Late in 2024, Satya Nadella made a remark on a podcast that business software vendors have been chewing over ever since. Paraphrased, his argument was that traditional business applications are essentially databases with business logic on top, and that in an agent-driven world much of that logic will move out of the application and into AI agents working across many systems at once. Coming from the chief executive of the company that sells Dynamics 365, it was an interesting thing to say out loud.

Nearly two years on, the ERP and CRM screens haven’t disappeared. But the direction Nadella described is visible in Microsoft’s own products, and it’s starting to change what businesses should expect from the systems that run their finances, sales and operations, and from the partners who implement them.

From forms to agents

For decades, business applications have been built around forms. A salesperson opens a record, types in the details of a call, updates a stage and saves. A finance clerk keys in an invoice, matches it to a purchase order and posts it. The software’s job was to store the data correctly and enforce the rules.

The first wave of AI in Dynamics 365 added help around those forms. Copilot features in Sales summarise account history before a meeting and draft follow-up emails. In Business Central, Microsoft’s ERP for smaller and mid-sized firms, Copilot has offered help with bank reconciliation, suggested sales lines and generated product descriptions. Useful, if incremental.

The second wave is more ambitious. Microsoft has been introducing agents that take on whole tasks rather than assisting with individual fields. In Business Central, a sales order agent can read a customer’s email, check stock and pricing, and prepare a draft quote for a person to review and send. On the CRM side, agents are handling lead qualification, drafting responses to routine service cases and chasing suppliers about late deliveries. The pattern is consistent: the agent does the legwork across several records and systems, and a human approves the result.

What changes for the business

The obvious benefit is time. Order entry, case triage and supplier follow-up are all tasks with high volume, predictable structure and low tolerance for delay, which is exactly where agents perform best. A small distribution business that handles hundreds of emailed orders a week has a real incentive to let an agent draft them.

The less obvious effect is on data. An agent is only as reliable as the records it reads, and ERP data in many businesses is less tidy than anyone likes to admit. Duplicate customer accounts, stale price lists, inconsistent units of measure and half-finished item records don’t matter much when a human is keying orders and quietly correcting as they go. They matter a great deal when an agent is doing it and nobody is looking closely at each line.

Process design changes too. When a person handles a task, exceptions get resolved through experience and a quick word with a colleague. An agent needs those exceptions spelled out, or at least needs to know when to stop and hand over. Businesses that have never written down how they handle a partial shipment or a customer on credit hold are about to find out how much of their process lives in people’s heads, and whether they have implementation partners who know the Dynamics platform well to help write it down.

Choosing a partner for the agent era

All of this changes what a good Dynamics implementation looks like, and therefore who should do it. The traditional selection criteria still apply: industry experience, a sensible methodology, references from businesses of a similar size. But they’re no longer enough.

The partners worth shortlisting now treat data migration and data quality as core work rather than a line item squeezed in before go-live. They should be able to talk comfortably about Dataverse, Power Platform and Copilot Studio, because that’s where much of the agent capability is built and extended. And they should be willing to map processes in enough detail for an agent to follow them, including the awkward exceptions. A quick test is to ask how they’d prepare your data and processes for agents, and listen for specifics.

It’s also worth asking about the long run. Microsoft releases updates to Dynamics 365 and Business Central continuously, with two major release waves a year, and AI features are moving faster still. An implementation partner that disappears after go-live leaves you to work out which new agents are worth switching on and how to govern them. One that stays involved, reviewing new capabilities against your processes, is likely to deliver more value over time.

The open question is how far Nadella’s prediction will run. If agents increasingly work across systems, drawing on Dynamics, Microsoft 365 and third-party applications in the same task, the value of any single business application starts to shift from its screens to its data model and its APIs. That’s a good outcome for customers with clean, well-structured data. For those without it, the agent era may start with a data clean-up project nobody budgeted for.