An AI agent for back office work

Count the exceptions, not the repetitions

The instinct is to automate whatever happens most often. For back office work that is the wrong first measurement.

A task performed 200 times a week where 40 of those need someone to look at them and decide is a poor candidate: the agent handles 160, and a person still has to stay in the loop for all 200 to spot which is which. A task performed 40 times a week with 2 exceptions is a much better one. The exception rate determines the return, and the volume does not.

Where an agent beats a plain integration

This is worth being honest about, because the cheaper answer is often the right one. If the mapping between two systems is fixed — this field goes to that field, always — build the integration. It is cheaper, faster and more reliable than anything with a language model in it.

An agent earns its place where the input is not consistent:

• A supplier who redesigns their invoice layout twice a year, breaking any template-based extraction.
• A field that has to be inferred from a free-text note rather than read from a column.
• Documents arriving in whatever form the sender felt like — PDF, photo of a printout, text in the body of an email.
• A process with rules that apply differently depending on context a rigid integration cannot see.

What it should not do on its own

Release money. Preparing a payment, matching it to an order and flagging what does not add up is useful. Executing it is a policy line worth keeping even where the technology could cross it.

Guess when unsure. In customer-facing work a slightly wrong answer is recoverable in the next sentence. In back office work a wrong entry propagates into reports and surfaces a month later, when nobody remembers the context. The escalation threshold belongs deliberately low here.

Work around a broken process. If figures are copied between two systems because a decision about which one is authoritative was never made, automating the copying makes the ambiguity permanent and harder to see.

The audit trail is part of the build

Every action gets logged with the input it saw, what it decided and on what basis. Not because a regulator asked, but because someone eventually will ask how a particular figure came to be where it is, and “the agent did it” is not an answer that survives that conversation.

This is also what makes an agent easier to trust than the spreadsheet macro it usually replaces, which typically has no log at all.

What it connects to

Whatever holds the data: accounting or ERP, document storage, the shared inbox where things arrive. This is the use case where the state of your systems matters most — an application with no usable interface turns a four-week build into a different project, and that is worth establishing before anything else.

How to tell whether it worked

• Time spent on the task per week, before and after.
• Error and rework rate — entries that had to be corrected later.
• Share of cases escalated to a person, tracked over time.
• Backlog age: how old the oldest unprocessed item is.

The third one should fall over the first weeks as edge cases get turned into rules, and then flatten. If it does not flatten, the process was less defined than everyone believed.