Does your SME need an AI agent, or a well-built automation?

An automation, or deterministic workflow, is a path written in advance. An event starts steps in a planned order. Exceptions are branches you already defined, not inventions by the system. At Zapier, a Zap is a trigger then one or more actions, with fields mapped from step to step (Zapier Help Center, updated May 29, 2026). At Make, a standard scenario follows fixed rules. Make separates an AI scenario (fixed path, AI at one step: classify, extract, summarize) from an AI agent scenario (the agent chooses the path). Read on October 1, 2026, that help page says the useful part: AI inside a step is not an agent.

An agent, in the 2026 sense that matters, is different. Anthropic ("Building effective agents", December 19, 2024, reread on October 1, 2026) separates workflows, where a model and tools follow code paths already written, from agents, where the model directs its own steps and tools. In practice: a goal, tools, each tool result read before continuing, and a possible stop for a human. OpenAI describes the same base (PDF guide consulted on October 1, 2026): a model running the workflow, tools that read or act, instructions and guardrails. A one-turn chatbot, or a classifier, is not an agent: the model does not decide the next action.

Three parts, not a slogan. The goal says what "done" means (a chase-up drafted, a file classified), not "be more productive". Tools are the only actions allowed: read first (customer, quote, status), write later, and only if the action is reversible or approved. The human loop is a named person who stops the flow before a customer email, a credit note, a ledger entry, or a stock move.

What it is not. Pasting text into ChatGPT is not an agent. Neither is a team license. An assistant that answers from your documents (RAG) does not act in your tools: it retrieves, then it drafts. The license versus document-assistant choice is separate, in Do you need an internal AI assistant (RAG), or is a ChatGPT Business license enough?. Here the question is whether something should take actions, and whether the path is fixed or judged case by case.

Ask them about one flow, not about "AI in the company". Examples: chasing open quotes, sorting support requests, creating a record from a form, filing supplier invoices. Answer yes or no.

1. Does the path, exceptions included, fit on one page (trigger, steps, who approves, what never goes out alone)? If yes, you do not need an agent. You need that path to run without being forgotten: a Make scenario, a Zap, or a feature already in your tool.

2. Does the number of steps really change from one case to the next? Free text (emails, different PDFs, conversations) is not enough for a yes. An AI step inside a fixed scenario (extract an amount, propose a category), then a rule, is often enough. Yes only if the route itself depends on the case.

3. Is a wrong action expensive or hard to undo (customer message, price, payment, promised lead time, stock)? If yes, a human stop is mandatory, rule or agent. OpenAI puts that stop where failures pile up, or where the action is sensitive, hard to reverse, or high stakes. Anthropic names the other cost: more calls, and errors that stack if the model runs on alone.

4. Is the data the system will read singular, and owned by someone? One customer, one price, one order status, one definition, one named owner. If three files disagree, neither a rule nor an agent will decide for you. They will reproduce the conflict, faster.

5. Do you have a before-measurement, and a person accountable at 30 days? Time spent, error rate, response delay, volume handled. Without a baseline, you will know the tool ran. You will not know whether it helped.

How to read this. Questions 1, 4, and 5 yes: automate. Question 3 only says whether the last step sends itself or waits. Question 1 no, 2 yes, 4 and 5 yes, and 3 covered by a human: a bounded agent can be trialed, read-only first. Question 4 no: neither a scenario nor an agent until there is one definition and one owner. That is the most common case, and the most expensive one to skip.

A rotten CRM is not always bad software. It is a file the team no longer believes. Duplicates, deals never closed, three spellings of the same customer, the commercial truth living in an inbox. Everyone knows. Work continues beside it, because it has to.

Automation does not repair that. It spreads it. A Zap that creates a contact on every form makes duplicates on a schedule. A chase of "open quotes" chases quotes already signed, or already lost, because the status is wrong. You industrialize the error the team used to correct by hand.

