The Microsoft Business Applications team has been running a particular demo for about a year now. A sales rep gets a Teams notification. An agent has noticed that a deal in their pipeline has gone quiet, cross-referenced the contact’s recent emails, found a question the customer raised that nobody answered, drafted a response, summarised the deal history, and put the whole thing in front of the rep with a recommended next step. The rep clicks accept. The deal moves. Total elapsed time: about eleven seconds.
It’s a good demo. It’s also nearly nothing like what the same agent does on its first day inside a real Dynamics 365 tenant.
Anyone who’s watched a real agent deployment knows what happens next, and it’s the kind of detail that doesn’t make it into a keynote. The agent runs. It finds twelve emails it can’t parse because they’re in a shared mailbox the agent doesn’t have permission to read. It hallucinates a customer name because the account record has two contacts both called Sarah and no clear primary. It suggests following up on a deal that was actually closed eight months ago and never archived. The sales manager, watching this happen in a UAT environment, asks the consultant the question every sales manager asks: “Is this going to embarrass us in front of a customer?”
Whether the answer is yes or no depends almost entirely on work that has nothing to do with the agent itself.
The dirty secret of the current AI-in-business-applications moment is that the agents are the easy part. Microsoft has done the hard engineering. Copilot Studio is, by the standards of agent platforms in 2026, genuinely capable. The reasoning is solid. The action-taking works. The integration story across Dynamics, Power Platform and the rest of the Microsoft estate is the best in the industry, which is faint praise but accurate. If you have a clean tenant with clean data and clean permissions, you can stand up an agent that does something useful in an afternoon.
Almost nobody has a clean tenant.
What Dynamics 365 and Power Platform delivery teams are actually doing for their clients in 2026 is the unglamorous work that has to happen before an agent is worth deploying. They are auditing security roles that have accumulated for eight years and now grant random users access to things nobody remembers granting. They are de-duplicating account records, reconciling the four different ways the company has spelled “BT Group” since 2019, and working out which custom fields are still in use and which are vestigial. They are writing the governance documentation that determines what an agent is allowed to email a customer without human review, which is a question most organisations have never had to answer in writing before. And they are doing user-acceptance testing of a kind that didn’t exist three years ago, because “does the software produce the correct output” is no longer a sufficient question when the software is making judgement calls.
This work is profoundly tedious. It is also where most of the value comes from. The agent doesn’t get smarter when you do it; it gets safer, and safe is what makes the difference between a tool that ships and a tool that gets pulled out of production after the first incident.
There’s a useful pattern emerging in how the better implementations handle this. The agent gets deployed against a deliberately narrow slice of the business first, one team and one workflow, and the team’s job for the first six weeks is to break it. Not to use it. To break it. To find the edge cases where it makes a decision the business would not have made. The implementations that go straight into broad rollout, on the strength of how well the demo went, are also the ones generating most of the cautionary stories doing the rounds in private CIO forums.
See also: Can SEO Be Guaranteed? The Truth Every Business Should Know
The optimistic version of this story is that the work being done now — the data cleansing, the role auditing, the governance writing — will pay compound dividends. A Dynamics tenant that’s clean enough to host a useful agent is also a tenant that’s measurably easier to migrate, integrate and audit. The organisations putting in the boring work this year are building a foundation that the ones chasing demo-quality results will eventually have to retrofit, at considerably greater expense.
The agents are coming. Most of them will be quietly disappointing for a couple of years, the way most enterprise software is quietly disappointing for a couple of years after the keynote. The handful that aren’t will be running on top of a Dynamics estate that someone, somewhere, spent six unsexy months getting ready for them.
