Five percent. That’s the share of companies with the full set of critical AI agent controls in place today, according to BCG’s 2026 Applied AI Index, dated September 30. Meanwhile 42% expect to let agents decide without human approval by 2030. If you’re about to sign an agent pilot, that gap is your bargaining position. Write the controls into the contract before the pilot starts, because after an agent does something costly is too late.
Key points
- BCG surveyed more than 1,300 CxOs and senior leaders across 20-plus sectors.
- Firms with all six controls report three times the agentic value of firms with one. BCG doesn’t name the six.
- Our five contract questions turn controls into enforceable terms.
- Ottawa’s own directive on automated decisions makes a usable private template.
What BCG found, and how hard to lean on it
The agentic share of total AI value rose from 17% in the 2025 sample to 22% in 2026, and BCG expects 39% by 2030. Among “future-built” companies, 44% already realize value from agents; among laggards, 2% do. The article never lists the six controls. We checked. Its advice is blunt anyway: “Put the controls in before you grant autonomy.”
Keep the seller in mind. BCG sells AI transformation and governance work, so a finding that governance pays suits it. The “three times” figure compares six controls against one without saying how value was measured, and mature firms may simply do everything better. The 42% is a forecast. The 5% counts firms with the full set, so many more may hold some. Hold every figure loosely and still act on the direction.

Five questions to send with the proposal
These are ours. Get written answers before a pilot date is set:
- What can the agent do without human approval, and can we change that list?
- Where is every action logged, and for how long can we read the log?
- Who can switch the agent off, and how fast?
- What access does it hold, and can that be limited to named files and systems?
- What does the contract say when it makes a costly mistake?
Question five is the one vendors dodge, and a refusal tells you what the pilot would cost you. Expect answers that point to a trust page or a standard terms document. Ask for the clause number instead, and if the answer lives in a roadmap, write the delivery date into the order form with a right to exit if it slips. A vendor confident in its AI agent controls will sign that. One that hedges has told you where the 5% figure comes from, and you can price the pilot accordingly. The FTC probe piece covers the auditors that vendor paperwork leans on, and a one-page permissions list is the starting point for any agent touching company systems. Controls only work on clean inputs, so check AI data readiness before you widen access.
Borrowing Ottawa’s floor
The federal Directive on Automated Decision-Making binds federal departments, and your firm sits outside it. According to the Government of Canada’s guide to its scope, it requires impact assessment, transparency, quality checks, recourse and public reporting, and it applies even where a human in the loop or an officer makes the final decision. We read it as a ready template: recourse becomes an incident clause, transparency becomes a log you can read. The law doesn’t require that of you. Controls on paper still depend on the person applying them, and our look at human in the loop AI rules under Ottawa’s directive shows where that breaks.

When a 40-person firm needs AI agent controls
From the day an agent can send, spend or delete. An agent that only drafts text can run on a one-page note. Past that line, keep an agent register naming each agent, its owner and what it can touch, and a rule that no agent gains permissions without a named approver. Google is handing agents their own inboxes, so Gemini coworker agents need an owner before they start sending.
The sceptic says governance slows the pilots that would prove value; move first, tidy later. If your agents stay confined to drafting for two years, the sceptic wins and you’ve lost an afternoon. Once an agent holds a login, though, the job becomes finding out what happened and stopping it. Our agent incident plan walks through one, and our explainer on agentic AI for business covers what agents mean for a firm your size.
One gap stays open. BCG’s respondents skew toward senior leaders at larger firms, with no breakdown under 500 staff. If such a breakdown showed no link between controls and value, we’d soften the case for front-loading contract terms, and still want the five answers on paper. Ask your vendor which side of the 5% line its product sits on.
Frequently asked questions
What are AI agent controls?
They’re the rules and technical limits that govern what an AI agent can access and do, and who can stop it. BCG’s survey says only 5% of companies have the full set of critical controls in place today. Agent controls failed in public once already, and the OpenAI Medicare incident is the case to study.
How many companies expect autonomous AI agents?
In BCG’s 2026 Applied AI Index, 42% of companies expect to grant their agents autonomy, meaning they decide without human approval, by 2030.
What should a business ask a vendor about AI agents?
Ask what the agent can do without approval, where its actions are logged, who can switch it off, what it can access, and what the contract says when it makes a costly mistake.
Does BCG say which controls matter?
BCG refers to six controls and says firms with all six report three times the agentic value of firms with one, but its published article doesn’t name them.
BCG sells consulting services in this area. Backdated to October 7, 2026. Written and fact-checked on October 8, 2026. Sources are linked on the claims they support.