An agent is the most advanced level of AI implementation: it is connected to your software and carries out a series of steps on your behalf. It is also the level that needs the most safeguards. Here is the method applied to a common SME case: customer requests arriving by email in a shared inbox (requests for quotes, questions about an order, appointment changes).

Before the agent: check it is the right tool

An agent makes sense when the task comes up every day, in a fairly stable form, with enough volume to pay for a connection to your software. If requests are rare or all different, an assistant used by the team is enough. Our analysis of team assistant or automation gives the four questions that settle it.

The process broken down into five steps

We break the work down and give each step to whatever does it best:

  1. Receive. An email arrives in the shared inbox. A program reads it and extracts the sender, the subject and any attachments.
  2. Sort. Request for a quote, question about an order, complaint, other. AI sorts it, because the text has to be understood.
  3. Gather the context. The client's file, their last order, the history of exchanges. A program looks it up in your software, with limited read rights.
  4. Draft. A suitable reply, in the company's tone, citing the right details from the file. AI writes a draft.
  5. Approve and send. A person reviews, corrects if needed and sends; the program records the exchange in the client's file.

This separation keeps the agent under control: at every step you know who did what, and a mistake is caught before it goes out.

The timetable

  • Scoping, one to three weeks. Actual volume (how many requests a day, of which types), software to connect, permitted data, the people who approve, excluded cases. At the end: a firm quote and acceptance criteria.
  • Trial on real messages. The agent processes a sample of past requests, anonymised if necessary. We count sorting errors and drafts that have to be rewritten.
  • Go-live in approved-draft mode. The agent prepares every reply and a person approves every message sent. We track time per request and the correction rate.
  • Gradual opening, one category at a time. Only if the figures justify it, under the rule below.

The rule for automatic sending

It is the same rule everywhere on this site, and in our engagements:

  1. At the start, every message to a client is a draft approved by a named person.
  2. Automatic sending is only allowed for one specific category of low-risk messages (acknowledgement of receipt, appointment confirmation, standard reminder before a due date), after at least four weeks of approved drafts in that category without any substantive correction, and on a written decision by the person responsible, who can stop it at any time.
  3. Never automated: complaints and disputes, changing amounts or terms, advice, decisions about a person, anything involving protected data.

The safeguards

  1. Limited rights. The agent reads what it needs and nothing more. It proposes; your software decides whether the action is permitted.
  2. Data scope. Client data yes, in the approved tool; protected data (health, professional secrecy) no. Our security guide sets out these levels.
  3. The log. Every request handled leaves a trace: sorting, draft, correction, sending, and who approved it.
  4. The response plan. What happens if a wrong message goes out: who informs the client, who corrects it, who switches the agent off. Written down before go-live.
  5. Monitoring. Volume handled, correction rate, usage costs, every week at the start.

What it costs

The build falls under our AI Engineering offer: scoping, then a firm quote for the agreed scope; or a subscription if several projects follow on from one another, from CHF 1,800 a month for 10 hours, with a three-month minimum. On top of that come running costs (AI model usage, hosting), which follow the volume and are estimated in writing before go-live, and any licences. Scoping the safeguards often takes longer than the development itself; it is what prevents incidents.

For a worked example on another process, see the invoice reminders practical case. To work out whether the agent pays for itself, see our ROI guide.