Understand

AI agents in business

Delegating to AI agents is not about adding one more assistant, but about entrusting complete tasks to software colleagues that see them through. The question is therefore not "which tool should I install", but "what work am I prepared to hand over, and under what limits". This page sets out what works, in what order, and what must be settled before starting.

1. Where to start

The first tasks to hand over share four traits: they recur often, they follow a rule that can be stated in a few sentences, their result can be checked at a glance, and a mistake can be undone without damage. Following up on a quote meets those criteria; signing a contract does not.

Two traps, on the other hand, are expensive. Starting with the company's most complex process, on the grounds that it costs the most: it is the hardest to frame, and failure there is the most visible. And handing over a task whose rule nobody can state: an agent will not guess what the organisation has never put into words.

2. Function by function

Here are the uses that come up most often, and the work actually delegated in each case.

Sales — qualify incoming enquiries, follow up on quotes left unanswered, keep the customer record current after every exchange, prepare meetings.
Support — answer requests already handled a hundred times, gather the context of a case before a human steps in, raise a flag when a customer hits the same difficulty repeatedly.
Finance — match invoices against payments, flag discrepancies, prepare overdue reminders, keep recurring closing items ready.
Operations — watch what should be moving, chase what is stalling, keep the people concerned informed without being asked.
Human resources — sort applications against written criteria, arrange interviews, run through the steps of welcoming a new arrival.

The common thread: these are tasks of coordination between tools, the ones that consume time without producing visible value.

3. A team, not an isolated agent

A single agent charged with everything soon becomes unmanageable: its instructions grow longer, its mistakes harder to attribute. The form that lasts is a team, modelled on a human organisation — defined roles, a scope per role, and a line of escalation for decisions.

Each agent then receives an understandable mission, only the tools that mission requires, and knows who to escalate to for what it cannot settle. That is the approach nullbot takes: you hire agents the way you would hire a team, you give them a role, and you follow their work from a shared dashboard.

4. The conditions to meet

Before handing agents the keys, four points must be settled — otherwise the rollout stops at the first incident.

The autonomy boundary — put in writing what goes out without approval and what requires a human decision.
The budget — set a maximum spend per agent and per period, enforced before the action rather than discovered on the invoice.
Access — grant each agent only the tools its mission requires, no more and no less.
Where the data lives — know which machines the work runs on, and what leaves the company.

These four points have a page of their own: governance and budget control. On the last one, nullbot takes a clear stance — the software installs and runs on your machines, with your own keys, as our security approach explains.

5. Frequently asked questions

How many agents should you start with?

One, on a task you can describe and check. The common mistake is to deploy a whole organisation before establishing that a single agent holds its post. The others follow once the first is reliable.

What happens if an agent gets it wrong?

It happens, as it does with a new colleague. What matters is that the mistake can be undone: sensitive actions go through human approval, and the rest leave a readable trace that lets you understand and correct the rule.

Do we have to replace our existing software?

No. Agents plug into the tools already in place and use them the way a person would. What disappears is the manual back-and-forth between those tools, not the tools themselves.

How long before the first result?

It depends above all on how clear the task is. A rule already written down and a verifiable result give you a useful agent quickly; a process nobody has ever formalised will first demand that work of formulation, which is the real cost of entry.

Going further

Read next: what is an agentic system, or follow the sector in our section devoted to artificial intelligence.