The story comes from Microsoft UK Stories, so it is a vendor-published case study. The figures are the company's own and haven't been independently audited. It is still a useful look at what happens after an AI pilot succeeds.
The rollout in numbers
According to Microsoft UK Stories, the engineering consultancy's Copilot journey went through these stages:
- It started with 50 licences after Copilot arrived in 2025.
- It then grew to 1,000 licences and then 5,000. Half of the new licences in the second stage went to early-career professionals.
- Copilot is now rolling out to all 20,000 employees under a new five-year agreement. The company describes this as a major investment in Microsoft 365 E7.
- Staff have created more than 8,000 agents. They cover project assurance, stakeholder engagement, design automation and data-driven decision-making.
The story gives no contract value, start date or licence mix for the five-year deal. Microsoft's published E7 list price is $99 per user per month in the US. That is a list price and shouldn't be read as what Mott MacDonald pays.
Global Head of AI and Data Nasrine Tomasi described a "reverse mentoring" effect. Younger staff who were already comfortable with generative AI learned from experienced engineers, and showed them what the tools could do. Many have become AI champions. The company's lesson was to give people access and let them apply their own expertise, instead of predicting in advance where AI would help.
What the agents actually do
The article offers three concrete examples:
- Requirements review: On an airport project, AI helped work through about 220,000 requirements and flag about 200 conflicts. The article doesn't name the airport, the method or how the flags were validated.
- Design analysis: On a large underground station, structural designers used Copilot-supported analysis to show that temporary works could safely become part of the permanent structure. This cut the extra construction needed. The wording is "Copilot-supported analysis", so engineers did the engineering and AI did not approve anything.
- Bid preparation: A hydropower team built agents that assemble requirements, calculate budgets and match capabilities. The team has set a target of cutting its proposal process from one week to 48 hours. That is a target, not a reported result. Tomasi says the team is winning more work, but the article gives no figures to back that up.
EMMA moves into workflows
EMMA stands for Every Mott MacDonald Answer. Microsoft's November 2025 profile described it as an internal assistant built on Azure AI Search and Azure AI Foundry technologies. It draws on company knowledge bases to answer questions on procedures, experts and compliance.
It is now moving beyond lookup. In a Teams-hosted app, EMMA captures a structured conversation between project managers and project principals for monthly reviews. It organises the notes around an agreed agenda and highlights lessons learned and future actions. The article reports more than 10,000 lessons captured since the app launched in May. It gives no usage denominator or quality assessment.
The sprawl problem
Success brought clutter. Tomasi says the 8,000-plus agents include overlapping solutions, dormant agents and knowledge gaps when creators leave or change roles. The company's response has two parts:
- A central AI library. Staff can check whether an agent already exists before building a new one, and the company can spot ideas worth scaling.
- Agent 365 oversight. The company needs to know who owns each agent, what data it can access, how it is used and whether it still delivers value. Tomasi calls Agent 365 a future "control tower".
The story presents Agent 365 as a planned capability. It does not say the product is fully deployed at Mott MacDonald.
What Agent 365 offers, per Microsoft
Microsoft's 1 May 2026 announcement says Agent 365 reached general availability that day. It is available in Microsoft 365 E7 or standalone at USD15 per user per month. Each licence covers an individual who manages or sponsors agents, or uses agents to do work on their behalf. Microsoft's blog groups the capabilities into observing, governing and securing agents. They include visibility into usage, onboarding agents with IT oversight, governing access to data, and securing agent identities.
Independent write-ups describe the registry as the foundation. One says the first thing most organisations discover when they enable it is that the registry contains agents they did not know were there. The same write-up notes that lifecycle flags still need a human decision. That means assigning a new owner or retiring the agent. That matches Mott MacDonald's problem of agents orphaned when their creators move on.
Admins should also check what is still in preview. Microsoft announced a public preview of registry sync with AWS Bedrock and Google Cloud connections. A third-party guide advises against building production governance on it until it is generally available. If your agents span clouds, that matters.
What other IT teams can take from this
This is a vendor story, so treat it with some scepticism. The strongest lesson is the governance sequence, not the productivity claims.
- Plan for inventory early. If thousands of agents appear within about a year, ownership and lifecycle become urgent.
- Make discovery easy. A library of reusable agents reduces duplicates and shows what is worth scaling.
- Tie agents to data access. The question is not just what an agent does, but what it can read.
- Keep humans accountable. Mott MacDonald says its adoption programme trains staff to review outputs critically, spot weak answers and know when human judgement takes over. The article offers no formal review policy or validation benchmarks.
- Prepare the data layer. The company is preparing to use Microsoft Fabric as a central data platform so Copilot and agents can reach better information. The article says this is still being prepared, and gives no architecture or timing.
Caveats
The case study doesn't show that AI made the engineering safer or the business more profitable. It doesn't report error rates for the airport requirements work, or how many of the 8,000 agents are active. Those gaps matter for anyone using it as a benchmark.
Still, it shows what scaled agentic AI looks like in a real enterprise. Licences and enthusiasm come first, and a registry, ownership model and human review follow close behind.