A woman oversees a glowing AI network linking robots, infrastructure, security, and sustainability dashboards in a modern office.
Arcadis has a lot of agents. The harder question is how it keeps track of them. Microsoft's customer story says the engineering and consultancy firm has built more than 1,000 agents in Copilot Studio. It also describes a client-facing tool, a governance stack and several unfinished projects. This piece sets out what the story supports and where its claims need caution. It is a vendor-published case study, so the figures are company-reported and have not been independently checked.

A woman oversees a glowing AI network linking robots, infrastructure, security, and sustainability dashboards in a modern office. The scale: what Arcadis says it has built​

Arcadis employs 34,000 people in more than 30 countries. Its 2027–2029 strategy, which Microsoft's story calls "Reset: Focused Growth, Sustainable Value," puts AI and digital innovation at the centre. Premium Power Platform licenses are available to all employees. That includes the premium connectors that link apps and agents to business data.

The reported estate is large:

  • More than 25,000 Power Automate cloud flows.
  • More than 10,000 agents built with Agent Builder.
  • More than 1,000 agents built in Copilot Studio.

The story doesn't define what counts as an agent. It doesn't say whether these are active or deployed, or over what period they built up. Treat the counts as evidence of enthusiasm and tracking effort. They don't measure production value.

Governance: the part worth stealing​

Arcadis's most transferable lesson is how it manages the estate. The story lists these elements:

  • Digital & AI Hub. This is the intake and discovery point. It shows which solutions already exist so teams reuse them. It also keeps a record of each solution, and Arcadis is extending that record to the whole estate. Arcadis is also extending the Hub's intake to agents, so reuse is checked before anything new is built.
  • Managed Environments. These apply the same rules across Power Platform environments.
  • Copilot Agent Kit. Arcadis says it helps makers build, test, govern and improve custom agents.
  • Community and training. Makers hold regular calls to share use cases and design principles. AI development and governance training is being added, with a path from basic Copilot skills to advanced use.

Phillip Brown, Global Low-Code Development Platform Director, says the kit's tracking capabilities give traceability into how each agent is designed, tested and operated. He calls that the evidence its GDPR and EU AI Act obligations call for. That is Arcadis's own view of its compliance approach. Nothing in the story says a toolkit alone makes an organization compliant.

What the Copilot Agent Kit does and what it needs​

Microsoft Learn describes the kit as an open-source toolkit for testing and evaluating Copilot Studio agents. It helps you define test cases, simulate user queries and validate responses. The wider feature set is governance-oriented:

  • Agent Inventory. It gives a tenant-wide view of all Copilot Studio custom agents. That view includes the features they use, authentication mode, knowledge sources and orchestration type.
  • Compliance Hub. It evaluates agent configurations collected through Agent Inventory against configured controls. Administrators can set thresholds, risk levels, SLA timers and enforcement actions (manual review, quarantine, delete). Makers are guided to remediate through Teams or Outlook notifications. Compliance Hub requires Agent Inventory.

The documented prerequisites matter if you're thinking of copying this model:

  1. A Power Platform environment with Dataverse.
  2. The system administrator role to install the kit.
  3. Power Apps code components and Code Apps enabled in the environment.
  4. Licensing for a model-driven Power App and for Power Automate cloud flows with premium connectors.
  5. The Power Platform Creator Kit deployed first.

Connector data policies must allow every connector shipped in the solution. The connections must be populated at import time, whichever features you plan to use. Some features consume AI Builder or Copilot credits. Microsoft's testing puts an average of about 50 credits on each Generative Answers test case. Azure Application Insights is needed for extra telemetry on AI-generated answers.

Installation is possible from Marketplace or by importing the managed solution from GitHub. One caveat applies to GitHub upgrades. Microsoft documents a known import error that mentions the cat_copilottestrunid attribute, and the suggested workaround is to stage the upgrade first.

The client case: Assurance AI across 200+ sites​

Arcadis's main client example is Assurance AI. It grew out of an assurance project for a major manufacturing client. The goal was consistent audit, inspection, testing and maintenance processes across more than 200 sites. The tool is meant to help users analyze evidence, spot emerging risks and recurring issues, and intervene earlier.

The architecture is the most useful technical detail in the story:

  • A Client Risk & Advisory orchestrator agent was built with Copilot Studio multi-agent orchestration.
  • It doesn't query one consolidated source. It routes questions dynamically to child agents, each focused on a specific dataset.
  • Those datasets are fetched from Dataverse using the Dataverse MCP server, according to the page's diagram description.
  • Dataverse is described as the layer connecting governed data to agents, with controls, traceability and human oversight.

Splitting data across child agents is a common pattern for keeping each agent's scope narrow. That is general industry knowledge, not something the story tests.

The numbers, read carefully​

Microsoft reports two separate results:

  • Standardizing the assurance process had already cut the client's compliance costs by 25%.
  • Assurance AI then cut the time Arcadis's team spends analysing site evidence by 40%.

The 40% is a time saving on Arcadis's own analysis work. It isn't a cut to the client's costs, and it isn't a measure of accuracy. No baseline, period, sample size or method is given for either figure. Notably, the cost reduction came from process standardization before the AI arrived. That supports Murthy's stated approach: start with the process, build a trusted data foundation in Dataverse, then build agents.

Facility Intelligence is still in development. It is meant to connect site data, audit outcomes, corrective and preventive actions, incidents and operational signals. It has no demonstrated results yet.

Employee-facing work​

  • Reimagining Arcadis Q&A agent (live). It is built in Copilot Studio and helps staff find program-approved answers about the company's transformation in Teams or a browser. Developer Chen Sheng says teams keep its knowledge current and review feedback. The aim is to support human communication, not replace it.
  • Account review agent (in development). It is built with Foundry and Copilot Studio and is planned for Teams. It would assemble a preparation pack before a project or account review. Afterwards it would analyze the meeting transcript and propose actions, risks and follow-ups. The team can approve, edit or reject these before anything is saved.
  • Microsoft 365 Copilot and Copilot Chat. All employees have access. Arcadis is exploring Copilot Cowork, which plans and carries out multi-step tasks across Microsoft 365 with approval checkpoints. That is exploration, not a finished rollout.

What IT administrators can take from this​

  • Inventory first. Olivieri says Arcadis has tracked its estate for years. Visibility into what people build, what gets reused and where to invest is the base for everything else.
  • Reuse before build. Putting agents through intake before development is a cheap way to cut sprawl.
  • Data layer before agents. The story ties trust in AI output to clean, governed Dataverse data.
  • Keep humans in the loop. The account review agent and Cowork both include approval checkpoints.
  • Be skeptical of headline numbers. A vendor customer story picks its best results. Ask for baselines and methods before using these figures in your own business case.

Bottom line​

The Arcadis story is more useful as a governance blueprint than as a performance claim. The pieces are a managed data layer, a reuse-first hub, Managed Environments and a toolkit for inventory and compliance. A multi-agent client workflow is built on top of them. The 40% and 25% figures are company-reported and unaudited. Facility Intelligence, the account review agent and Cowork are all still unproven.

 

References

  1. Arcadis scales client delivery with Microsoft Copilot Studio and Power Platform - Microsoft Microsoft Fri, 20 Nov 2026 11:05:54 GMT
  2. Copilot Agent Kit marketplace.microsoft.com
  3. Install Copilot Agent Kit - Microsoft Copilot Studio learn.microsoft.com