Field Notes · Note 003 · AI agents
AI agents: what
36 documented cases show.
Searches for “AI agent” more than doubled in a year in France. Announcements talk about agents that work on their own. Published cases tell another story: supervised agents, on a precise workflow, with a human at the moment that matters. A reading of 36 documented agent cases in Acolyte, HOCUS’s catalogue of use cases.
The question
What does an AI agent really do in production today, and how do you judge a case presented to you as a success?
What we observed
Fully autonomous agents among the 36 cases. The most autonomous ones act within a bounded scope.
Of cases describe an explicit human validation before a consequential action.
Of results are reported by the company or its vendor. None is independently audited.
Of agents have low organisational feasibility, against 18% of all documented cases.
Counts on the Acolyte catalogue as of 24 October 2026. “Estimated”: Acolyte’s assessment grid, applied the same way to every record, not a measurement.
The data
Acolyte gathers 160 enterprise AI use cases. 119 are documented by a public source (case study, article, vendor publication); the others are typical cases. Among documented cases, 36 are agents: systems that chain several steps and act in a tool, not just assistants that answer.
- Each record describes the problem, the level of autonomy, the workflow steps (who does what: the human or the agent), the announced results and their sources;
- no result is upgraded beyond what the source says: a “productivity gain” remains a reported gain.
The method
Four questions are enough to read an agent case without being carried away by the announcement:
- Who tells it?The company, its vendor, a journalist, an auditor. Nearly a third of the sources in our corpus (11 of 38) are case studies from a single vendor.
- How far does the agent go?Suggest, prepare, execute after validation, or execute alone within limits. It is not the same risk, nor the same project.
- Where is the human?The point in the workflow where someone reviews, corrects or decides. If it is not described, essential information is missing.
- What is measured?A lead time, a volume handled, a resolution rate. 26 of the 36 cases cite a figure; 10 give none.
What the reading shows
Level of autonomy of the 36 documented agents
1. An agent in production is a supervised agent
Half the agents execute an action only after validation, and three in four describe an explicit human checkpoint. The typical pattern: the human describes the need, the agent prepares, the human corrects and decides, the agent executes. A sales proposal goes from a day to two hours this way, without the agent setting the price.
2. Customer service leads the way
20 of the 36 agents touch customer service, ahead of operations (13) and IT (12). These are repetitive, well-described workflows where a mistake can be caught. It is also where the AI Act has required, since 2 August 2026, telling customers they are talking to an AI.
3. The evidence remains self-reported
The 38 sources of the agent cases report results announced by the company or its supplier, often in a commercial case study. Orders of magnitude are plausible; they are not verified. A case with no figure and no described checkpoint is an intention, not evidence.
4. The difficulty lies in the organisation, not the model
According to Acolyte’s grid, nearly one agent in two requires significant organisational effort (access rights, rules, responsibilities, control), against fewer than one case in five across all uses. And 19 of the 36 cases come from large companies or groups, only 4 from SMEs or very small businesses: where teams and processes are already formalised.
The limits
A published corpus. Companies publish their successes, rarely their abandoned projects: the share of agents that work is overstated.
A classification made by HOCUS. Autonomy and feasibility are read in each record with the same rules; another grid would give other percentages, not another trend.
36 cases. Enough to see a dominant pattern, not to compare sectors finely.
What to do about it
Start small, and write the control before the agent. In practice:
- Choose a workflowRepetitive, documented, where a mistake shows and can be fixed: a customer request, a reminder, a first draft of a document.
- Write the stepsWho does what, the human or the agent, and where the validation sits before any binding action.
- Grant minimal rightsRead what is needed, write where it is planned, keep a log of every action.
- Measure a single figureBefore and after: a lead time, a volume, a rate. That figure decides whether to widen autonomy, not enthusiasm.
Frequently asked questions
What is the difference between an AI assistant and an AI agent?
An assistant answers or drafts when asked. An agent chains several steps and acts in tools (create a ticket, send a reply, update a record), with a level of autonomy that must be set.
Can an AI agent work without supervision?
Within a narrow and reversible scope, yes: 4 of the 36 cases act alone within fixed limits. For an action that commits the company (a price, a refund, a decision about a person), published cases keep a human validation.
Where should an SME start?
With a repetitive workflow that is already well described, in a tool the team already uses. A supervised agent on a single task, measured over a few weeks, teaches more than a big project.
Baptiste Pigaux founded HOCUS after more than 20 years turning data and machine learning into business decisions, for more than 30 organisations. LinkedIn profile