By Sekhar Ramarao
.jpg)
For the last few years, most people's experience of AI inside enterprise applications has been the copilot.
You ask for a summary, get a recommendation, draft a reply, or find something faster. The work gets quicker, but you still decide what happens next.
That line is starting to move.
In business applications like Microsoft Dynamics 365, AI is moving from helping users to actually taking part in how business processes run. Agents can now pick up an event, work with live operational data, kick off tasks and push routine work forward across sales, service, finance and supply chain.
That's a bigger shift than it sounds. AI stops being a productivity layer sitting around the process and becomes a participant inside it.
So the question I keep hearing from leadership teams, "How many of our people should have access to AI?" is already the wrong one. The better questions are: where should AI be allowed to move work forward on its own, where must human judgment stay in charge, and how do you increase autonomy without losing control of the process?
In my experience, that makes the next phase of enterprise AI as much a process-design problem as a technology one.
Copilots have already proven their worth. They cut the effort involved in searching, analysing, drafting and responding.
Agents are a different proposition. They don't wait for someone to start every interaction. An agent can respond to a business event, work out the next step and carry it out within the applications and rules that already support the process.
Microsoft is clearly heading this way with Dynamics 365. Its 2026 roadmap brings agentic capabilities into sales, customer service, finance, supply chain, HR and more. Customer Service is getting both semi-autonomous and fully autonomous workflows, and Business Central already has agents that can take on work like accounts payable, with a human keeping watch.
What changes here is the unit of transformation. With a copilot, you ask how AI can make an employee more productive. With an autonomous workflow, you have to ask how to redesign the work itself.
The market is already feeling this. Deloitte's 2026 research found that 74% of leaders surveyed expect nearly half their business processes to be redesigned or rebuilt around AI agents within four years. Only 5% said their processes were highly prepared for agents.
That gap is the real story. Dropping agents into processes built around manual hand-offs, scattered data and old approval chains won't give you an autonomous enterprise. More often than not, the process has to change before the technology can deliver.
One trap I see is treating autonomy as the finish line every process should eventually cross.
Real operations don't work like that.
In some processes, AI is most useful for gathering information and preparing the next step for a person. In others, it can weigh up the context and recommend a decision. Where outcomes are predictable, and the boundaries are well understood, it can start carrying out approved actions. Only once a process has proven itself reliable should routine execution move towards full autonomy.
These aren't just different levels of AI sophistication. They are different operating models.
A few examples:
Dynamics 365 is increasingly able to support all of these patterns. Microsoft describes its agents as working inside the transaction layer, using the same data, business rules, permissions and audit trails the applications already rely on.
The practical upshot is that you don't have to choose between manual work and full autonomy. You can give each process the degree of autonomy it has earned, based on what's at stake, how predictable it is and what the evidence shows.
Traditional automation programmes usually started by hunting for high-volume, repetitive tasks. That still helps, but agentic AI needs a sharper filter.
A good candidate for more autonomy usually has a clearly defined outcome, dependable data, business rules you can write down, and exceptions you can spot before they cause damage.
The cost of getting it wrong matters just as much. Resolving a standard service request and committing the company to an unusual financial or contractual position are not the same thing, even if an agent is technically capable of both.
That's why I'd encourage teams to look at the individual decisions and actions inside a Dynamics 365 process, rather than labelling a whole function "suitable" or "unsuitable" for AI.
Take procurement. Much of it can be heavily automated: interpreting a requirement, pulling approved supplier information, comparing quotations, checking inventory and preparing the transaction. But an unusual commercial condition, or an approval above an agreed threshold, can still come back to a person.
Customer service works the same way. Routine cases with known resolution paths can gradually move to autonomous handling, while complaints, sensitive customers and anything ambiguous stay with people.
The goal isn't end-to-end autonomy for its own sake. It's taking people out of the parts of a process where their judgement adds little, and keeping them firmly in the parts where it matters.
Putting a person in front of every AI action feels like the safe option. At scale, it creates a different problem.
If someone has to approve hundreds of low-risk actions that follow the rules every single time, the approval becomes a reflex. The human is technically in the loop but adds almost nothing.
A better model puts people where their involvement actually changes the outcome: unusual financial exposure, regulatory interpretation, sensitive customer situations, significant contractual commitments, ethical questions, or anything that can't be undone.
For routine scenarios that consistently produce the right result, people can gradually step back from doing the transaction and move into supervising and handling exceptions.
Leaders seem to agree. In the same Deloitte research, 75% said collaboration between people and AI agents creates more value than agent-driven automation on its own.
So people don't disappear from Dynamics 365 processes as agents get better. They move up the decision chain, spending less time pushing routine work through the system and more time on ambiguity, exceptions, relationships and the judgment calls where business context really counts.
Agent demos almost always show the perfect transaction. Real operations don't stay perfect for long.
A customer record conflicts with another system. A supplier's details are incomplete. Stock changes halfway through an order. An integration goes down. An approval expires. A transaction lands just outside the expected threshold.
These aren't edge cases to deal with after go-live. They are part of the design.
