Mott MacDonald's people built more than 8,000 agents. The next problem was finding, merging and retiring them
The engineering consultancy went from 50 Copilot licenses to a rollout for 20,000 people. Its agents brought duplicates, dormant agents and orphans when creators left. A catalog became the control.
Listen to this article · 4 min
AI-generated narration of the full article.
Mott MacDonald, the global engineering consultancy, has about 20,000 people. According to Microsoft’s account of its rollout, those people have created more than 8,000 AI agents. That is one agent for every two or three employees, and nobody planned that number.
The company’s story is useful because it does not stop at adoption. It describes what happened next.
From 50 licenses to everyone
The rollout was gradual. Mott MacDonald started with 50 Microsoft 365 Copilot licenses after introducing it in 2025, expanded to 1,000, then to 5,000, and gave half of the second wave to early-career professionals. It has now signed a five-year agreement to bring Copilot to all employees, as part of a larger investment in Microsoft 365 E7, and is preparing Microsoft Fabric as the central data platform for Copilot and agents.
The work examples are specific to engineering. On one airport project, AI processed around 220,000 requirements and flagged about 200 conflicts for engineers to review. The company aims to cut its average proposal process from a week to 48 hours. An app called EMMA has captured more than 10,000 lessons learned from project reviews since May. These figures come from Microsoft’s story and are not independently audited.
What 8,000 agents did to the organization
The candid part is the list of problems. Fast growth produced overlapping solutions, with several agents built for the same task; dormant agents that nobody used anymore; and knowledge gaps when creators left or changed roles, leaving agents that still ran but that nobody understood.
None of these is a model problem. They are the same problems companies had with spreadsheets, Access databases and SharePoint sites, arriving faster. An agent takes minutes to create, so the inventory grows at the speed of curiosity.
| What it costs | The control | |
|---|---|---|
| Duplicates | Effort spent twice; inconsistent answers to the same question | A searchable library, checked before building |
| Dormant agents | Access and connectors left open for nobody | Expiry unless someone renews |
| Orphans | Running logic nobody can explain or fix | A named owner, transferred when people move |
| Promising local agents | Good ideas stay in one team | A path to review and scale the best ones |
The library is the governance
Mott MacDonald’s response is a central AI library that makes agents discoverable and helps identify the ones worth scaling. Microsoft describes Agent 365 as a future control tower for overseeing the agent estate. Nasrine Tomasi, the firm’s global head of AI and data, frames the human side too: employees need to review AI output critically and know when their own judgment must take over.
The library is the right first move because it serves two purposes with one artifact. For employees, it answers “has someone already built this?” For governance, it is the inventory, and every other control depends on an inventory. We made that argument when SAS started governance with an inventory, and Mott MacDonald shows what it looks like when the inventory is needed after the fact.
The orphan problem deserves its own rule. An agent built by a project engineer who has moved to another office keeps running with that person’s design choices, and sometimes with connections to that person’s data sources. We discussed who owns a bot when its author leaves; the answer has to exist before the person leaves, not after.
For companies a year behind
The lesson is not to slow agent creation. Mott MacDonald’s spread of agents is how it found work worth automating. The lesson is to put the catalog, the owner field and the expiry date in place while the count is in the hundreds. Retiring agents nobody owns is far easier at 300 than at 8,000.