EPC Shadow AI Case Study: The 38% Productivity Mandate
AI Decision Governance · an engineering, procurement and construction (EPC) firm serving energy · Field observation · November 2025
This is a field observation, described by profile to protect the people involved. It records what ModalPoint saw directly. It does not claim an outcome: the productivity gain was never measured, and that turns out to be part of the story.
/01: The Firm
An EPC firm doing engineering work for energy projects. The person at the center of this is its lead engineer: the one whose output sets the pace for the rest of the engineering team, and whose documents carry the firm's name to its clients.
/02: The Mandate
In November 2025 the lead engineer had a target from above: increase productivity 38% by the end of the year. The tool the company supplied to get there was Microsoft Copilot.
/03: What Was Actually Happening
The work was not running on Copilot. It was running on ChatGPT and Claude Code.
What went into those tools was the firm's core work product: client specifications, drawings, calculations, bid and cost data, emails and reports.
That is the whole finding, and it is enough. The company had a sanctioned AI tool, a hard productivity number, and a senior engineer hitting that number with tools the company had not chosen. On paper the firm had an AI rollout. In practice it had two AI stacks: the one IT issued and the one the work ran on.
Claude Code is worth singling out. It is not a chat window. It is a coding agent that runs on the engineer's own machine, reads and writes files, and executes commands. When that is the tool doing the work, the governance questions are no longer only about what text was pasted into a browser. They are about what files an agent could read, what it produced, and whether anyone reviews that output before it reaches a deliverable.
Did the engineer hit 38%? Nobody could say. The gain was hard to measure. So the firm had a hard target it could not score, running on a tool stack it could not see.
/04: Why It Happens
The engineer is not the problem here. The mandate is. A leadership team set a 38% target and a tool budget, and left the gap between them for the engineer to close. People who are measured on output use whatever produces output. If the approved tool had been the best one for this engineer's work, there would have been no reason to go around it.
It wasn't. Copilot lost for two reasons. It was not a tool like Claude Code, which works as an agent on the engineer's own machine. And it arrived without enough training to get real work out of it.
This is the pattern ModalPoint's shadow AI work is built around: you don't have a governance problem. You have a tool-quality problem creating a governance problem. A policy memo does not fix it, and neither does blocking the tools, because the target is still there on Monday.
/05: The Questions the Firm Needs to Answer
Every one of these has a real answer at this firm. The point of governance is that someone inside the company knows it.
- Which of those documents were client-confidential? Specifications, drawings and calculations are often covered by confidentiality terms in the client contract, and bid and cost data is commercially sensitive on its own.
- Whose accounts were they? A company-issued account and a personal one are governed by different data terms.
- Who reviews the output? Engineering work carries professional responsibility. An AI-assisted calculation still needs a named human who checked it.
- How was the 38% going to be measured? It was hard to measure, so no one can say whether it was hit. A target that cannot be scored cannot be governed either, and any gain that did come from unsanctioned tools is one the firm cannot reproduce with the tools it paid for.
- Would training have closed the gap? Copilot arrived without enough of it. Whether a trained team would still have gone around it decides what the firm should buy next.
/06: What Governing It Looks Like
Start from the tools people already use, not the ones on the purchase order. Inventory them, find out why the sanctioned tool lost, and then choose: approve the tool that works under company accounts and data terms, or fix the sanctioned one until it wins. Put a named reviewer on anything AI-assisted that leaves the building. And tie the productivity target to the approved stack, so the number and the governance measure the same thing.
ModalPoint's Shadow AI Exposure Assessment does the inventory and classification step, and AI Decision Governance covers the review and sign-off layer.
/07: A Question to Ask Your Own Team This Week
Ask your three most productive engineers which AI tool they used last. If any answer is not the one you issued, you have the same two stacks this firm had.