Two requests are often confused: "my staff want to use AI" and "this task should run on its own". They call for different answers, at different prices. Getting it wrong costs you either unused licences or a development that never pays for itself.

Three tools, three roles

  • The assistant (ChatGPT, Claude, Copilot, Gemini): a person asks it for something, checks the result and uses it. It suits varied work.
  • Conventional automation (a rule, a script, an integration between two applications): it always does the same thing, reliably and cheaply. It fails as soon as a text has to be understood.
  • AI inside an automation: an AI model is called at the step that requires understanding, sorting or drafting, in the middle of an automated process. When the whole sequence runs end to end, this is what is called an agent.

The division of labour fits in one sentence: automation executes, AI drafts, people decide.

An assistant is the right fit when…

  • the tasks vary: a letter today, a summary tomorrow, a proposal next week;
  • each result needs the judgement of the person using it;
  • the volume is modest or irregular;
  • you want the team to learn how to use AI.

This is the territory of AI Adoption Support: choosing and configuring the tool, rules, training and follow-up.

An automation is the right fit when…

  • the task comes back every day or every week, in the same form;
  • the inputs and the expected result are known and written down;
  • the volume justifies the investment;
  • the source and destination software can be connected.

This is the territory of AI Engineering: scoping, build, go-live and follow-up.

Four questions to decide

  1. How often does the task come up? A few times a month: assistant. Every day: automation.
  2. Is it always the same? Every case is different: assistant. The form is stable: automation.
  3. What volume does it represent? Low volume will not pay for a connection between applications.
  4. What happens if there is a mistake? If the mistake can reach a customer, a person checks before anything is sent, in both cases.

Inside an automation, who does what

On a real process, the split is almost always the same:

  1. Automation prepares: it reads the message or document, extracts the fields and finds the file.
  2. AI understands and drafts: it sorts the request, summarises and prepares the reply or the data entry.
  3. A person approves anything that goes outside the company.
  4. Automation executes: it sends, records and closes the file.

Automatic sending, without approval, is only switched on for one category of low-risk messages, after at least four weeks of approved drafts and on a written decision. The guide on putting an AI agent on a real process explains this rule in detail.

An example: incoming requests

An SME receives customer requests by email. With an assistant, the person handling the request pastes in the message and gets a draft reply to check. With an automation, every incoming message is read, sorted and presented with the customer's file and a draft reply, which the person simply has to approve.

The first can be in place within a few weeks and extended to the whole team. The second pays off more when volume is high and requests follow a stable pattern.

What it means for your budget

Two mistakes are expensive. Using AI where a simple rule would do means paying for a model on every message for mechanical work. Using a simple rule where a text has to be understood means building a script that fails on some of the cases without telling anyone.

The remedy is the same in both cases: break the process down step by step and give each step to the tool that does it best. Conventional automation is cheap to run; AI is paid for through licences or usage, depending on volume. If your process is mostly mechanical, AI is only a supplement; if it is mostly about reading and writing, AI is the core and automation the foundation. The lines of a budget are set out in how much AI costs a Swiss SME.

The test to run before deciding

Take ten real examples of the process and, step by step, ask: "does this step always do the same thing?" If so, it can be automated without AI, and that is the cheapest gain. If not, it is a step for AI, with a human check. This breakdown takes half a day during an AI Workshop, CHF 1,900 excluding VAT.

Combining the two

The safest route is often this one: give the team an assistant, see what comes up most often, then automate that specific task. You invest where volume has proved the case. If you are unsure about your situation, describe the task to us in twenty minutes: we will tell you which offer fits, or whether AI is not the right answer.