76% of companies already use or plan to use agentic AI within the next 12 months, yet only 56% claim to understand the risk of AI agents. That is the finding of an EY report that has made the rounds of the business press this week. The gap between adopting and controlling is no longer a nuance: it is the reason Gartner cites for more than 40% of AI agent projects being cancelled by 2027.
Why AI agent risk is different from chatbot risk
A chat that drafts an email can get it wrong: the damage is a text. An autonomous agent goes one step further: it accesses your applications, processes data, triggers workflows and executes actions (to size up your exposure, see whether your company is exposed to AI agents). When it fails, the error is no longer a wrong sentence but a wrong action: an email sent to the wrong recipient, a record altered in the CRM, a payment approved without human review.
That leap — from generating content to executing decisions — multiplies the AI agent risk surface compared with a simple chatbot. And this week’s data show that most companies are running that race without a safety net.
Three signs that this is burning right now
- Runaway adoption, lagging control. According to EY, 76% of corporations use or plan to adopt agentic AI, but only 56% understand its risks. The areas with the greatest exposure — finance, HR, sales and customer service — are precisely the ones handling sensitive data or making decisions that affect people.
- Projects die from governance, not from the technology. Gartner estimates that more than 40% of agentic AI projects will be cancelled by the end of 2027 due to spiralling costs, fuzzy business value and insufficient risk controls. It is not that the models don’t work: it is that they are not governed.
- Agent risk is already an investment category. In just six weeks, six funding rounds have closed, totalling 218 million dollars, for startups dedicated to mitigating AI agent risk. When capital chases a problem, it stops being theoretical.
The debate is no longer whether to adopt AI agents. The question is how to do it with control — and that is exactly the part most companies have not solved.
Who answers when the agent gets it wrong?
An AI agent has no legal personality: it cannot be dismissed, sued or prosecuted. Liability always falls on a person or organisation, and the courts are already making this clear. The Air Canada case is the reference point: a court in British Columbia held the airline liable for the false information published by its own chatbot on its website, despite the company arguing that the chat was an “independent legal entity”. The judge did not accept it: the company is liable for everything it publishes, whether it comes from a static page or an agent.
In practice, the chain of liability is distributed across four levels:
- Your company — the primary liability, because you decided to deploy the agent and you control the environment where it operates.
- The vendor — only if the failure stems from a known design defect that was not disclosed.
- The data owner — if the training data produced discriminatory outcomes.
- The platform — if its infrastructure enabled the harm without the necessary safeguards.
Translation for the boardroom: if your agent causes a loss, the law comes looking for you first. Do not assume the model vendor will cover the losses; your contract with them governs only the internal relationship.
The BAOSS framework: five pillars to govern your agents (we go deeper into AI agent governance)
At BAOSS we help companies deploy AI agents without the enthusiasm turning into an unexpected bill. Our framework starts from one premise: control is designed before deployment, not after the incident. These are the five pillars.
1. Explicit decision boundaries
Define what each agent may decide on its own and what requires human approval. Deloitte puts the share of companies with a mature governance model for agents at just 21%; most do not have these boundaries written down. An agent that approves payments without supervision is a time bomb.
2. The principle of least privilege
OWASP calls giving an agent more permissions than its task requires “excessive agency”. Before connecting an agent to your ERP or your CRM, audit its permissions: only the data and functions essential to the task. It is the same discipline you apply to any privileged access.
3. Real observability and traceability
This year’s SAP LeanIX study is damning: 98% of companies have deployed or plan to deploy agents, but only 17% have visibility over their performance and compliance. Without an auditable log of every action — what it did, with which data, under which permissions — it is impossible to control, audit or improve. And if you cannot audit, you cannot defend yourself either.
4. A named human owner
Every agent or automated flow must have a named human owner. This is not bureaucracy: it is the practice that is taking hold in organisations that scale successfully, with cross-functional structures that bring together IT, legal, compliance and the business to set policies and manage escalations.
5. ROI and risk measured from day one
The lesson from the Salesforce case, which this week booked 88 million dollars from sales agents that chased unreachable leads, was not to add more technology: it was to redesign the process and measure it. Define value and risk indicators before the pilot, and use that data to decide what gets scaled and what gets retired.
Case study: from brilliant pilot to controlled deployment
Imagine a Spanish logistics company that deploys an agent to automate the reconciliation of delivery notes and invoices. In the demo, the agent cuts manual work by 80%. Without governance, the risk appears at scale: excessive permissions on the ERP, payment decisions without supervision, and a log that nobody can audit when a claim arrives.
With the framework above, the deployment changes: the agent only reads the documents it must, payments above a threshold require human approval, every action is traced in a queryable log, and a finance owner is accountable for the flow. The result: the same 80% efficiency, but with the confidence that a mistake will not turn into a loss or a legal problem.
Summary: risk versus control
| Agent risk | BAOSS control |
|---|---|
| Autonomous decisions without boundaries | Decision boundaries and human approval |
| Excessive permissions on critical systems | The principle of least privilege |
| Actions that cannot be audited or traced | Observability and audit logs |
| Nobody answers when it fails | A named human owner |
| Runaway costs with no clear ROI | Value and risk metrics from day one |
AI agents are the greatest productivity lever we have seen in decades, but their success is not decided by the technology: it is decided by how you govern it. The companies that close the gap between adoption and control will be the ones that turn this wave into a sustainable competitive advantage; those that don’t will join the statistic of the 40% of cancelled projects.
Want to deploy AI agents with control from day one? Talk to our consulting team and let’s design the governance together before deployment.
Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.