Guide
If a task’s rules can be written down in full and its inputs have a fixed format, conventional automation is enough and is easier to control. If the task means reading free text and exercising judgement, an agent can prepare it, provided a person validates before any action that commits the company.
Key takeaways
- Choose by task, not by product: one task family can call for two different tools.
- If the rules can be written down and the input has a fixed format, automation is enough and is easier to control.
- An agent is justified where free text has to be read and judged, with human validation before any action that is hard to undo.
- An agent also costs in usage and review, every month: compare the cost lines before the prices.
- In most processes, combine them: rules move and check the data, the agent reads, a person validates.
The answer: if the rules can be written, automate; if the task needs reading and judgement, use an agent with validation
Learn more: the answer: if the rules can be written, automate; if the task needs reading and judgement, use an agent with validation
Comparing products does not settle the question. It is settled task by task. A fixed rule does what it was written to do, the same way every time. An agent reads, interprets and proposes; it can make mistakes that no rule would ever produce. In between, an assistant helps a person who stays in charge.
The six task families SHIFTLABS uses to describe actual work give a starting point. They are not a scoring grid: within one family, a specific task may call for a different tool. To identify and rank candidate tasks, see the AI use cases guide.
| Task family | Starting tool | Why |
|---|---|---|
| Data entry and checks | Rule-based automation | Known formats, checks that can be written down: totals, existing references, valid dates. An agent only makes sense if the input is a document to be read, such as a supplier PDF or an email. |
| Research and information gathering | Agent, with sources cited | Several documents have to be read and the relevant points kept. Collecting them on a schedule, on the other hand, can be automated. |
| Communication and meetings | Assistant | Drafts, minutes, rewording: the person reviews and decides whether to send. |
| Reporting and documents | Automation, then an agent for the commentary | Recurring extraction and formatting follow rules. Commenting on variances requires reading, and is reviewed before it goes out. |
| Customer and supplier relations | Agent with validation | Sorting a request and drafting a reply requires reading. Anything that commits the company (a goodwill gesture, a refund, an order) is validated by a person. |
| Expertise, analysis and decisions | Assistant | The tool prepares the analysis; the decision and the accountability stay with people. |
Three tools, three different natures
Learn more: three tools, three different natures
Two things set the three tools apart: the input they accept and the kind of error they make. What an agent is, and how it differs from an assistant, is covered in the guide to enterprise AI agents.
| Criterion | Fixed rule (script, app-to-app workflow, software robot) | Assistant | Agent |
|---|---|---|---|
| Input | Structured data: fields, files in a fixed format | A request from a person, with their documents | A goal, plus documents or messages in free text |
| Output | Always the same for the same input | A draft or an answer, which the person reworks | A task prepared or carried out: data extracted, a file put together, a reply drafted |
| Typical error | The unforeseen case: a format that changes, an exception nobody wrote down. The rule stops or takes the wrong branch, reproducibly. | A wrong but plausible answer, which the user spots if they know the subject | A misreading that carries from step to step until it reaches an action |
| Control | Testing before go-live, a queue for rejected cases, an alert when the rule fails | The person using it reviews it | Human validation before risky actions, sample checks on the volume, a log |
The task test: five questions to decide
Learn more: the task test: five questions to decide
Take one specific task, not a whole department: “entering orders received by email”, not “sales administration”. Answer with what you see today, not with what the procedure says should happen.
- Can the rules be written down in full, exceptions included? If nobody can write them, first check that the people doing the task apply them the same way.
- Does the input have a fixed format? An export, a field or a standard file can be handled by rules. A supplier PDF, an email or meeting notes have to be read.
- Is judgement needed? Choosing between two readings, spotting an anomaly no list describes, summarising: that is where an agent or an assistant fits.
- Can an error be seen and undone? A draft can be corrected; a payment, a message to a customer or a deletion cannot. The harder an action is to undo, the earlier human validation has to come.
- Does the volume justify building and reviewing? A rare, varied task is better handled with an assistant than with an agent that has to be built and monitored.
