What Makes a Business Task Ready for an AI Agent?

Monday's management meeting needs a margin explanation. Finance also has overdue invoices waiting for follow-up, while Support keeps chasing Billing to finish service changes. Which job should go to an AI agent first?

Consider this as a planning exercise. You could rank the jobs by the hours they consume and pick the biggest one. But that leaves out a fairly expensive question: how much work will your team still have to do after the agent finishes?

A report produced in seconds is a poor bargain if Finance spends the afternoon rebuilding it. A less impressive job might be the better starting point if someone can quickly tell whether it was done properly.

Start with the person who receives the work

The finance manager needs to explain the margin change in Monday's meeting. They need comparable revenue and cost figures, the adjustments used and an explanation they can defend. If the agent supplies a confident paragraph without those connections, the manager still has the investigation to do.

The collections owner needs a different output: which balances need a reminder, which are disputed and which might already have been paid. Writing the emails is only part of that job.

Support needs the requested service changed and the customer told what happened. Instructions for making the change may be helpful, but somebody still has to carry them out.

Before looking at products, ask the person receiving the work what would let them use it. Their answer exposes the gap between a convincing demonstration and a job they can actually hand over.

A useful work assignment names the information needed, the expected result and the decisions people retain. Those details make different candidates comparable before the team commits to an implementation.

A report may need more preparation than an action

The margin report looks like the safer project because its output is a document. Suppose, however, that Finance and Operations use different definitions of services revenue. One system records delivery costs immediately; another allocates them after month end. The agent can assemble a report before those differences have been resolved.

Which definition should the report use? That's a management decision the agent cannot resolve by writing a better paragraph. Finance and Operations have to settle the reporting period, definitions and source totals.

A recurring service change might be further along. The policy exists, account records are reliable and Support knows how to check the finished change. It needs permission to act, of course. Even so, the company may be better prepared to delegate that job than its supposedly simple report.

The receivables queue sits between these cases. An aging export may be accessible while payment matching remains incomplete. The first useful assignment could therefore be preparation with explicit unresolved items, rather than sending reminders. That smaller result can be worthwhile if it removes enough work from the collection owner.

Put that preparation in the estimate. Otherwise, people will spend the pilot doing work the project was supposed to take off their hands.

Product labels do not settle the assignment

A chatbot may answer a policy question. An assistant with connected tools may read account records and act. An agent may intentionally stop at a prepared result because the business wants to retain a particular decision.

Return to the invoice queue. Can the product read current receipts? Does it separate a disputed charge? Can it only draft the reminder, or can it send it? Those answers tell you more about the proposed job than the product category does.

For agents, chatbots and assistants, keep the invoice request fixed while comparing what each deployment can actually do. The differences become easier to inspect than vendors' category names.

Once the job is clear, the product discussion gets easier. You can compare what each offer delivers and what your team will have to operate afterward.

Run the awkward case while the pilot is still small

For the report, introduce a total that disagrees with the supporting figures. An acceptable response may preserve the disagreement and prepare a reconciliation question. Silently selecting the more convenient figure produces a prettier report and a worse decision.

For the invoice queue, introduce a payment received after the export. Inspect whether the proposed reminder survives a current-state check. If the system cannot access recent receipts, its result needs to preserve that limitation for the reviewer.

For the service change, consider a hypothetical timeout after payment has been verified. The team needs to establish whether the service changed before allowing a retry. A case that reaches a specialist with the relevant investigation attached may still remove useful work. A case that forces the specialist to start again may not.

Watch how much help the reviewer needs. If the original process owner has to explain every output, the pilot may depend on that person's attention just as much as the old process did.

Compare the work remaining next month

During a pilot, a specialist may repair the data, the founder may review every result and someone may fix a connection before anyone notices it broke. Count those contributions. They won't necessarily be available every Monday.

For each candidate, compare the whole process before and after delegation: collecting inputs, producing the result, reviewing it, correcting it and following it through. Include the maintenance needed when a policy, reporting definition or source system changes. Do not treat hypothetical staff minutes as measured savings.

A job becomes a better candidate when recurrence spreads those costs across useful results. High volume cannot compensate for an output the team cannot accept. Equally, a less frequent report may justify the work if it supports an important recurring decision and meaningfully reduces preparation.

Choosing the first assignment comes down to a concrete decision: this person will accept this result from these inputs, and unresolved cases will return with enough information to continue. That decision applies to reporting, receivables and service work alike.

Take that job into the pilot. Have its recipient use the output, return the bad cases and record the work still required. You'll learn more from that than from another demonstration with clean inputs.