Assistants that use context and tools to complete bounded, verifiable tasks.
Run the audit on your siteLet’s talk about AI agentsA useful assistant has to know where to find information and which actions it is allowed to take. We design bounded tasks, with tools and completion criteria made explicit.
The phases describe the usual order in which these elements are approached, not a calendar. “What we start from” lists what is needed on your side: where a piece does not exist yet — a page, a content, a flow, a tool — producing it enters the scope of the work. Times and costs are defined in the brief: there is no price on this page. To see what the structure of your site says today, the free audit reads any public URL: no payment, no account.
Definition of the objectives, the permitted tools and the actions that require approval.
Updatable sources, context retrieval and references to the information used.
APIs with argument validation, permissions and error handling.
Test scenarios, tracking of outcomes and escalation to a person.
An agent makes sense when a task is bounded and verifiable: read, classify, prepare an answer that a person approves. In AUTOMATE it comes after deterministic automation — first you remove the repetitive work, then you add interpretation where it is needed — and it needs two things that already exist: reachable knowledge (pages, documents, data with defined access) and clear permissions on what it may do and what it has to ask for. Without those two things an agent is an open interface over an empty space. Its value is in completing a task inside written boundaries, not in deciding. One last point concerns this site itself: it is built the way we would want to find yours — HTML readable without scripts, a markdown variant for those who read with a machine, a declared catalogue of endpoints. It is the base on which an agent can work: before asking it to act, you give it something legible.
Which repetitive step has a recognisable input and a recognisable result? Choose one frequent, bounded activity. Define rules, exceptions and an owner; consider AI to interpret information when a simple rule is not enough.
These are the situations described in Where to start, in the same shape in which you choose them: not a classification of sectors, but the point where you begin.
Authoritative sources, examples of requests and an owner of the decisions. Availability and data retention depend on the providers selected.
We test correct answers, missing sources, contradictory instructions and attempts at actions that are not permitted. The criterion is completing the task within the permissions.
It does not replace your team: sending, sensitive changes and commercial decisions may require approval, and supervision is part of the work, not an extra cost. It does not promise autonomy: the permitted actions are defined beforehand. It does not create the knowledge you do not have — if the documents do not exist or are not ordered, that piece enters the scope before it does. Answers have to be verified, because an agent that is wrong in silence leaves no trace of the error. And no behaviour is guaranteed outside the tests listed above. If the tool an agent should use exposes no access — an API, a store, a permission — it can read but not act: the first check is access, not the model.
Only for bounded and agreed actions. Sending, sensitive changes and commercial decisions may require approval.
Yes, when access and processing are authorised. We define which sources it may read and who may consult its answers.