thing as a retrospective evidence trail. The audit does not create the evidence. It only finds out whether you did.
- 14Decoding AIMay 5, 2026 · 7 min
Sovereignty is control, not location. Tell that to the EMA inspector.
A follow-up to "The Hidden Cost of AI: Data Sovereignty" — why the hyperscaler reframe of sovereignty as control is incomplete for pharma.
- 15Layer · ProcessSeptember 29, 2026 · 6 min
Start a New Value Creation Plan
Bain tells private equity to start with the value creation plan. The plan is written before anyone can see the work AI changes, so it caps the answer.
Batch Record issues in this layer
The archiveThe weekly newsletter, not the essays. Numbered and dated.
- No. 1026 October 2026OpenAI Did Not Buy a Molecule. It Bought the Place Where the Programme Is Run.
- No. 0812 October 2026Pharma Commercial Does Not Have an AI Problem. It Has a Decision Problem.
- No. 075 October 2026Cool Does Not Scale: Novartis Says What Every Pharma Steering Committee Needs To Hear
- No. 0521 September 2026Pharma Does Not Get the Oversight Discount
- No. 0414 September 2026The Supplier Wrote the Rule
- No. 0125 August 2026The Pilot Trap now has a number