The Pilot Trap now has a number
For three years the Pilot Trap has been something I could describe but not measure. Accenture has now measured it, across nineteen industries, and the direction of travel is worse than I expected.
By Mark Rhys Thomas
The Lead
For years I have argued that pharma's AI problem is not technical. A company runs a clever pilot, the demo lands, leadership applauds, and nothing scales. A year later they run another pilot. I call it the Pilot Trap. Until now my evidence for it has been what I have seen inside client organisations, which is a weak position to argue from in a boardroom.
Accenture has just measured it.
Across 3,000 C-suite leaders and 3,000 employees, in 19 industries and 20 countries, all at companies above $500 million revenue, the share reporting widespread and sustained business value from AI is 23 percent.
Earlier in the same year it was 32 percent.
Over that same period, 82 percent of the C-suite increased AI investment.
Read those together. More money going in. Less value coming out.
That is not a technology curve. It is not a confidence problem either, because confidence is high and rising. It is a conversion problem, and conversion is an organisational function, not a technical one.
The survey says as much itself. Asked why they cannot link AI spend to business value, 57 percent cited weak business-technology alignment, 51 percent insufficient data and analytics capability, and 50 percent an inability to isolate AI's contribution from everything else moving at the same time.
Not one of those is a model problem. All three are plant problems. No shared foundation, no instrumentation, no defined ownership.
The finding I keep returning to is where the skills gap sits. Leaders and employees agreed, and named the same place: middle management, at 36 percent from both groups. That is the layer that has to turn a capability into a changed way of working. In pharma it is brand leads and regional commercial managers. It is also the layer nobody has trained, because AI budgets go to the data teams who need the training least.
Here is the part that should trouble a pharma board. Twenty-three percent is a cross-industry number, and pharma is not the average. On the maturity curve I use with clients, most pharma companies sit at stage one or two of four, below the cross-industry midpoint. If 23 percent is what nineteen industries manage between them, the pharma figure is lower than that.
Nobody in manufacturing calls a single good batch a success. Pharma has spent three years applauding good batches.
Also on the Line
COMPUTE
Cheap models are not replacing expensive ones. They are being added to them. The assumption that inference costs will fall on their own is now measurably wrong. Enterprises are adopting routers, which send each request to the cheapest model that can actually do the job, and yet frontier model spend is still rising among the fastest-growing companies. Accenture's own internal platform runs roughly 8.7 trillion tokens a week at about one sixth of frontier cost, and it estimates only 10 to 20 percent of enterprise tasks are complex enough to need a frontier model at all. Most companies cannot tell which of their tasks are which, so they buy the expensive option for everything and call it prudence.
DATA
64 percent of drug launch delays are manufacturing data problems. Accenture analysed FDA complete response letters from 2019 to 2024 and found 64 percent of launch delays traced to chemistry, manufacturing and control issues. Only 35 percent of surveyed executives describe their technical operations as a connected organisation. A launch delayed by fragmented data costs exactly the same exclusivity months as one delayed by a failed trial. Boards treat the two completely differently, and only one of them gets an executive sponsor.
PROCESS
Deloitte's verdict on agents: the technology is ready, the leadership is not. “While agentic capabilities are technically ready, most organisations have yet to activate them beyond experiments or simple use cases.” That sentence comes from a firm that sells agent implementations, which is what makes it worth more than any vendor benchmark. The named constraint is coordinated executive decision-making. Not model quality, not readiness, not cost.
What Did Not Meet Spec
Every week I read roughly sixty items and publish four. This is the in-process control.
Rejected this week: twelve sponsored placements written to read as editorial, three recruitment ads, and the usual tooling minutiae, quantisation formats, IDE changelogs and model-serving internals, which matter only if you are writing code this week.
One rejection worth naming, because it cuts against my own lead story.
I used six figures from Accenture's tokenomics report. I rejected a seventh. The report claims only 15 percent of organisations would bank the savings if token prices fell 25 percent. That figure came from a simulation of 6,000 synthetic executive profiles, not from real respondents. It is the softest number on the page, and if I put it in front of a pharma audience, the one person in the room who checks would be entitled to doubt the other six.
Showing the batch record means showing the rejects.
The Line
Investment in AI went up this year. Realised value went down. Both numbers came from the same survey.
Sources
- 1. Accenture, Pulse of Change: What's top of mind for today's leaders, research report, 26 July 2026. Survey of 3,000 C-suite and 3,000 employees, 19 industries, 20 countries, companies above $500 million revenue, fielded April to June 2026.
- 2. Accenture, AI Tokenomics for Enterprise Value, perspective, 29 July 2026. Figures verified verbatim against the published page on 25 August 2026.
- 3. Accenture, Reinventing biopharma from lab to line, research report, 10 February 2026.
- 4. Deloitte, The executive decisions shaping the value of the digital workforce, perspective, 4 March 2026.
Every figure in this issue was checked against the primary source. Where a number originates with a third party, it is attributed to them and not to the firm quoting it.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.