For years, enterprises have added more and more technology to increasingly complicated processes.
But as AI agents become capable of taking action rather than simply providing assistance, Chad Scheller, ServiceNow’s area vice president for global Microsoft business, believes it’s time to ask a more fundamental question: why does the process exist in the first place?
What changes have you seen that give you confidence enterprises are now ready for the shift to agentic AI?
The question has changed. A year ago, customers asked me whether the technology worked. Now they ask who is accountable when it does. I had a chief information officer (CIO) ask me how he would attest to an agent’s actions in an audit and who owns the identity that the agent runs under. Nobody builds an audit trail for a science project. That is a deployment question.
The other change is where the money sits. Agentic AI moved from the innovation budget to the operating budget. Once something is in the run budget, it has to work, be supported and have somebody’s name on it.
And frankly, the models stopped being the hard part. The bottleneck moved to data, permissions and process. Those are problems enterprises already know how to solve.
Where are you seeing the strongest business cases for the use of autonomous AI today?
IT and employee service, by a wide margin. High volume, well documented, already measured and contained risk. Rolls-Royce is the example I keep coming back to. It has achieved roughly 54 per cent help desk deflection and about 5,000 hours given back. That is not a demo, that is a profit and loss statement.
Customer service operations are right behind it, and the back office is catching up fast in finance and procurement.
The pattern is consistent. The work is repetitive, the source of truth is already digital and somebody already tracks cycle time. If those three things are true, agents work. Where I tell people to slow down is anything requiring novel judgment in a process nobody has ever measured. You will not be able to prove you won.
Are enterprises making a mistake by putting AI into existing processes? Should they instead be asking whether those processes should exist?
Yes, that is a real mistake, and it is everywhere. Most enterprise processes exist because software used to be dumb. Four approval steps because we could not trust a system to check three fields. If you point an agent at that process, you have automated the tax instead of removing it. The right question is what this process would look like if it were designed today, then build for that.
What is preventing enterprises from scaling AI agents right now?
The biggest barrier to scale is sprawl. Every function is standing up agents right now. Marketing has some, engineering has some and finance bought a tool. Nobody has a list. Non-human identities already outnumber human ones in most enterprises by many orders of magnitude, and each has credentials and access to something. That is a security and accountability problem before it is a productivity problem, and it is the thing that stops a successful pilot from becoming a programme.
What does an organisation actually need to have in place before it is ready for autonomous AI?
Four things. A system of record and a knowledge base that are actually clean, because agents inherit your data quality. An identity and permission model for non-human actors, with least privilege scoping. An inventory of every agent with a named human owner. And a baseline measurement of the process before you touch it.
That last one gets skipped constantly. If you cannot say what the process costs today, you will never prove the agent helped, and the funding dries up in the second year.
I would add one non-technical item. Someone senior has to own the business outcome, not the technology. If the AI program reports up through a lab, it stays a lab.
How can companies ensure measurable outcomes? What metric separates a successful deployment from an impressive demo?
Autonomous resolution rate. The percentage of work that gets completed end-to-end with no human touching it. That is the number. Seats deployed, prompts entered, agents built – those are vanity metrics, and they tell you nothing.
Pair it with the cost per resolved unit of work and track both over time. A demo proves an agent can do the task once. Production means it does it 10,000 times, and you know your exception rate.
Measure the exceptions as hard as you measure the wins. The failures tell you where the process is broken, and that is usually more valuable than the deflection number.
What does an enterprise architecture look like when you have hundreds of AI agents operating across it?
It stops looking like a stack and starts looking like an org chart. You still have your systems of record and your models underneath, but the interesting part is above them.
You need a registry, so you know what exists. An orchestration layer so agents can hand work to each other instead of every one of them being a dead end. A permission layer scoped per agent. Observability into how an agent reasoned, not just what it output. And a kill switch.
The workflow platform becomes the substrate. Agents are the workers. The platform is their management system. Companies that treat agents as features bolted onto individual apps will hit a wall around a few dozen.
How does the AI Control Tower help organisations govern an increasingly complex ecosystem of agents?
It gives you one place to see every agent in the enterprise, ours and everyone else’s, who owns it, what it is allowed to touch, what it did, and what it delivered.
The pieces that matter most to customers right now are risk frameworks mapped to the National Institute of Standards and Technology (NIST) and the EU AI Act out of the box, runtime observability so you can see how an agent reasoned and where it made a decision, identity and least-privilege enforcement across connected systems, and the ability to detect an agent operating outside its permissions and shut it down in real time.
That last capability is what unlocks the rest. Governance is not the brake pedal. It is the thing that lets a board approve putting agents to work.
Scheller (left) with the Microsoft and ServiceNow partner team at Knowledge 2026 (Photo: ServiceNow)
How does the ServiceNow and Microsoft partnership benefit your customers?
Microsoft is where employees work and where many of the models and infrastructure reside. ServiceNow is where work gets governed and executed end-to-end. Those are complementary, and we have spent the last year making that concrete rather than directional.
At Knowledge 2026, we extended AI Control Tower governance across Microsoft Foundry, Copilot Studio, and Agent 365, and our AI specialists are coming to the Agent 365 Marketplace so they run inside the tools employees already use. We are also taking a joint autonomous IT motion to market.
In practice, for a CIO, it means you do not govern your Copilot agents in one console and your ServiceNow agents in another. One inventory, one policy set and one audit trail across both.
There is a commercial piece people underrate, too. Customers can transact through Azure Marketplace and draw down existing Azure commitments. Procurement is often the slowest part of an AI strategy, which can remove a quarter or two.
At Knowledge 2026, ServiceNow announced its extended AI Control Tower governance (Photo: ServiceNow)
Read more like this, including news, interviews, case studies and insight from industry experts, in the Autumn 2026 issue of Technology Record. To get content like this on a regular basis, subscribe to the print edition or free digital edition of the quarterly Technology Record magazine.