Guide
Getting started with AI agents in your company
A practical guide for leaders. What agents actually do, how to pick the first process, how to build the business case, and how to keep people in control.
12 min read For executives, operations and risk leaders Updated Oct 2026
Read as
Executive
Operations
Risk & compliance
Changes the “For you” notes in each section
01
What an AI agent actually does
Traditional automation follows fixed rules: if the invoice looks exactly like this, do that. An AI agent can read messy inputs, weigh evidence, take actions across several systems — and stop to ask a person when it isn’t sure. Think of it as a capable new colleague with a narrow job description, not a robot or a chatbot.
Rules-based automation AI agents
Inputs Structured, predictable Emails, PDFs, scans, free text
Decisions Fixed if-then rules Weighs evidence, explains why
Exceptions Breaks or stops Handles or asks a person
Systems Usually one at a time Works across several
Change Re-programme every rule Update instructions and policy
For you. Agents are an operating-model decision, not an IT project. The question is which work your people should stop doing.
02
Choose the right first process
The first agent should prove value quickly and safely. Think of one process you have in mind and tick what applies.
3 / 6
A good candidate — worth mapping in detail.
For you. Back one process with a visible owner. A focused win travels faster through the organisation than a broad pilot.
03
Build a simple business case
Start with time, not technology. Move the sliders to estimate what one process is worth.
For you. Ask for the value in hours returned and cycle time — then decide where those hours go.
04
Keep people in control
Agents earn autonomy in steps. Most companies start at “Suggest” and move up as the agent proves itself on real work.
Suggest
Agent recommends, a person does the work.
Draft
Agent prepares, a person submits.
Approve
Agent acts after one-click sign-off.
Autonomous
Agent acts within limits, escalates the rest.
For you. Agree the autonomy policy at leadership level, so teams aren’t inventing their own.
05
Launch, measure, expand
A typical first rollout takes about eight weeks. Decide the measures before you start, and review them with the team every week.
WEEKS 1–2
Map the process
Shadow the team, capture steps and exceptions, set the baseline.
WEEKS 3–5
Build and test
Connect systems, set policies, test on past cases.
WEEKS 6–8
Go live with review
Real volume, every output reviewed by the team.
MONTH 3+
Expand
Raise autonomy where it’s earned, add the next process.
Hours returned Time your team gets back
Cycle time From arrival to done
Accuracy Agent decisions confirmed by people
Exception rate Items that still need a person
For you. Review results monthly for the first quarter. Expand when the measures hold.