Questions and answers

What does an AI agent need to work well in a business?

Updated:

The short answer

Treat it as a new employee, not a new program. An AI agent reads, decides and acts. So it needs what any new employee needs: written instructions, a memory, its own desk and keys, and a second pair of eyes with a person who signs.

Why a new employee

You install a program once and it does the same thing every day. An AI agent is different: it reads, decides and acts. It can answer a customer, update a record, chase a late payment or prepare tomorrow’s schedule.

That makes it closer to a new employee than to a new program. And a new employee who gets no onboarding does poor work, however talented.

We run agents every day in our own work. They follow up on new enquiries, track open tasks, prepare daily updates and maintain the open-source tools we publish. They are useful because of four things you would give any new employee. Which AI sits inside matters less.

Written instructions

A new employee gets a job description. They learn how you answer customers, what you never promise, and who to ask when unsure.

An agent needs the same thing in writing. When an agent disappoints, the AI inside is rarely the cause. The agent was given two lines of instructions, or two instructions that contradict each other, and it guessed.

Instructions also age: prices, people and rules change. We built a tool that checks agent instructions for mistakes, conflicts and missing parts, the way a spell checker checks a letter. It is open source, so anyone can inspect and use it, and it is downloaded more than 100,000 times a month.

Ask your vendor: Where are the agent’s instructions written? Can we read them? Who updates them when our business changes?

A memory

Think of an employee who forgets every morning what you told them yesterday. You would explain the same pricing rule fifty times.

Many agents work exactly like that. Each conversation starts from zero.

Ours keep a written record. When we make a decision, it is saved in our own words, with the date. Before an agent makes a change, it looks up what was already decided. That way a rule is stated once, not fifty times.

Ask your vendor: If we correct the agent today, will it still know next month? Where is that kept, and can we see it?

Its own desk and its own keys

You do not hand a new employee your personal phone and your bank password. You give them a desk, a user account and access to what the job needs.

An agent should get the same. When we build an agent, it gets its own workspace and its own user account, with access only to what its task needs. For example, an agent that handles only supplier invoices sees only supplier invoices.

The desk can be in your building. An agent can run on your own computers or in your own cloud account, next to the information it works on. Sensitive information can stay in-house.

Ask your vendor: What exactly can the agent open and change? Does it run on our systems or on yours?

A second pair of eyes, and a person who signs

In a well-run office, the person who prepares a payment is not the person who approves it.

We build agents the same way. One agent does the work. A second one, which sees only the finished work, checks it. Then a person approves the actions that matter. Spending money, deleting something or publishing under the company name waits for a yes from a person.

You decide where that line sits: which routine actions the agent may take alone, and which wait for approval. A good setup makes it easy to move the line as trust grows.

Ask your vendor: Who checks the agent’s work before it reaches us? Which actions does it take alone, and which wait for approval?

How to start

The technology is the smaller part of the work. The larger part is the onboarding: writing down how your business works, deciding what the agent may touch, and choosing where a person has the final say.

That is why we start every project the same way. We learn where your day gets stuck before we propose anything. Then we build the agent around that. You get instructions you can read, a memory you can inspect, access you control and approvals where you want them.

An agent set up this way takes the follow-ups, the reminders and the copying between systems off your team’s desks.

Have a question about your own business?

Tell us how you work today. We will answer in plain words.