The case study comes from Microsoft UK Stories and describes an engineering consultancy that took the least cautious possible route into agents: it let everyone build them.
From 50 licences to a company-wide rollout
Mott MacDonald is a UK-headquartered engineering and development consultancy with a global workforce. It adopted Microsoft 365 E7 in late July 2026, moving from localized pilots to decentralized agent creation across the whole company. Internal adoption produced thousands of specialised agents for bidding, compliance and engineering work, built by staff rather than a central team.
The company also carried over EMMA, its internal assistant, and embedded it inside Microsoft Teams. EMMA now runs structured project debriefs and indexes knowledge across disciplines. Since May, it has logged more than 10,000 lessons learned and action items.
The per-employee arithmetic
Divide the published figures and you get a picture the press release does not spell out. Eight thousand agents across 20,000 employees is roughly 0.4 agents per person. That is not a tool rolled out to everyone. It is a tool rolled out to everyone plus a long tail of small automations that individual engineers, bid managers and compliance staff decided they needed.
Licence growth tells the same story from the other end. Fifty Copilot seats to 20,000 E7 seats is a 400-fold expansion between the pilot and the July 2026 E7 rollout — the time span is anchored to those two documented points rather than to a round claim of "under a year". The 10,000 lessons logged in Teams work out to roughly 1,650 a month, or about one per twelve employees per month. Modest per head. Large in aggregate.
One airport, 220,000 requirements
The most concrete result in the case study comes from a single airport infrastructure project. AI parsed dense technical contracts and reviewed roughly 220,000 requirements, flagging about 200 conflicts.
Two hundred out of 220,000 is a hit rate of about 0.09%, or one conflict per 1,100 requirements. That sounds tiny until you price a conflict found late. Requirement conflicts in infrastructure contracts lead to rework, change orders and schedule slips, and the cost of finding one in a document review is a fraction of finding it on site. The value sits in early detection, not volume.
It also sets a limit. Two hundred flagged items still need an engineer to judge each one. The case study describes human oversight as part of the workflow rather than something automated away, which is consistent with what we see running our own AI Arena pipelines: the model produces candidates, the human decides.
A week of bid work in 48 hours
A hydropower bidding team used specialised bid-assembly agents to cut proposal preparation from a week to 48 hours. On a seven-day week that is a 71% reduction; on a five-day working week it is 60%. Both are large, and both depend on how much of the week was actually spent assembling text versus negotiating commercial terms, which the case study does not break down.
Why the control tower became the product
Agent 365 is Microsoft's governance layer, and in this deployment it functions as the register: who owns each agent, what it can access, what safety constraints apply. Once thousands of employees can spin up an agent in an afternoon, the risk stops being model quality and becomes bookkeeping. Unregistered agents touching customer data are the enterprise version of shadow IT.
For European organisations the bookkeeping is not optional. Agents that process personal data fall under GDPR regardless of which model runs underneath, and an agent drafting a stakeholder consent summary is processing personal data. You need a legal basis, a record of processing, retention rules and a way to answer an access request. A governance console helps produce that evidence, but it does not create the legal basis for you.
What EU rules already require
Since 2 August 2026, Article 50 transparency obligations under the EU AI Act apply. Synthetic content must carry machine-readable marking, and users interacting with a conversational system must be told they are doing so.
The high-risk deadlines moved. Annex III obligations, originally scheduled for August 2026, now apply from 2 December 2027, and Annex I embedded systems from August 2028, according to the current implementation timeline. For a consultancy doing infrastructure design, that delay matters: AI assisting the design of safety-relevant systems may land in Annex III scope, but not this year. What binds in 2026 is transparency plus the GPAI regime, which the European AI Office now enforces directly.
Microsoft's stack against the alternatives
The E7 bundle is the argument. Copilot, agents and governance arrive on one contract and one tenant. The trade-off is lock-in and per-seat cost — and the published list prices do allow a gross calculation. Twenty thousand seats at USD 99 per user per month is roughly USD 1.98 million a month at list. Licensing only the Agent 365 governance layer at USD 15 per user per month on top of an existing E5 and Copilot estate would instead come to about USD 300,000 a month for the same headcount.
Both figures are incomplete in the same direction. E7 replaces per-seat spend Mott MacDonald was already carrying for separate productivity, security and Copilot licences, so the net delta is smaller than the gross number — potentially much smaller. And these are Microsoft's USD list prices, not the five-year deal Mott MacDonald negotiated with 20,000 seats behind it. At the exchange rate we use for this conversion, USD 1.08 per euro, the two gross figures come to roughly €1.83 million and €278,000 a month respectively.
For firms with data residency constraints there are cheaper routes to the model layer. Mistral Large 4 weights have been downloadable since 29 October 2026 under the Mistral Ecosystem Open-Weight licence, a modified MIT licence, with API list pricing at USD 1.36 per million input tokens and USD 4.18 per million output tokens; self-hosting in an EU datacentre is possible, in euros, on your own hardware. Hosted Llama 4 APIs sit around USD 0.10 to USD 0.27 per million tokens. Neither gives you Agent 365. What you get instead is control of where the inference happens, which is the part European compliance officers ask about first.
Does the EU AI Act require companies to register internal AI agents?
Not by name. There is no EU-level agent registry. The obligation that pushes firms toward one is GDPR accountability: you must be able to show what processed personal data, on what basis, and for how long. A central register of agents is the practical way to answer that.
Is Agent 365 a separate licence from Microsoft 365 E7?
It is both. Agent 365 can be bought standalone at USD 15 per user per month at general availability, and it is also bundled inside Microsoft 365 E7 at USD 99 per user per month. The case study describes Mott MacDonald's shift as following its E7 adoption in July 2026.
Could a 200-person firm copy this model?
The pattern scales down more easily than the licence does. Letting staff build small agents is cheap; governing them is where the effort goes. Below a few hundred users, a spreadsheet plus strict data-access rules covers much of what the console does, at the cost of manual review.
The case study does not publish what Mott MacDonald actually pays for its E7 deployment beyond list, how many of the 8,000 agents are used in a given week, or how the 200 flagged conflicts were resolved. Those three gaps are the ones to ask about before treating the numbers as a benchmark.