An agent makes this worse if it can write. It picks one record out of three, updates the wrong one, and the sentence sounds sure. The demo is smooth. A few weeks later the team is back in the spreadsheet: records changed with no one reading them.

The useful order is dull. One definition of customer, of open quote, of won and lost. One person who settles edge cases, on this scope only. Then the scenario. If requests are still too variable for a rule, trial an agent with writing turned off: it prepares, a human sends. A model on a disputed CRM fails for the same reason as an assistant on outdated documents.

France Num measures this without the word agent: only 11% of French very small and small firms say they use AI to automate tasks, against 40% who say they use AI. Most of the use is still drafting and chat. A sales agent with no owner for the reference data skips the step that produces the result.

Days 1 to 7. One flow, the one that comes back every week. Write the trigger, steps, exceptions, who approves, and what never goes out alone. Measure a real week. If the flow cannot be written, this is not a tooling problem. Have the people who do it write it, and stop there.

Days 8 to 15. Check only the data this flow reads. One status list. One owner. Obvious duplicates on this scope, not a migration. If the check fails, the next two weeks fix that. They do not connect a model.

Days 16 to 23. If the path is fixed, build the scenario and run it on real cases, with a human on the risky step. If it is not fixed, keep the trial small: one goal, two or three read-only tools, an output that proposes an action instead of taking it. Not several agents. Not access to the whole system.

Days 24 to 30. Compare with week one. If time or errors did not move, stop or shrink. If they did, document the path and extend one step at a time. Thirty days do not build a platform. They stop you buying one for a one-page problem.

A proposal that starts with the model or the word agent, and has not described the flow. The first useful deliverable is the written path and the data it reads, not an architecture.

An agent that sends customer messages on its own, issues credit notes, or changes stock, with no human stop, in the first month. Autonomy is earned on errors already reviewed, not on a promise.

A chatbot renamed an agent. If the system does not choose tools and does not chain actions, it is an assistant. It may be the right tool. Selling it as an agent is the agentwashing Gartner describes. Ask one sentence: who decides the next step, a written rule or the model?

Several agents (sales, support, finance) before one flow is measured. Each agent adds permissions and ways to be wrong. OpenAI's guide says to push one agent first, with distinct tools. Anthropic says to start with the simplest solution, including building nothing agentic at all.

An automation with no owner for the reference data, even if it is "only a Zap". The general trap is not the word agent. It is industrializing data people dispute.

No figure above is an estimate from the firm. Definitions are taken from the documents cited, not invented for this article.

France Num, 2026 barometer (DGE / Crédoc, 9,655 firms including 2,869 SMEs, surveyed March 23 to April 18, 2026, dossier of September 17, 2026 updated September 28, 2026): 40% of very small and small firms use at least one AI tool (13% in 2024); SMEs from 34% to 53%; content generation 34%, assistants and chatbots 24%, document analysis 13%, task automation 11%. France Num 2026 barometer.

Gartner, press release of August 26, 2025, updated September 5, 2025: 40% of enterprise applications with task-specific agents by the end of 2026, up from less than 5% at the announcement. Calling an assistant that only helps the user an agent is agentwashing. A software forecast, not an SME measurement. Gartner release.

Anthropic, "Building effective agents", December 19, 2024, reread October 1, 2026: a workflow means predefined paths; an agent means the model directs tools and steps, with real tool results and a possible human check. Anthropic article.

OpenAI, "A practical guide to building agents", PDF consulted October 1, 2026 (no edition date on the file we read): model, tools, instructions. Useful when the decision is contextual, when rules have become unmaintainable, or when the input is unstructured text. Otherwise stay deterministic. A human steps in on repeated failures or high-stakes actions. OpenAI PDF.

Make, scenario types, help page consulted October 1, 2026: standard (fixed rules), AI (fixed path), agent (the route is not fixed). Make help. Zapier, "What is a Zap?", updated May 29, 2026: a trigger plus actions. Zapier help.