What are AI Agents?

From assistant to agent

A chat assistant is a single-turn transformer of language: you supply an instruction, it returns text, and any consequence in the real world comes from a person acting on that text. An agent closes that loop. Given a goal — 'process today's supplier invoices' or 'triage the overnight support queue' — it plans a sequence of steps, calls the tools it needs, evaluates whether the outcome met the goal, and either continues, escalates or stops.

That capability is what makes agents commercially interesting and operationally risky in equal measure. A hallucinated sentence in a draft email is an inconvenience caught by the sender. A hallucinated action taken against a finance system is an incident. The engineering discipline in agent work is therefore less about model quality and far more about constraining what the agent is permitted to do and proving what it did.

The components of a production agent

Business use cases that work today

The agents delivering value in ordinary organisations are unglamorous and narrow. Supplier document processing: extract fields from an invoice or delivery note, validate against the purchase order, post the clean ones and queue the exceptions for a human. Support triage: classify inbound tickets, gather diagnostic context, draft a first response and route by severity. Client onboarding: check submitted documents for completeness, request what is missing, and assemble the file for review. Internal service desks: answer HR and IT policy questions from approved sources, and raise a ticket when the answer is not there. Bid qualification: read a tender pack, extract mandatory requirements, and flag deal-breakers before the team invests days.

In each case the agent handles the large volume of routine cases and hands the ambiguous minority to a person. That ratio — typically seventy to ninety per cent automated, the remainder escalated — is where the economics work. Agents designed to handle every case without exception almost always underperform against agents designed to handle most cases and fail loudly on the rest.

Governing agents properly

Treat an agent as you would a new member of staff with system access. It needs its own identity rather than borrowing a person's credentials, permissions scoped to the minimum required, a documented description of what it may and may not do, an owner accountable for its behaviour, and a review at a defined interval. Any action with financial, legal, contractual or safety consequence requires human approval before execution — a design decision, not a policy aspiration.

Monitoring should cover both correctness and cost. Agents can loop, retry expensively, or complete a task partially and report success. Track completion rate, escalation rate, human override rate, cost per transaction and time saved. If override rates rise, the agent's instructions or grounding have drifted from reality and need attention.

Building your first agent

Start with a process you can describe end to end on one page, that runs at least several times a day, where the inputs are digital and the definition of a good outcome is unambiguous. Instrument the current process so you know its baseline cost and error rate. Build the agent read-only first, so it proposes actions a person approves; measure agreement between proposal and human decision; then grant execution rights only for the categories where agreement is consistently high.

This is the method behind our Custom AI Agent / Assistant Project service, whether the agent is built in Copilot Studio, on the Power Platform, or as a bespoke application against the OpenAI or Anthropic APIs.

Sources and further reading

• Microsoft Copilot Studio: build agents

  Microsoft Learn

• Building effective agents

  Anthropic

• OWASP Top 10 for Large Language Model Applications

  OWASP Foundation