82% of Executives Think They Can Govern Their AI Agents. 26% Can Handle Two of Them Disagreeing.

Ask executives whether they can govern their AI agents and 82% say yes. Ask the same people whether they are well prepared to resolve conflicts across agents, systems and automated workflows and the number falls to 26%.

Those figures come from research OutSystems published this week with CIO Dive and KPMG, based on a survey of more than 850 IT decision makers conducted in June. The gap between them is the most useful thing anyone has published about agents this month, because it points at a problem almost nobody is preparing for.

Everyone is braced for the wrong failure

When people worry about agents, they picture one agent going wrong. It hallucinates, it exceeds its permissions, it does something nobody sanctioned. That is the scenario the entire safety conversation is built around, and the industry has responded with sandboxes, kill switches, approval gates and audit trails.

All of that is necessary. None of it helps with the failure the 26% figure describes.

Picture two agents, both working exactly as designed. Your support agent receives a complaint, checks the policy, and issues a refund. It is correct. Your finance agent runs its daily pass, finds an order marked paid, and reconciles it. It is also correct. Neither has misbehaved. Neither has exceeded its permissions. Nobody wrote down which of them wins when they touch the same record forty seconds apart, because when each was deployed it was the only agent in the room.

That is not a model problem, and no model release will fix it. It is a question of precedence between two systems that were each specified in isolation.

Why this arrives suddenly

Conflicts of this kind do not scale linearly. One agent has no one to disagree with. Two agents have one relationship to get wrong. Five have ten. The first agent in a company is a pilot and the fifth is an operating environment, and most organisations pass through that transition without noticing it happened.

The research shows the same pattern in how people describe their own position. 36% report a fragmented approach to deciding where and how to use agentic AI, which is what it looks like when agents are adopted department by department rather than planned as a set. Each one is sensible on its own. Nobody owns the seams between them.

You cannot resolve what you cannot see

The measurement numbers in the same research explain why conflicts stay invisible until they are expensive.

35% say they struggle to measure the effect of agents across multistep workflows. 31% lack reliable baseline data. 31% have limited visibility after deployment.

Put those together and you get an organisation that can tell you an agent ran, and cannot tell you what changed as a result. In that state a conflict does not announce itself. It shows up weeks later as a reconciliation that will not balance, a customer who was refunded twice, or a record that two systems disagree about, and the investigation starts from scratch because nothing recorded who touched what and in which order.

The honest version of the 26% number is not that three quarters of companies would handle a conflict badly. It is that most of them would not find out a conflict had occurred.

What a conflict rule actually looks like

This is cheaper than it sounds, and the work is clerical rather than technical. For any two agents that can touch the same thing, you need three decisions.

Who has precedence on this object. If both your support agent and your finance agent can modify an order, one of them wins by default and the other defers. Write down which, and when.

What constitutes a contested action. Most actions are not contested. The ones that are tend to involve money, customer-visible communication, or a status that other processes read. Name those specifically rather than trying to govern everything.

What happens on a collision. The three workable answers are: the second agent stops and asks a human, the second agent defers to the first and logs it, or the action is queued until the first finishes. Any of these is fine. Having none is the problem.

Then make sure the log records the actor, not just the action. A line saying the order was refunded is nearly useless. A line saying which agent refunded it, acting under whose authority, and what the record looked like before, is the difference between a five minute explanation and a two day investigation.

The part worth copying from KPMG

Michael Harper of KPMG put it better than we could: assurance for agentic systems has to be engineered from day one, because you cannot inspect it afterwards.

That is the same argument we made about evaluation criteria a fortnight ago, when we looked at why 88% of agent pilots never reach production. Define what done means before you build, because inventing the definition retrospectively is a different and much harder job. Conflict rules work the same way. Deciding precedence between two agents while both are already running in production, with a disputed record in front of you, is the worst possible moment to be having that conversation for the first time.

One more number, for perspective

Only 12% of these organisations say agentic AI has accelerated the launch of new products, services or customer experiences.

That is worth sitting with, because it tells you where almost everyone actually is. The transformation stories are still mostly stories. What is real today is error reduction, which 64% name as their leading impact metric, and that is a perfectly respectable thing for a technology to be doing in its first couple of years.

It also explains the confidence gap. When your agents are reducing errors in separate departments, governance genuinely does feel manageable, because each one is doing a contained job. The 26% is a preview of what happens when those contained jobs start overlapping.

The practical version

If you run one agent, you have time. Use it. Write down what it may touch, what it must ask about, and what the log shows afterwards, while the answer is still obvious.

If you run three or more, the seams already exist whether or not anyone owns them. The question to put to your team this week is a simple one: if two of our agents acted on the same record within a minute of each other, how would we find out?

If the answer is that a customer would tell us, you have your project.

Want to test the most advanced AI employees, with scoped permissions and an audit trail that records who acted? Try it here: https://geta.team