What an AI agent needs to work inside a business

The difference emerges inside the workflow

In ThinkUp, I described how Mr. Anderson began working within our engineering process, using Discord, Jira and the context of our previous decisions. What mattered to me was that the agent could find its place in that system and understand how we worked.

That account describes our own experience at SourcingUp and CommerceUp. It does not establish what an agent will do in another organization, or mean that a working demo is a reliable day-to-day operation.

Four things to define

1. A verifiable task. Resolving a ticket has a clearer scope than an instruction such as improving the company. There has to be a way to recognize if the work turned out well.
2. Context the agent can retrieve. Rules, previous decisions and documentation help interpret what is requested. It is also necessary to distinguish what information is still valid.
3. Appropriate tools and permissions. Reading documentation and modifying a system are different actions. Access must respond to what the task requires.
4. A defined review process. You have to decide who checks the result and what happens when information is missing or an error appears.

What does continuity change?

Part of working in a company is remembering why earlier decisions were made. With that context available, we can check whether an agent respects constraints and makes appropriate use of previous decisions. We can also see where the documentation needs correcting.

That edition described a run lasting more than 35 hours. It was an observation from our work, not a general benchmark or a promise of autonomy. When evaluating an agent, I care about what it resolved and how much work it left for the team, as well as how long it ran.

What I would check before expanding its scope

I would review completed tasks, errors, human interventions and rework. I would also test how the agent handles an incomplete instruction or a failing tool. If we cannot reconstruct its actions, it will be difficult to improve the process.

Define autonomy task by task. A team may let an agent prepare a proposal while requiring approval before it is applied. That boundary needs to be enforced in the workflow, not simply written in an instruction.

Start with a part of the work

For a business getting started, I prefer a focused workflow, identified sources and a person responsible for it. That gives you a way to learn what an agent contributes and where the team still needs to be involved. If you have not chosen the first process yet, here is an approach to start with AI from a specific problem.