Pharma Commercial Does Not Have an AI Problem. It Has a Decision Problem.
Two BCG papers describe the same gap from opposite ends. Patients are lost where no one owns the next decision, and agents cannot be trusted until someone writes down which decisions they may take.
By Mark Rhys Thomas
The Lead
Most pharma commercial teams do not have an AI problem. They have a decision-making problem that AI keeps exposing.
Two BCG papers, both read this week, show where the problem sits. Neither says it this way. Both describe it exactly.
The first is about patients. In “Listening to Patients in the Digital Age”, published in June, BCG says pharma and medtech brands lose prospective patients because the signals patients send land in separate systems. Web visits, survey answers, hub activity and call centre conversations never meet. Outreach runs on a fixed calendar instead of on what the patient just did. BCG puts the share of brands that decide at the level of the individual, across channels, at about 2 percent.
Its diagnosis of the rest is the useful part. The bottleneck is rarely the platform. It is the absence of a shared patient data model and a journey roadmap. And where digital, marketing and data and analytics all touch the same programme with no clear owner, patient progress stalls.
Read that as an operator. The technology is bought. The data exists somewhere. What is missing is a named owner for the question: what happens to this patient next.
The second paper is about agents. In Part 4 of its next-best-action series, published in July, BCG describes marketing in which software agents choose the action, channel and timing for each customer from a shelf of pre-approved content. Then it draws the line that matters. Agents may act alone on high-frequency, low-stakes choices from approved assets. They recommend, and a human approves, when the stakes rise or the shelf has a gap. And four things, objective setting, budget, shelf strategy and any widening of the agent’s own autonomy, “should never be delegated to AI”. BCG’s warning is that firms which deploy agents without settling these boundaries meet a crisis of trust that stops them scaling.
The examples are banks and retailers. The mapping to pharma is mine, and it is short. The pharma shelf is MLR-approved content, and BCG’s patient paper names it directly, noting that the review cycle is what pushes teams back to one message for everyone. An agent can pick between approved variants on a Tuesday afternoon. Who decides what goes on the shelf, what the system is optimising for, and when it may act without a person is a question of decision rights. Most commercial organisations have never written those rights down for people. They cannot hand them to software.
That is the Pilot Trap seen from the commercial side. Last week’s issue ended on a test: count decisions, not demos. Roche counts the share of portfolio decisions where AI moved the outcome. A commercial team that cannot say who owns the next-best-action decision cannot run that count, because there is no decision to count. The pilot works because a pilot is allowed to ignore the question. Production is not.
The rebuttal is fair, so take it. BCG sells this work, and both papers carry the interest that implies. The 85 percent lift in patient-reported activation in a cardiac monitoring programme is one client case, with no baseline or sample size in the article, and I have not used it. The autonomy list is written for retail banking. And a sceptic will say this is a data problem. BCG says it is not, and it would sell more data work if it thought it was.
What survives the discount is a structure, not a number. Signals that no one joins. A shelf that no one is accountable for. Boundaries that no one has written down. None of it needs a better model. All of it stops a good model scaling.
Any commercial leader can test this without a vendor. Pick one patient journey. Write down who decides the next action today, what they may decide alone, and what must go to a person. If that takes more than a page, or three people give three answers, the agent is not the problem you are about to have.
Decide who decides, then buy the model.
Also on the Line
COMPUTE
The chip build is financed with debt, and the grid is saying no. Reported this week: Broadcom arranging more than $50bn of financing for OpenAI’s custom chip, Oracle in talks with Apollo and Goldman Sachs over a chip purchase, and SpaceX approaching lenders about $40bn for Nvidia chips. Separately, Texas paused new data centre permits because speculative projects are overwhelming its grid. For a pharma buyer the link is indirect. The price and supply of the compute under every AI contract now depends on credit markets and electricity, neither of which sits in a procurement plan. These are press reports relayed by a newsletter. I have not read the primaries.
DATA
Watermarks reach text. OpenAI will add invisible watermarks to eligible ChatGPT and Codex text in the EU to meet AI Act transparency rules, and says editing, translation and short passages weaken detection. Google has widened its SynthID Detector globally to read Google, OpenAI and NVIDIA marks. Issue 07 covered SynthID Bio for proteins. The same logic now reaches documents: provenance is becoming a property of the input. For regulated writing, such as submissions and quality records, a mark that breaks on a light edit is a weak control. Both items come via a newsletter. The vendors’ own notices were not read.