An enterprise-ready workflow must know when it lacks information, when an action is outside its boundaries, or when something it depends on has failed. Then it needs a controlled way to pause, retry, ask for more information or hand the case to a person, with the full context intact.
This is where the business application layer earns its place. Agents need more than intelligence. They need reliable access to the operational context the business actually runs on: customers, orders, cases, inventory, suppliers, financial transactions, service history, rules and approvals.
Microsoft's agent-ready Dynamics 365 strategy is built around exactly this. The company says Dynamics 365 now exposes more than 650,000 Model Context Protocol actions across sales, finance, supply chain, HR, field service, customer service and project operations, so agents can work with the application while keeping the transaction context intact.
Put simply, an agent is only as good as its environment.
Trust in AI is often discussed as if it were a philosophical choice: do we trust agents or not?
For business processes, trust can be far more concrete. You can measure how consistently an agent completes its task, how often someone has to step in, which exceptions trip it up, whether outcomes stay within policy, and whether the process actually gets faster, cheaper, better or more pleasant for customers.
That evidence is what should drive decisions about giving an agent more room.
Picture a customer service process in Dynamics 365. At first, the agent analyses a case and recommends a resolution to an employee. Once you've seen how it performs across enough real cases, selected categories move to controlled execution. Proven, low-risk cases then become candidates for fully autonomous resolution, while sensitive or unusual ones keep getting escalated.
The same logic applies in finance, supply chain, sales and elsewhere. Autonomy is something a process earns through performance.
It should also be reversible. If conditions change, exceptions rise or outcomes slip, you should be able to narrow the agent's scope and hand more decisions back to people.
Gartner's 2026 research makes a similar point. It warns against governing agents as either fully locked down or fully trusted, and argues that controls should match each agent's level of autonomy and scope. It also predicts that by 2027, 40% of enterprises will demote or decommission autonomous agents after production incidents expose governance gaps.
The lesson is simple: more autonomy should be the result of evidence, not the starting assumption.
As agents start doing real work inside business applications, you need to see what they're doing and what outcomes they're influencing.
Governance, auditability and access control are still essential, but not only for security. They generate the evidence that tells you whether an agent is behaving as intended and whether a process is ready for more autonomy.
In Dynamics 365, agents work against the same business data, rules, permissions and audit trails as everything else, rather than in a separate environment. Microsoft's business-process architecture also pairs autonomous agents with deterministic workflows for activities that must stay predictable and governed.
For leaders, the basics are clear ownership of every agent-enabled process, defined operating boundaries, and enough traceability to reconstruct any significant decision or action.
Identity and permissions are part of that foundation, but they can't be the whole strategy. The bigger issue is keeping technical authority aligned with process authority.
An agent may be able to execute an action. The business still has to decide whether it should.
Moving from copilot to autonomous workflows doesn't have to happen in one big transformation programme. A more practical path starts with the process.
First, find where AI can remove information friction and make people more productive. Then introduce recommendations where AI can add useful judgement while people stay accountable for the decision.
As performance becomes measurable, move selected activities into controlled execution. The agent acts within well-understood scenarios; people keep the exceptions and the consequential calls.
Full autonomy only makes sense once you have enough evidence that the process, data, controls and exception handling can support it reliably.
This path also avoids a common mistake: measuring progress by the number of agents deployed. Gartner's current advice to CIOs points the same way: redesign the work before automating it, and measure outcomes and trust together instead of treating agents as standalone technology assets.
For Dynamics 365 customers, that gives a far more useful goal. Success isn't an agent in every department. It's less wasted effort, shorter cycle times, better decisions, and human expertise focused where it adds the most value.
Dynamics 365 is evolving from AI-assisted applications into an agent-ready business platform, and that changes what transformation can look like.
Microsoft now positions it as delivering agentic capabilities across ERP and CRM, with AI able to analyse data, automate tasks and support real-time decisions across sales, service, finance and supply chain.
The technology is moving fast. Operating models now need to catch up.
Organisations need to work out where autonomous execution genuinely improves a process, which workflows need redesigning, how exceptions get back to people, and what results will show an agent is ready for more responsibility.
This is the conversation we are increasingly having with customers at Intertec. It goes well beyond switching on AI features. As a Microsoft Business Applications partner, we help enterprises find the processes with the strongest case for agentic transformation, connect AI to the right Dynamics 365 business context, redesign workflows around people and agents working together, and introduce autonomy step by step, within the right controls.
It matters most in processes that span customer engagement, finance, supply chain and operations, where the value of AI depends less on how clever the agent is and more on how well it works with the applications, data, approvals and operating models that already run the business.
The next phase of Dynamics 365 transformation won't be defined by simply moving from copilot to agents.
It will be defined by how wisely enterprises choose which parts of the business move with them.
----------------------------------------------------------------------------------------------------
Sekhar Ramarao leads the Microsoft Business Applications practice at Intertec Systems. He has 28 years of experience across ERP, CRM and enterprise AI in Government, Healthcare, Real Estate and Construction, working with clients across India and the Middle East.
