How Do I Run My Back Office on AI Agents?

You run a back office on AI agents by converting one named process at a time, never the department in one go. Write the function down as five to eight processes, each with a trigger, an input, a rule, an output and one system of record, rank them by monthly volume multiplied by rule density, then rebuild the top one as a queue with an agent doing the work, scoped write access into the system of record, and a named person owning the exceptions. Budget six to eight weeks per process. There is no platform you buy that does this for you, because the hard part is the decomposition and the ownership, not the model.

Why companies get this wrong

The usual attempt is a department sized one. Someone says the back office should run on AI, a vendor demos an assistant that sits on top of everything, and six months later there is a chat box that answers questions about invoices while every invoice is still keyed by a person. Nothing moved, because a chat box is a retrieval product and a back office is a queue. The work is items arriving, being decided, and landing in a system. Until an agent writes into that system, you have bought commentary on the work rather than the work.

The second mistake is starting where the pain is loudest instead of where the rules are clearest. The noisiest process in most companies is the one with angry humans on both ends and no written rule anywhere, which makes it the worst first candidate. I have seen a finance team spend four months on supplier dispute handling, the process everyone complained about, and get nowhere, while three thousand supplier invoices a month with a documented three way match sat untouched next to it.

The framework I use with clients

Five steps. The order matters more than the tooling, and the first two happen on paper before anyone opens a model.

One: write the back office down as processes, not departments. A department is a budget line. A process is a thing an agent can be given. For each one, name five fields: what triggers it, what arrives, what rule decides it, what comes out, and which single system holds the answer afterwards. Most mid size companies land on five to eight processes per function. If a process needs a paragraph to explain the rule, split it into two until each half fits one sentence.

Two: rank by volume multiplied by rule density. Volume is items per month. Rule density is the share of those items decided by something written down rather than by someone deciding how they feel about it. Both are needed. A thousand items a month decided on judgement is a bad first project, and forty items decided by a perfect rule is not worth the integration. My working thresholds:

Three: rebuild the top process as a queue, and run it in shadow for four weeks. Items arrive into a queue. The agent produces the answer. A human still does the real work in parallel and the two answers are compared every day. Four weeks, because one week is noise. The promotion bar goes in writing before you start: on rule shaped work with documentary evidence behind it I want 90 to 95 percent agreement across two full cycles before the agent touches production. Write that number down in advance, or a sponsor with momentum will argue a 70 percent agent into live.

Four: give the agent the systems, not the screenshots. One system of record per process, scoped write access with its own service account, and every action logged with the inputs that produced it. This is the step that turns a demo into operations. If the agent cannot post the journal, raise the credit note, or move the ticket to closed, a person is still doing the job and you have automated the thinking rather than the work. Scope the credentials tightly, because an agent with your finance team's full permissions is a control failure waiting for a bad week.

Five: staff the exceptions desk on purpose. Between 10 and 20 percent of volume will not go through the agent, and that is the design rather than a defect. Name one person per process who owns that queue, with a service level on it. The exceptions desk is also where you learn: every exception is either a rule you had not written down or a genuine judgement call, and sorting them into those two piles each month is what moves the automated share up over the following quarters. Expect nine to twelve months to cover a full function at one process every six to eight weeks.

From my operating seat

Inside one client I run, the pipeline research that used to eat a morning now happens overnight, and the team arrives to a queue of briefed accounts rather than a list of names to look up. Cash runway is live in the same system instead of being rebuilt in a spreadsheet each month end, and onboarding runs in the joiner's own language on day one. None of that arrived as a platform. Each was one process, decomposed, shadowed, then promoted, with one person named against it.

Twenty years in the C-suite, three exits and three FT100 listings taught me what this looks like from the other side of a data room. A buyer does not pay for an AI story. They pay for a function that runs with fewer people and can be handed over, which means documented processes, an auditable log of what the agents did, and no single person holding the knowledge in their head. The decomposition work above produces exactly that documentation as a side effect. That is the part I would do even if the agents never got built.

Will AI agents replace my back office team?

They replace roughly half the task volume and change what the rest of the team does. On a process with real rule density, 40 to 70 percent of the human hours come out, and the remainder sits in exceptions, judgement calls and the supplier or customer conversations a person handles better. Six people processing items becomes three people running an exceptions desk and owning agent outputs. In a growing company it usually shows up as headcount you did not add rather than headcount you removed.

What does it cost to run a back office on AI agents?

Budget per process and count both halves. Model and infrastructure spend on a mid volume process is normally the cheapest line on the sheet. The real costs are the integration into the system of record, the review time of whoever signs the output, and maintenance at 20 to 40 percent of build every year, because upstream systems change and an unmaintained agent degrades quietly until someone spots a wrong output in front of a customer. Put that total next to the fully loaded cost of the hours removed and you have the only comparison worth making.

Do I need to replace my back office software to use AI agents?

No, and replacing it first is the most reliable way to lose a year. Agents work against the systems you already run, through the API where one exists and a scoped service account where it does not. What matters is that each process has exactly one system of record the agent writes into, so there is a single place to audit what happened. If a core system has no API at all, fix that one integration point. Do not run a migration programme before a single agent goes live.

Where I come in

This is the work I do inside companies: sit with the function, write the processes down, rank them, then build and run the agents that take the top one, with the exceptions desk and the reporting attached. The same method carries across finance, sales pipeline, onboarding and compliance evidence. If your back office is five people keying items that a written rule already decides, the first process is worth converting before the next hiring round rather than after it. See how I work as an AI agent operator, or start at meethayat.com.