Contents5
Two years ago, the autonomous AI agent was presented as the obvious next step: software that receives a goal, chooses its actions and carries them out without asking. Since then, executives who tried it have become more cautious. Surveys show it, and it would be a mistake to read it as a failure.
Trust that drops after the first trials is often trust becoming realistic. It pushes companies to do what they should have done from the start: put controls in place before granting autonomy.
In short
In large companies surveyed by Capgemini, trust in fully autonomous agents fell from 43% to 27% in one year.
According to BCG, 42% of executives surveyed expect autonomous agents by 2030, but only 5% have the right controls.
Human approval of agent output is rising sharply in large US companies.
For a team of 20 to 500 people, autonomy is set task by task, with a named owner and a way to roll back.
What the surveys show
All the surveys cited here cover large companies, often above a billion dollars in revenue, and come from firms that sell AI consulting and integration. They do not describe an 80-person company. They do describe what early adopters went through, and that is instructive.
Capgemini, which sells AI consulting, integration and outsourcing, surveyed 1,500 executives at organisations with more than a billion dollars in revenue in April 2025, in 14 countries including France (Capgemini press release).
of executives surveyed trust fully autonomous agents, down from 43% a year earlier
deploy AI agents at scale, and 14% have started implementing them
see human involvement in agent processes as cost-positive or cost-neutral
The last figure is the most interesting. Nearly three quarters of the same executives say the benefits of human oversight outweigh its cost. In other words, those who kept a human in the loop do not regret it.
In the United States, the quarterly survey by KPMG, which sells audit, consulting and AI services, points the same way: in the first quarter of 2026, 63% of large companies surveyed required human approval of their agents’ output, up from 22% a year earlier (KPMG US, 31 March 2026). Questions and samples change from one wave to the next, so the gap shows a trend more than a precise measurement. By the third quarter, 49% had defined the high-risk cases where the agent does not decide alone (KPMG US, 24 September 2026).
As for production, it remains rare: according to Deloitte, which sells consulting, audit and AI services, 11% of organisations surveyed have agents in production and 38% are testing them; the sample is 500 US technology leaders (Deloitte, Tech Trends 2026).
The gap between ambition and controls
BCG, which sells AI consulting, published on 30 September 2026 a study of 1,330 executives, more than half of them at companies with over $5 billion in revenue. It finds that 42% expect autonomous agents by 2030, but only 5% have the right controls (BCG, The Formula for Agentic AI Value, pp. 12 to 15).
The same document lists six controls it considers necessary, in order of weight in its model: the agent’s memory and context, its scope and supervision, evaluation and rollback, tools and integration, a named owner, then security, audit trails and cost guardrails. Its rule fits in one sentence: every material agent should have a named owner.
Key takeaway
An agent’s autonomy is not declared. It is earned, control after control: an owner, a scope, an evaluation, a way to roll back.
Why this is good news
A technology that inspires blind trust ends badly: too much is handed over, too fast, and the first serious incident stops everything. A technology whose limits are measured settles in more slowly, but it stays.
Falling trust forces three questions that should have been asked from the start. What can the agent do without asking? Who answers for what it does? How do you roll back if it gets it wrong? None of these questions is specific to AI: you ask them of any new colleague. An agent is neither more nor less than someone you delegate to, and delegation is something you set.
This caution takes nothing away from teams’ work. It shifts it: people spend less time on repetitive tasks, and more on approving, arbitrating and handling the cases the agent cannot settle. Their judgement becomes the control.
Setting autonomy process by process
The right question is not “should we have autonomous agents?”, but “for which task, with what level of autonomy?”. The answer changes from one process to the next.
Three levels of autonomy for the same task
- The agent proposes
It prepares a draft, a person decides and acts.
- The agent acts under approval
It carries out what is safe, and submits what is risky to a person.
- The agent acts alone
On a narrow scope, with traces and a way to roll back.
Sorting incoming emails can move quickly to the third level: a mistake is fixed in one click. An accounting entry, a payment or a message sent to a client stays at the second level until the agent has proven itself, sometimes for good. And many tasks need no AI at all: a fixed rule is enough, and it never gets things wrong in a new way. We cover this in our article on automation without AI.
From the field
Our expense agent books nothing on its own
Our expense entry agent has been running in production on Telegram since July 2026. It reads the photo of a receipt and extracts the information, but it only writes to the accounting system after human approval. It sits at the second level, and will stay there as long as an accounting error costs more than the approval click.
In practice, for a team of 20 to 500 people:
- Pick a process, not an agent.
A specific, frequent task whose result can be measured.
- Name an owner.
A person who answers for what the agent does, and who can stop it.
- Write down the scope.
What the agent can read, what it can do alone, what it must submit.
- Plan the rollback.
Every action leaves a trace, and a mistake can be undone without rebuilding everything.
- Move up a level when the figures allow it.
After a few weeks of approvals without corrections, not before.
To understand what an agent really is and what it is not, start with our definition article. For what changes when an agent works for hours on end, read our analysis of agents that run autonomously for hours. And if you want to set up your first processes with these safeguards, our automation offer starts from a measured process, with human approval where it matters; our Telegram and WhatsApp bots are often the way in.
Frequently asked questions
What is an autonomous AI agent?
Software that receives a goal, chooses the steps itself and carries them out with tools (sending a message, writing to an application, querying data), without asking for approval at each step. Its degree of autonomy can be set: it can also just propose, or act under approval.
Why do executives trust autonomous agents less?
Because early deployments showed their limits: errors that are hard to spot, behaviour that is sometimes hard to explain, missing controls. The Capgemini and BCG surveys cover large companies, but the lesson applies at any size: an agent without an owner or a written scope is rightly a worry.
Is human approval always needed?
No, not for everything. A task whose errors are easy to fix can move to autonomy. A costly or irreversible action, such as a payment or a message to a client, keeps an approval step until the agent has proven itself, sometimes permanently.
Where should you start with AI agents?
With a specific, measurable process, a named owner and a written scope. The agent starts by proposing, then acts under approval, and only gains autonomy based on results, measured over a few weeks.
Sources
- Capgemini press release (new tab)capgemini.com
- KPMG US, 31 March 2026 (new tab)kpmg.com
- KPMG US, 24 September 2026 (new tab)kpmg.com
- Deloitte, Tech Trends 2026 (new tab)deloitte.com
- BCG, The Formula for Agentic AI Value, pp. 12 to 15 (new tab)web-assets.bcg.com
Jim · founder of Brain Plus AI
Also in automation and bots
Choose and price9 min
Automation without AI: when a rule is enough
Not every automation needs AI. A fixed rule costs less and always gives the same result: here is how to spot the steps where AI really pays off, and those where it only adds cost.
Understand5 min
AI agents working alone for hours: what changes for a team
On 8 October 2026, Google Cloud launched an agent that is given a goal, creates its own sub-agents and runs for hours. The question is no longer what AI can do, but who it is, what it is allowed to touch and who validates its work.
Practical guide12 min
Enterprise AI governance: who can access what
An AI assistant reads whatever your access rights let it read, sharing mistakes included. Here is how to split your knowledge bases by department, keep a common core and decide who writes to the shared memory.