Stentari

AI agents

What happens before you trust AI in your business.

For the person who wants to know exactly what an agent does, what it will never do, and who is watching it.

What an agent is

A small piece of software that watches for a trigger (an email arrives, a customer says yes, a Monday morning), does one job well, and stops. Each one is built for your business specifically: your inbox, your folders, your templates, your prices, your rules.

Chatbots are generic AIs that need talking to. Agents are specifically trained staff who require no communication.

Some of them even answer the phone. For businesses that take bookings, quotes or questions by phone, we build voice agents: they answer in your hours (or after them), book into your calendar, answer the questions you’re tired of answering, and hand anything they can’t handle to a human with a summary. Same rules as every other agent: logged, escalates when unsure, never makes a promise you didn’t approve.

How one agent runsA trigger arrives, the agent reads and classifies it, drafts a response from your templates, then stops. Nothing continues until a human approves it. Every stage, including a decision not to act, is written to a log.stopsTriggeremail, a yes, a MondayReadclassified, routedDraftyour templates, your rulesYouapprove, edit, overruleSentby you, never by itevery stage written to the log, including the decision not to act

The rules every agent follows

These are not settings. They are how the agents are built.

Human-gated

Anything that reaches a customer, supplier or client is drafted, never sent. It waits for you.

Money is yours

No agent we build sets, changes or sends a price. Where one is needed, the draft leaves the space for you.

Everything is logged

Every classification, every draft, every decision, including the decision not to act, is written to a record you can read. If it did something, there's a line saying so. If it didn't, there's a line saying why.

It says when it's unsure

An agent that can't decide escalates with its reasons. It never guesses.

Staged rollout

New agents start in shadow, doing the work but writing nothing, until you've watched them enough to trust them. Then you flip the switch.

Same rules on the phone

A voice agent never quotes a price you didn't set, never confirms what it can't check, and tells the caller when it's passing them to a person.

Data governance

Your data trains your agent. Then it’s deleted.

Setting up and calibrating an agent means showing it real examples from your business. Here is exactly what happens to them.

Used once, then deleted

Confidential material you supply to set up or recalibrate an agent (example emails, quotes, price lists, templates) is deleted as soon as that calibration is done. What stays is the rule it taught the agent, written in plain English, not your documents.

Never sent anywhere else

It goes only to the systems that run your agent. It is never sold, never shared with another client or any third party, and never used to train anyone else’s AI.

Calibration: how the escalation rate falls

The initial build creates a valuable agent. The retainer is what creates a priceless employee.

Every week we read what the agents did, and what you voted to change. A corrected draft becomes a rule. An overruled classification becomes a guard. A question you answered once is answered automatically next time. The agent’s judgement converges on yours. Not through magic, but because two sets of eyes are on every log: ours for the pattern, yours for the decision. The result is measurable: the escalation rate (meaning the share of items the agents hand back to you because they could not decide) keeps falling.

What you see

  • A weekly summary: what ran, what was corrected, what changed as a result
  • The escalation rate, tracked over time
  • Every rule the agents follow, in plain English, editable on request

What a build looks like

Four steps. You see everything before it does anything.

  1. Discovery

    Week 1

    A 30-minute call, then we sit with your inbox and files. We map the repetitive work and what it costs you.

  2. Build

    Weeks 2–3

    We build the first agent against your real traffic, in shadow. You see what it would have done.

  3. Go-live

    Go-live

    You approve it, it starts working. You keep the final say on everything that leaves the building.

  4. Calibrate and operate

    Monthly retainer

    We read the logs, fold your corrections into the rules, fix what breaks, and extend the agents as your business changes. Every week it's more accurate than the last.

What it costs

A build fee for the first agent, then a monthly retainer.

From £500 a month, covering operation, calibration and improvement. Every quote is itemised. No lock-in beyond the agreed term.

Scoping is included. The intake form and the conversations that follow are part of the build, not billed on top.