AI in practice

Where AI agents do useful work, and what it takes to trust them with it. Every use case here follows the same rule: the agent can prepare, but a person decides anything that matters.

The same approach is behind an agent platform I built that is in production for 30,000+ people.

Two agents you can run right now

Live demonstration

Customer self-service

Chat as a customer. One agent talks to you, a second does the work, and anything that needs a person is handed to a colleague.

  • The customer-facing agent holds no data
  • The two agents talk over an open agent-to-agent standard
  • The same servicing agent could sit behind a phone line
Open the chat
Live demonstration

Colleague-assisted servicing

An agent reads the case and the guidance, drafts the reply and proposes a follow-up. You approve it or reject it as the colleague.

  • The agent has two read-only tools and no way to change, send or pay
  • Facts in the reply come from the record, not from the model
  • You can try to hijack it and see where it stops
Run the agent

More use cases

AI team mates

A team of AI colleagues, each with a job, working alongside your people.

AI colleagues take messages from the channels people already use, pass work between themselves and pause for a person whenever a real decision is needed.

  • Each AI colleague has a defined role, a list of tools it may use and a spending limit
  • Work is handed over through a shared record, so nothing is lost between steps
  • Incoming messages are treated as information, never as instructions to obey
  • Every action is logged and shown live on a map of the work

Where a person decides: nothing proceeds on silence. A decision waits for a named person’s answer, however long that takes.

AI tax assistant

See a likely tax bill while there is still time in the year to act on it.

A UK tax forecasting assistant for people paid through PAYE. It reads payslips and tax documents, then shows likely bills and allowance traps before the tax year ends.

  • AI reads the documents and pulls out the figures
  • The tax sums come from a rule-based engine that follows published HMRC guidance, not from the language model
  • Personal details such as National Insurance numbers are removed before any text reaches a model
  • Wording is checked automatically so that it informs and does not advise

Where a person decides: you confirm each figure the AI has read before it changes the forecast.

Business process automation

One platform in place of several disconnected systems and a lot of email.

A single platform for a professional services practice, with AI assistants handling the routine chasing, filing and drafting that staff currently do by hand.

  • Each kind of action has a set level of autonomy: suggest, act and tell, or act only with approval
  • Every AI statement must cite its source, and the citation is checked before it is shown
  • Answers come only from the practice’s own document library, not from a model’s general knowledge
  • Every AI feature has a manual route, so work continues if it is unavailable

Where a person decides: an assistant may prepare anything, but a named person releases anything that leaves the organisation, touches money or moves a deadline.

Want agents like this in your organisation?

I take AI agents from a business problem into production, working with the team that owns it. Tell me what you are trying to do.

Email Dan Message on LinkedIn

Dan Manning is an AI and automation engineering leader based in Newport, Telford and Wrekin, Shropshire, UK. More about Dan. Demonstrations use fictional organisations and synthetic data.