[JETPACKS THEY SAID]

The one thing AI training
programs get wrong

Most AI adoption programs teach people how to use the tools. A field experiment across 515 startups found that was never the constraint. The firms that pulled ahead did so because they learned where to deploy AI, not how. Seeing what reorganized production actually looks like, in concrete detail, across other firms, was worth 1.9x more revenue and $220K less capital needed. Access was never the problem. Discovery was.

Based on: Kim, Kim & Koning  /  "Mapping AI into Production" /  INSEAD Working Paper 2026/20/STR

Ten key findings
01 / 10
The bottleneck is search, not access

Both groups had identical tools, API credits, and technical training. Access was equal. Outcomes were not. The constraint was the cognitive work of discovering where AI creates value in a specific production process.

INPUTS (both groups identical)
API creditsModelsTrainingMentorship
OUTCOME
CONTROL
TREATED
02 / 10
Task-level gains don't aggregate automatically

Abundant evidence shows AI improves individual tasks. What remained unresolved: does that compound to the firm? Yes, but only when firms solve the search problem first. Without it, gains dissipate before reaching the bottom line.

Task
+30%
Firm
flat
mapping not solved
03 / 10
Treated firms found 44% more use cases

Control firms typically reported zero or one new use case per week. The gap emerged after week 3 and widened each week. Cumulative discovery, not a one-time knowledge transfer.

CONTROL6.1
TREATED6.1
Baseline (week 2)
04 / 10
Gains landed in product and strategy, not email

The largest differentials were in areas that required rethinking how work is organized. Research, writing, and sales barely moved. The intervention shifted firms from adopting tools to redesigning their production process.

control   treated
Product dev
+0.67
Prod / strategy
+0.41
Biz operations
+0.47
Marketing
+0.33
Tech infra
+0.27
Sales
+0.13
05 / 10
The performance effects are large

Not marginal efficiency improvements. Treated firms completed more tasks, acquired more paying customers, and generated substantially higher revenue over the same ten weeks.

?
more tasks
?
customers
?
revenue
tap each to reveal
06 / 10
Gains concentrate at the top, not across the board

Revenue effects are small through most of the distribution and spike at the 90th-95th percentile. AI raises the ceiling of what the best-positioned firms can reach. It does not lift the floor.

P90P10P50P95Move the dot to where the treatment effect lands
07 / 10
Faster growth, same headcount, less capital

Treated firms demanded $220K less in outside capital (39.5% reduction) with no change in labor demand. The capital reduction was sharpest above the 60th percentile.

INPUTS
Headcount
Capital ask
OUTPUTS
Revenue
Customers
08 / 10
Technical background made no difference

Technical founders did not benefit less. High-traction firms did not benefit more. The constraint was search scope, not skill or prior performance. Tap each group.

Technical / high tractionsignificant (+p<0.05)
Non-technical / high tractionsignificant (+p<0.05)
Technical / low tractionsignificant (+p<0.05)
Non-technical / low tractionsignificant (+p<0.05)
09 / 10
Automating one step moves the bottleneck, it doesn't remove it

An eight-step AR process with one automated step still has seven human handoffs. Because firm activities are complementary, partial integration preserves the original constraint. The full sequence must be redesigned.

Partial automation (1 of 8 steps)
Bottleneck: preserved
10 / 10
Better models make the search harder, not easier

As AI capabilities expand, the space of possible applications grows larger. A firm that has mapped today's AI into production will face the same discovery problem when the next generation arrives. The bottleneck is managerial, not technological.

Current generation
Ten findings from a field experiment  /  INSEAD Working Paper 2026/20/STR  /  ssrn.com/abstract=6513481
Learn Where to Deploy AI

Want to talk about this?

If something here resonated with a problem you're working on, let's spend 15 minutes on it.

Book a Free Call
Book a Free Call