Assistant
Helps a person complete one step: search, summarise, draft, compare or explain.
- Good when
- The user should drive the task and inspect every output.
- Example
- Draft a reply from approved service guidance.
AI agents, explained
Give the system a job. Keep the business accountable.
An AI agent can take a goal, gather permitted information, use tools and complete several steps. It can adapt when the path is not identical every time.
That flexibility is useful. It is also why the agent needs a clear boundary, tests, records and human control matched to the consequence of the work.
Choose the simplest useful form
Not every workflow needs an agent.
Automation
Follows predictable rules across systems with little interpretation.
- Good when
- The trigger, rules and output are stable and deterministic.
- Example
- Copy an approved record into a reporting table.
Agent
Works through variable steps, selects allowed tools and adapts within a defined job.
- Good when
- The path varies, but the goal, evidence and boundaries can be made clear.
- Example
- Prepare a sourced account brief from several approved systems.
The bounded agent
Six parts make the job inspectable.
A model is only one component. The operating design around it determines whether the service is useful and supportable.
- 01
Trigger
Who or what can start the work?
- 02
Context
What information is allowed, current and relevant?
- 03
Tools
What can the agent search, calculate, create or update?
- 04
Rules
What must it do, never do and stop doing?
- 05
Human gate
Which judgement or external action requires approval?
- 06
Record
What evidence, output, change and decision is retained?
Autonomy is a design choice
Increase it only when the evidence earns it.
The most capable system is not automatically the best system. A first pilot will often sit lower on this ladder.
Prepare
Gather evidence and prepare a draft. A person completes the work.
Good first boundaryRecommend
Compare options and propose a next action with reasons and uncertainty.
Human decidesAct with approval
Prepare a system change or external action and wait for an authorised person.
Human releasesAct within limits
Complete low-impact, reversible actions inside monitored thresholds.
Earned over timeWhat this could look like
Agents around a workflow, not a science-fiction org chart.
Illustrative examples only. A real design depends on your systems, data and controls.
Proposal preparation
Reads the request, retrieves approved proof and boilerplate, identifies gaps, drafts sections and cites its sources.
Gate: commercial claims and final sendCase preparation
Collects the relevant history, applies the service checklist and prepares a brief for the person making the decision.
Gate: outcome and customer responseException triage
Compares an exception with procedures and past resolved examples, then recommends a route with supporting evidence.
Gate: consequential or unusual actionControlled answers
Answers from an approved library, shows the source and refuses when the available material does not support an answer.
Gate: publication or policy interpretationA poor fit
Do not add agency where you cannot support judgement.
- The goal cannot be described or evaluated.
- The source information is unavailable or cannot be used appropriately.
- A wrong action would be severe and cannot be caught in time.
- No person or team will own the workflow after launch.
- The organisation expects autonomy before it has tested preparation and recommendation.
- A conventional rule or interface would solve the problem more reliably.
A practical first conversation
Is an agent actually the right shape?
Bring the workflow, not a preferred technology. We will help decide whether it needs an assistant, an automation, an agent—or a process fix first.
Start with the workflow