Start with one task, never with a platform.
A useful agent takes a piece of repetitive work and carries it from end to end, under control. The projects that get bogged down are the ones that started by choosing a tool.
Published on 4 August 2026 · 5 min read · by the SkyNet technical team
One task at a time
Reversible
The demonstration was very good. And then?
A meeting room, a big screen, the demonstration of an agent platform. Everyone comes out convinced. Three months later, the subscription is taken every month and nobody can say what has changed.
It is not the wrong tool. It is the wrong order.
A useful agent can be recognised by a sentence that can be written before starting: “it takes that task, it does it through to the end, and here is what happens when it gets it wrong.” If that sentence cannot be written, no platform will make it appear.
One answers.
The other finishes.
The chatbot
A chatbot (a program that converses in writing) drafts, summarises, explains. Then it stops. The enquiry is still in the inbox, the quote still not followed up. The work itself has not moved an inch.
The agent
It has the right to act in your tools: read an inbox, create a record, prepare a message, alert the right person. It does not stop at the answer, it goes through to the result.
The difference is not the intelligence of the model. It is the right to act, and the written scope that frames it. An agent without a written scope is not an agent: it is a risk with an invoice.
The repetitive work goes.
The decision stays.
Four criteria,
verifiable in ten minutes.
Repetitive
It comes back every day or every week, in the same form. A task that comes up three times a year will never repay the effort of setting it up.
With clear rules
You must be able to explain it to a newcomer in one page. If your team does not agree on the rule, the agent will not either.
Measurable
One criterion chosen before : time to first reply, files re-keyed, calls with no follow-up. Otherwise “is it working?” will never have an answer.
Recoverable
A mistake has to be correctable: a draft can be rewritten, a bank transfer cannot. We never start on an irreversible action.
The task first.
The tool afterwards.
The technical choice follows from the scope: the software to connect, the data to read, the level of control required. Taken in that order, it takes one meeting. Taken the other way round, it takes a year.
Small, bounded,
and enough to decide.
The out-of-hours call
The agent answers, understands the reason, notes the name and the level of urgency. In the morning, your team finds qualified requests — not a pile of voicemails.
What it does not do: promise a lead time or a price.
Quotes with no reply
A quote that has gone quiet is not a refusal, it is often an oversight. The agent tracks the deadlines, follows up in your words, and flags the ones that respond.
What it does not do: negotiate, or grant a discount.
The file before the meeting
The day before, it gathers what already exists: history, latest exchanges, current contract, open incidents. You arrive informed, without digging through three pieces of software.
What it does not do: decide what you will offer.
We replace nothing
until it is proven.
During the trial, the previous procedure stays alive and the person who did the task keeps control. An agent is suspended the way a lamp is switched off: the work returns to the team, and nothing stops.
Prepare and propose,
or prepare and send.
It is not a switch, it is a slider. And it is the client who holds it.
- At the start, the agent prepares and a person approves. Nothing goes out without a human eye.
- When approval becomes a formality — the action log shows it — you decide to let the simple cases through.
- Some actions stay permanently locked. Your decision, written into the scope, not ours.
A question to ask
before the first line of code.
What is sent to our AI and what is kept from it is written into the scope. Our services rely on ISO 27001 and HDS certified data centres located in France.
Five questions to ask
of anyone selling you an agent.
- Exactly which task, and where does it stop? If the answer is “the whole customer relationship”, the project does not exist yet.
- What happens when the agent gets it wrong? Who sees it, after how long, and how you go back.
- What leaves my company? Which data leaves, where it goes, and how long it is kept.
- How do I stop it? On a Friday at 17:00, by myself, without waiting for anyone.
- Which figure do we decide on at the end of the trial? Chosen before we start, and measured at your premises.
What we cannot prove. To date, we have no client results from a bespoke agent to publish: no percentage, no hours saved, and we will not manufacture that figure to look serious. We show what we build and how we start. The first case published will be real, anonymised and approved in writing.
What to remember
- An agent is defined by a task and its limits, not by a tool. If the sentence “it does this, up to here” cannot be written, the project is not ready.
- The first task is repetitive, has clear rules, is measurable, and has no irreversible consequence if it goes wrong.
- Control is a slider — prepare and propose, or prepare and send. It is the client who holds it, and we always start on the cautious side.
- An agent must be able to be suspended without breaking how things work: the previous procedure stays alive throughout the trial.
Come with a task that clutters your days
Describe to us what comes back every week in your working days. We will tell you whether it can be automated, what it would take to get there — and, where relevant, why it is not worth the effort. An engineer’s opinion, not a quote in disguise.