- Over 40% of agentic AI projects will be cancelled by the end of 2027, largely due to inadequate risk controls, according to Gartner (2025).
- Gartner predicts that at least 15% of daily work decisions will be made autonomously by agentic AI in 2028, up from 0% in 2024.
- According to Statistics Canada (Q2 2026), 19.2% of Canadian businesses use AI, with the rate reaching 40.4% in finance and insurance.
- PlanAxion governs every AI agent in finance with four controls: a written scope of action, human approval thresholds, an audit log, and exception review.
It’s Friday at the end of the month at a distributor. Overnight, an AI agent matched 312 payments, posted the entries, and closed the batches. Monday morning, the controller asks a simple question: who approved all of this?
No one has a clear answer. The agent is working, the numbers balance, but no document specifies what it is authorized to do on its own. This is exactly the gap that AI agent governance fills.
“More than 40% of agentic AI projects will be cancelled by the end of 2027 due to spiralling costs, unclear business value, or inadequate risk controls.” Source: Gartner, June 2025
What is AI agent governance?
AI agent governance is the set of rules that define what an agent can do on its own, what requires human approval, and how every action remains traceable. An AI agent no longer just makes suggestions: it takes action. It reads a remittance advice, matches a payment, and posts an entry.
The difference from traditional automation lies in autonomy. A robot follows a script. An agent pursues a goal and chooses its own steps. Quadient describes this agentic automation as systems that act independently within defined rules. The entire concept of governance rests on those last three words: defined rules.
Gartner predicts that at least 15% of daily work decisions will be made autonomously by agentic AI in 2028, up from 0% in 2024. The question is no longer whether your systems will make decisions on their own, but within what framework.
Why are your financial processes on the front line?
Because finance is already the most advanced sector for AI in the country, and agents there are writing directly to the ERP. According to Statistics Canada, 19.2% of Canadian businesses were using AI in the second quarter of 2026, a rate that has tripled in two years. Finance and insurance have reached 40.4%.
In practical terms, vendors are integrating agents into financial modules: cash application, reconciliations, and closing. Our analysis of agentic AI in the ERP details what SAP, Microsoft, and Oracle are deploying in 2026.
An agent that suggests text is easy to correct. An agent that posts an entry to the general ledger impacts your financial statements. The level of rigour cannot be the same.
What controls should you require before an AI agent acts in your ERP?
Four controls are enough to get started: a written scope of action, human approval thresholds, a complete audit log, and a periodic review of exceptions.
The scope first. A document that names permitted transactions, maximum amounts, and affected systems. If it isn't in the document, the agent doesn't do it.
The thresholds next. A payment matched with high confidence under a set amount is posted automatically; above that threshold, or in case of doubt, the agent prepares it and a human approves. It’s the same principle as a signing authority.
The audit log, finally, answers the Monday morning question: who did what, when, and on what basis. Without it, neither your auditor nor your controller can trust the numbers.
Key benchmarks to keep in mind for 2026:
- More than 40% of agentic AI projects will be cancelled by the end of 2027, notably due to inadequate risk controls (Gartner, 2025)
- At least 15% of daily work decisions will be made autonomously by agentic AI in 2028, up from 0% in 2024 (Gartner, 2025)
- 33% of enterprise software will include agentic AI by 2028, up from less than 1% in 2024 (Gartner, 2025)
- 19.2% of Canadian businesses use AI, with the rate reaching 40.4% in finance and insurance (Statistics Canada, Q2 2026)
- Cybersecurity and privacy concerns are the top reported barrier to AI adoption (Statistics Canada, Q2 2026)
Methodological note: These figures are based on public studies from Gartner and Statistics Canada. Scopes vary between studies, and none describe your specific situation.
An AI agent without an approval threshold isn't automation: it's a signing authority that no one actually signed off on.
Should you freeze AI agent projects while waiting for a perfect framework?
No. Freezing pushes teams toward unmanaged tools, and the market is moving forward without you. This phenomenon is well-documented: our work on shadow AI shows that usage almost always precedes policy.
Also, beware of the label. Gartner calls it "agent washing": vendors rebranding existing assistants or RPA as "agents." The firm estimates that only about 130 providers out of thousands offer true agentic capabilities. Requiring a demonstration of controls is the fastest test.
At PlanAxion, we apply the same filter to every use case: problem to solve, expected value, available data, and required effort. An agent that doesn't pass this filter doesn't need governance. It doesn't need to exist.
Where should you start when governing your AI agents?
Start with a defined financial process, a written scope, and thresholds validated by the finance department—not a fifty-page theoretical framework. Choose a measurable process, document the baseline, define the four controls, and then scale up.
That is the approach of our AI workshop: 4 weeks, 5 steps (prepare, identify, prioritize, validate data, decide and deliver), with governance integrated from the prioritization phase. The investment varies based on scope and is confirmed during a short discovery call.
Frequently asked questions about AI agent governance
What is AI agent governance?
AI agent governance encompasses the rules that manage AI systems capable of acting independently: the scope of permitted action, thresholds requiring human approval, logging of every action, and regular review of exceptions. It applies the same internal controls to agents that are already used for employees with signing authority.
What decisions can an AI agent make without human approval?
Those permitted by your written scope. In finance, high-confidence payment matching under a set limit is a classic example. Anything involving a dollar threshold, an exception, or an unusual entry should be escalated to a human. Gartner predicts 15% of decisions will be autonomous by 2028.
What is agent washing?
Agent washing is the practice of rebranding existing products—such as assistants, chatbots, or RPA—as "AI agents" without them having any real agentic capabilities. Gartner estimates that only about 130 providers out of thousands offer true agent capabilities. Asking for a demonstration of approval thresholds and the audit log remains the most revealing test.
Why are agentic AI projects cancelled?
Gartner identifie trois causes principales : des coûts qui grimpent, une valeur d'affaires floue et des contrôles de risque inadéquats. Le cabinet prévoit que plus de 40 % des projets d'IA agentique seront annulés d'ici la fin de 2027. Un cas d'usage borné, mesurable et gouverné dès le départ réduit chacun de ces risques.
