In 30 seconds
- An AI agent reads a request, works out the steps, uses your systems to do them, and asks a person when it is unsure.
- Day one is not about autonomy. It is about clearing the routine pile while a person approves anything risky.
- Start with one workflow, measure it for a few weeks, then grow.
“AI agent” has become a word that means everything and nothing. So here is something concrete: one morning in an operations team, with one agent added.
“AI agent” has become a word that means everything and nothing. Here is something concrete.
One request, handled by an agent
How it works
One request, handled by an agent
The steps, in words
- A request arrives: “Please change the delivery address on order 4821.”
- The agent reads and plans: Find the order, check it has not shipped, update the address, confirm.
- It checks your systems: Order status from your order system, through a permitted connection.
- It proposes the change (a person decides): Shipped orders, refunds or bulk changes wait for a person.
- It updates and replies: Record updated, customer told, everything logged.
9:00: the pile
Forty requests overnight: address changes, “where is my order?”, invoice copies, a complaint, two new enquiries, a supplier asking about a payment.
9:05: the agent sorts it
It reads every request and sorts them:
| Type | Count (example) | What the agent does |
|---|---|---|
| Simple updates | Many | Does them, within its permissions |
| Status questions | Many | Answers from your systems |
| Invoice copies | Some | Sends them from your accounting system |
| Refunds and exceptions | A few | Prepares, then waits for approval |
| Complaints | A few | Hands to a person with a summary |
9:30: the approvals
The operations manager opens a short list: each proposed action, the reason, and the data behind it. Approve, edit or decline, in seconds each.
11:00: the routine pile is done
What remains are the conversations that need a person: the complaint, the supplier, the new enquiries. That is where the team’s morning goes now.
The rules that make it work
- Clear permissions. The agent can only use the systems and actions it has been given.
- Approval for anything risky. Money, deletions, messages to many customers.
- Ask when unsure. Uncertainty goes to a person, never a guess.
- Everything logged. Who asked, what the agent did, who approved.

We have built this
AI agents, with a person in the loop
We design, build and run AI agents for real businesses: they read each request, check your systems, update the record, and ask a person when they are not sure.
Anything risky waits for approval. Start with one workflow, measure it, then grow.
Day one, week one, month one
Day one
the agent proposes, a person approves everything.
Week one
simple updates go through on their own; everything else is still approved.
Month one
measure time saved and mistakes caught, then choose the next workflow.
Frequently asked questions
Which AI do agents run on?
Agents can run on Claude, GPT or other models, with your data and your rules. The choice depends on the task, cost and where your data must stay.
Can the agent use our existing systems?
Yes, through their APIs or secure connections such as MCP, limited to the actions it needs.
What if it makes a mistake?
Risky actions need approval, and everything is logged, so mistakes are caught before they matter and corrected quickly.
How long does a first agent take?
One workflow with approvals is a focused project of weeks. Our free automation audit tells you which workflow to start with.
Want this for your business?
Ask ET on our website, or book a free 30-minute call with a project lead, not a salesperson.