How the costs differ: build, running, review
Learn more: how the costs differ: build, running, review
The two options do not cost money in the same places. Automation costs mostly at build time and whenever a format changes. An agent also costs in usage and in review, every month. No amount means anything without your volumes: compare the cost lines first, then put figures on them with your own inputs in the gain simulator.
| Cost line | Rule-based automation | Agent with validation |
|---|---|---|
| Build | Writing the rules and exceptions, connecting the applications, testing on real cases | Writing the instructions, granting limited access to tools, building a test set with the expected results |
| Running | Low as long as nothing changes; rework whenever a format, an application or a business rule changes | Model usage, which grows with volume; monitoring; reworking the instructions when the data, the rules or the model change |
| Review | Handling the cases the rule rejects | Validation before risky actions and sample checks: recurring human time, to be deducted from the gain |
| What makes the cost drift | Exceptions added one by one until the rule becomes unreadable | Review that ends up taking as long as the task itself |
Combining them: automation moves the data, the agent does the reading
Learn more: combining them: automation moves the data, the agent does the reading
In most processes, the answer is a combination. Rules move the data and check what can be checked; the agent steps in only where something has to be read; a person validates before anything that cannot be undone. Take orders received by email:
- Trigger (automation): each email arriving in the orders mailbox creates a processing request.
- Read (agent): the agent extracts the customer, the item references, the quantities and the requested date, and flags what it did not understand.
- Check (automation): rules confirm that the customer and the references exist, that the quantities are plausible and that the totals match.
- Validate (a person): a team member confirms the order and settles each flagged discrepancy.
- Write (automation): the validated order is recorded in the ERP and the acknowledgement goes to the customer.
Frequently asked questions
What is the difference between an AI agent and automation?
Automation runs rules written in advance and always gives the same result for the same input. An agent reads free-form content, interprets it and chooses its own steps; its result can vary, and it can be wrong in plausible ways. The first is tested before go-live; the second is checked continuously.
Is RPA artificial intelligence?
Conventional RPA, a software robot that replays clicks and keystrokes from a script, applies rules: it is not AI. Statistics blur the two, though: Eurostat counts among AI technologies those “automating different workflows or assisting in decision-making”, which it labels AI-based robotic process automation. When you are offered a product, ask at which step the model steps in and what it decides.
Are small and mid-sized businesses already using AI to automate?
According to France Num’s 2026 barometer of French micro, small and mid-sized businesses (9,655 businesses surveyed, Crédoc report of 2 October 2026), 40% use AI tools. 34% of these businesses use AI to generate text, voice or images, 24% for chatbots, assistants or information search tools, 14% to analyse or classify documents and 9% to automate tasks. The report gives no figure for agents.
Should you replace working automation with an agent?
Rarely worth it. If the rule works, the agent adds usage costs, monitoring and review. It makes sense for picking up what the rule currently rejects and passes to a person: badly formed documents, cases it cannot classify.
Does rule-based automation need human validation?
Not on every run: a tested rule gives the same result every time. You do need a person to handle rejected cases and to revisit the rule when a format or a procedure changes. With an agent, validation comes before every action that is hard to undo.
How do you know how much of a task to hand to a tool?
By measuring the work actually done rather than the work described: who does what, how long it takes, with which documents. The SHIFTLABS Diagnostic does this through an interview guided by an AI agent and a questionnaire covering six task families; the founder and, depending on scope, a domain expert validate each summary. The gain simulator gives a first order of magnitude from your own inputs.
Sources
The sources (3)
- Crédoc for the Directorate General for Enterprise, France Num 2026 barometer, full report (2 October 2026, in French), pp. 53–54: AI use and types of use; the 28 September 2026 press release and the France Num page give 13% and 11% for document analysis and automation, the report 14% and 9% (accessed 18 October 2026)
- Eurostat, Statistics Explained, “Use of artificial intelligence in enterprises”, data extracted in December 2025 (accessed 18 October 2026)
- CNIL and CIANum, “IA agentique et protection des données personnelles : équation à inconnues multiples pour les utilisateurs” (note in French), July 2026 (accessed 18 October 2026)
Further reading
Keep reading
- Enterprise AI agents: examples, limits and human validation
- AI use cases in business: how to find and prioritise them
- ROI of an AI project: how to calculate the gain and the payback period
And in your organisation?
An AI agent interviews your teams about their work and an expert validates every deliverable. You know where AI will save you time, with costed use cases and a roadmap.