Enterprise AI Governance Gap Widens as Agent Conflict Resolution Lags, Report Finds

Ask a room of enterprise executives whether they can keep their AI agents in line and most will nod. Ask them what happens when two of those agents disagree, and the room goes quiet.

That, in essence, is the finding of a new report from OutSystems, the Lisbon-born low-code company that now markets itself as an “Agentic Systems Platform.” Produced with trade publication CIO Dive and consultancy KPMG, Engineering Business Outcomes: The Shift to Agentic Systems surveyed more than 850 IT decision-makers. Some 82 per cent said their organisations can keep pace with governing, integrating and monitoring AI agents as they move into production. Only 26 per cent said they were well prepared to resolve conflicts across agents, systems and automated workflows.

That 56-percentage-point gap is the story. Confidence is cheap. Conflict resolution is not.

What an agent conflict looks like

An AI agent, unlike a chatbot, does not just answer. It acts: it updates a record, triggers a refund, books a shipment. Put several of them inside one company and they start treading on each other’s toes. A procurement agent trims inventory while a sales agent promises next-day delivery. A risk agent freezes an account that a customer-service agent has just reactivated. Each is following its own rules. Together, they create a mess, and someone has to decide which action stands and how to unwind the other.

The report suggests most enterprises are not wired to make that call quickly. While 83 per cent rely on existing IT, security or compliance processes, specialised reviews or centralised policies to oversee agents that can act autonomously, just five per cent have automated controls, audit trails and formal human review points in place. Oversight, in other words, is still largely a person reading logs after the damage is done.

Southeast Asia has seen this film before

None of this will surprise anyone tracking enterprise AI in the region. In June, research by Sinch found that Singapore led Asia Pacific in deploying AI agents and also in pulling them back after going live, with 82 per cent of the city-state’s enterprises reporting a rollback or shutdown. In August, a Sumsub and Singapore Fintech Association study found that 94 per cent of Singapore businesses were using or piloting multi-step AI systems, yet only 29 per cent could produce an audit trail for AI-driven decisions.

The engineering floor is warier than the boardroom. The Agoda AI Developer Report 2026, which polled more than 800 developers across Southeast Asia and India, found that 79 per cent require human approval before AI-generated work goes live. Agents are trusted, but not with production. Thailand’s SCB 10X even caught agents confidently declaring tasks complete when the underlying code had failed its specifications.

Regulators have taken note. Last week, Singapore’s Infocomm Media Development Authority (IMDA) unveiled a starter kit for testing agentic systems and an accreditation scheme for third-party AI testers, effectively building an industry whose job is to check the agents. One of the kit’s metrics, handoff accuracy, measures whether an agent passed a task to the right agent or human. That is precisely the seam where the OutSystems respondents admit they are weakest.

The data problem underneath

Agents collide most often when they are working from different versions of the truth. More than a third of respondents (35 per cent) cited data consistency and context sharing as a major orchestration hurdle, which is a polite way of saying agents are too often running on generic model knowledge rather than governed company data and live application states.

Singapore has its own version of this headache. A Confluent survey in July found that 78 per cent of the city-state’s IT leaders flagged a lack of real-time data infrastructure, leaving agentic ambitions stuck on legacy systems and data quality gaps.

Buyers also want an exit door. Some 74 per cent prioritise integration with existing applications, 63 per cent want agent orchestration across systems, and 62 per cent value the ability to switch models, vendors and platforms over time. After a decade of cloud lock-in, enterprises are in no mood to marry a single AI supplier.

Read the fine print

A dose of scepticism is warranted. This is vendor-commissioned research, and its central prescription, that governance must be built into the platform and enforced while agents run rather than reviewed afterwards, happens to describe what OutSystems sells.

Tiago Azevedo, the company’s CIO, makes the pitch plainly: “Safety comes from building control into the platform, not reviewing it afterward.” KPMG Managing Director Michael Harper argues that logging and procedural governance were designed for static software, and that enterprises now need policy-as-code and runtime guardrails. Both points are reasonable. Both also lead neatly to a purchase order.

The outcome data is more revealing than the quotes. Error reduction (64 per cent) is the leading measure of impact, followed by compliance and risk reduction, productivity and cost. Yet only 12 per cent say agentic AI has sped up the launch of new products, services or customer experiences. Agents are tidying up existing operations, not inventing new businesses. Meanwhile, 81 per cent rate testing and validating agent behaviour as very or extremely important when choosing a platform, which tells you how little they trust what is already running.

Context matters, too. Last year’s edition of the same study, based on 550 software executives, found 93 per cent of organisations building or planning custom agents. The ambition has arrived. The plumbing has not.

A crowded referee’s chair

OutSystems is hardly alone in offering to referee. Salesforce is pushing Agentforce, Microsoft has Copilot Studio, ServiceNow and UiPath are pitching orchestration layers, and low-code rivals such as Siemens-owned Mendix and Appian are making near-identical governance promises. Each claims its platform is the safest place for agents to work together. The 62 per cent of buyers who want the freedom to switch suggest few are ready to hand the whistle to any single vendor.

In Southeast Asia, banks are drawing the line in their own way, as UOB’s research shows: AI agents can move money, as long as someone is accountable. Even model makers are wrestling with the problem; OpenAI shelved its GPT-6.1 Astra model this month amid concerns over how honestly the agent reported its own work.

The harder question is who referees the agents at companies that cannot afford a KPMG engagement. A mid-sized retailer in Jakarta or a lender in Manila deploying off-the-shelf agents will rely on whatever controls its vendors ship, plus free templates such as IMDA’s. As AI governance moves from promises to proof, that is where the next failures, and the next regulations, are likely to emerge.

Eighty-two per cent of executives believe they can govern their agents. The other number, 26, is probably closer to the truth.

Source link