The Mapping Problem: Why Enterprise AI Fails Before It Starts

---

The research

The paper, "Mapping AI into Production" by Hyunjin Kim, Dahyeon Kim, and Rembrand Koning (INSEAD / Harvard Business School, March 2026), studied 515 high-growth startups in a randomized controlled trial. The treated group received one thing the control group did not: case studies showing how other firms had reorganized their production processes around AI.

The results should change how every enterprise leader thinks about AI investment.

And here is the number that will change boardroom conversations: treated firms demanded 39.5% less external capital than the control group, with no change in workforce size. They achieved more with less.

The researchers call the core friction the mapping problem: the challenge of discovering where and how AI creates value within a firm's production process. It is not an access problem. Both groups had the same tools, the same API credits, the same technical training. The difference was that treated firms were shown how to search more broadly across their production chain for where AI could reorganize work — not just assist with individual tasks.

---

Why this is harder in enterprises than startups

The paper studied early-stage startups, where organizational inertia is low. The researchers acknowledge this directly: "if these frictions bind even in early-stage ventures with minimal organizational complexity, the challenge is likely greater still for established firms."

Having spent 20 years building enterprise data platforms at Cisco, Samsung, and across healthcare and retail clients, I can confirm: the mapping problem in enterprises is not just harder. It compounds.

The FazeShift case: replacing human glue with AI

The paper illustrates this beautifully. In accounts receivable, the conventional process alternates between software systems and human clerks who bridge between them. Automating one step still leaves the clerk manually entering data and sending emails. Only when the full chain is automated does the process transform from a labor-intensive service into a scalable system.

---

What mapping actually looks like in practice

The paper names what enterprise architects do intuitively: search across the full production process, not just the obvious pain point. In my work, mapping follows four patterns I have applied at every scale:

---

The governance gap the paper does not address

The paper identifies discovery as the bottleneck. I would add a second bottleneck that binds immediately after discovery: governance.

Once a firm discovers that AI can reorganize a production process, it needs to answer a harder question: how do we deploy this safely, auditably, and in compliance with regulatory requirements?

In healthcare, discovering that an AI agent can conduct clinical intake in eight minutes instead of an hour is the mapping problem solved. But deploying that agent requires PHI-aware data zoning, row-level security scoped by provider and site, audit lineage documenting every decision the agent made, and conformance profiles defining what the agent is allowed to do before it runs.

---

What this means for enterprise leaders

The paper's practical implication is clear: "Teaching managers and entrepreneurs how to solve the mapping problem may be at least as important as ensuring they have access to the technology."

For enterprise leaders, I would extend this:

Stop starting with the technology. Do not ask "how can we use AI?" Ask "where in our production process are humans bridging between systems, and what happens if we automate the full chain, not just one step?"

Show your teams how others reorganized. The paper's treatment was simply showing firms case studies of how other firms reorganized around AI. That intervention alone produced 44% more use cases and 1.9x higher revenue. In enterprises, this means sharing concrete examples of AI-driven reorganization across your industry — not generic AI demos.

Invest in the data foundation first. The paper assumes tool access is equal. In enterprises, it is not. If your data is fragmented, ungoverned, and untrustworthy, no amount of AI mapping will produce firm-level gains. The data platform is the AI platform.

Build governance into the mapping process. Every AI use case discovered through mapping should be evaluated not just for value but for deployability: can we govern this? Can we audit it? Can we explain it to a regulator? If not, it is not ready for production regardless of how much value it promises.

Measure firm-level outcomes, not task-level productivity. The paper's most important contribution is showing that task-level gains do not automatically aggregate to firm-level performance. If you are measuring AI success by how fast a report generates or how many emails an LLM drafts, you are measuring the wrong thing. Measure revenue. Measure customer acquisition. Measure capital efficiency. Those are the outcomes that tell you whether your mapping is working.

---

The mapping problem is not going away

The paper closes with an insight that should keep every enterprise leader awake:

"As models grow more capable, the set of activities they can perform expands, making the search space larger and the mapping problem harder. A firm that has figured out where to use today's AI will likely face the same discovery problem again when next year's models can do more."

This is not a one-time exercise. It is a continuous discipline. The firms that build the organizational muscle to map AI into production, govern it, and measure firm-level outcomes will compound their advantage. The firms that wait for the technology to become obvious will always be one cycle behind.

---

Further reading

• Kim, H., Kim, D., & Koning, R. (2026). "Mapping AI into Production: A Field Experiment on Firm Performance." INSEAD Working Paper 2026/20/STR.
• Your Agents Need a Contract — Spec-driven governance framework for enterprise AI agents, with a nine-framework ADK comparison scorecard.

---

This post is part of "the inference," a series on enterprise AI strategy and architecture.