PROCESS
Controls up, value up, says BCG. Issue 07 reported KPMG’s finding that the share of firms building controls into their agents fell from 43 to 30 percent. This week BCG’s Applied AI Index, drawn from more than 1,300 senior leaders across 20-plus sectors, reports that companies with all six AI controls in place generate three times the agentic value of those with one, and calls controls an enabler, not a brake. It is a correlation from a survey run by a firm that sells the remedy, and the six controls are defined in a full report I have not read. Set beside KPMG, it points the way the lead does.
What Did Not Meet Spec
Thirty-five emails came through this week’s pull, across five consultancies and two TLDR editions, plus hubs at McKinsey, Accenture, Deloitte and BCG checked directly and Brightbeam’s week 41 roundup read on the web. All seven TLDR issues and eight McKinsey emails were read in full, and eleven pages were read at source across McKinsey and BCG. Rejected by category: about thirty sponsored placements and job adverts in the TLDR editions; consumer hardware, robotics demos, benchmark papers and non-AI science; McKinsey digests on energy, beauty, dairy and retail; KPMG’s board digest and EY’s Budget webcast, neither with any AI content. Accenture’s email channel and Deloitte’s remained empty, and nothing on either hub could be dated.
Three named items.
BCG’s “almost half create value”. The Applied AI Index says almost 50 percent of companies generate value from AI, counting the 7.5 percent it calls future-built and the 41 percent it calls scaling, who create “some value”. McKinsey’s 6 percent, used in issues 02 and 07, is companies attributing at least 5 percent of EBIT to AI. They measure different things at different thresholds. I have not set them side by side in this issue, and I would not recommend anyone does.
The cardiac device case. BCG reports an 85 percent lift in patient-reported activation, and a reversal of a multiyear decline in device returns, from a single AI-powered programme. It is the most striking number in the patient paper and it cuts in favour of this issue’s lead. It names no client and gives no baseline or sample. It stays out.
“Fewer than 10 percent of drugmakers are scaling AI in regulated manufacturing.” Brightbeam’s week 41 roundup states this with no originator. It would be a clean Pilot Trap figure for the shop floor. No originator is given and I have not traced one, so it is not used. Tracing it is on the list for next week.
The Line
Pharma commercial does not need a smarter model. It needs someone who owns the decision the model is about to make.
Sources
- 1. Boston Consulting Group, “Listening to Patients in the Digital Age”, bcg.com, 18 June 2026. Full article read. Source of the signals landing in separate systems, the roughly 2 percent of brands deciding at the individual level across channels, the bottleneck not being the platform but a shared patient data model and journey roadmap, ownership across digital, marketing and data and analytics, and the MLR review cycle as a bottleneck. The cardiac device case (85 percent lift in patient-reported activation) is a BCG client case, unnamed and with no baseline, and is rejected in this issue. BCG sells patient engagement and AI work; both of its papers here carry that interest.
- 2. Boston Consulting Group, “How Agent-Native Marketing Transforms Operations”, Part 4 of a five-part series on next-best action, bcg.com, 2 July 2026. Full article read. Source of the autonomy boundary (agents act alone on high-frequency, low-stakes decisions from pre-approved assets; recommend for human approval at higher stakes or shelf gaps), the four decisions that “should never be delegated to AI” (objective setting, budget, shelf strategy and expansion of agent autonomy), and the crisis of trust warning. The article’s worked examples are retail banking. The application to pharma commercial operations and MLR-approved content as the “shelf” is my inference, not BCG’s claim.
- 3. Boston Consulting Group, “The Formula for Agentic AI Value” (2026 Applied AI Index), bcg.com, 30 September 2026. Survey of more than 1,300 CxOs and senior leaders across 20-plus sectors. Page read; full report not read. Source of 3x agentic value with all six AI controls versus one, 7.5 percent future-built, 41 percent scaling and almost 50 percent generating value. The six controls are not enumerated on the page. A correlation, not a causal finding.
- 4. TLDR AI, 8 October 2026, items on Broadcom, Oracle and SpaceX chip financing, Texas data centre permits and the SynthID Detector expansion; and 6 October 2026, OpenAI EU text watermarking. Newsletter reads only. Primary reports not read. The financing talks are reported, not confirmed by the companies in the newsletter.
- 5. Brightbeam, Second Nature Intelligence, Week 41, brightbeam.com, 12 October 2026. Source of the unsourced manufacturing figure rejected above.
- 6. Batch Record issue 07, 5 October 2026, for the count-decisions-not-demos test, KPMG US AI pulse (43 to 30 percent of firms building controls into agents; 314 US leaders, fielded July and August 2026), and SynthID Bio. Issues 02 and 07 for the McKinsey 6 percent definition (McKinsey, “The state of AI in 2026: On the road to ROI”, 25 August 2026).
Applying BCG’s patient-engagement and marketing-governance findings to pharma commercial decision rights is my inference, not BCG’s claim.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.