Picture Monday at nine. Your controller opens one page. Every overdue invoice has a follow-up drafted in your firm’s tone, every payment that does not match is held with its evidence, and the three decisions only she can make sit at the top. The review that used to fill her Monday is done before the leadership meeting, and the team spends the week on the clients and questions that need them. That is 10X what a finance team can do, with the people you have.
You have decided to put AI to work on a process. The next question is how far to take it, and you answer it one piece of work at a time.
Three levels for one piece of work.
- Better assistance.The work runs in the same order, with the same people. Each step starts from a prepared draft instead of a blank page.
- A redesigned workflow.The order of the work, the handoffs and the roles change around what AI prepares. People spend their time on the exceptions and the decisions.
- Deeper automation.Steps start without a person starting them, and a narrow class of routine work finishes inside limits a person wrote and can change.
Each level is a destination in its own right. Some work belongs at the first level for good, and some earns the third within a quarter. Two neighbouring choices have their own pages: what to use and who helps you do it.
The process: a Monday receivables review.
An invented 60-person architecture firm reviews its receivables every Monday. The accounts receivable specialist matches the week’s receipts to open invoices, refreshes the aging report, writes follow-ups for overdue invoices and lists anything that does not add up. The controller reads the list, decides each exception and sends a summary to the partners. Project leads own the client relationships, so a follow-up on a sensitive account goes out under their name or not at all.
One case runs through all three levels. A client owes $1,200 on an invoice and pays $1,000. The $200 might be a short payment, a disputed line or a part payment with more to come. Deciding which, and what happens next, belongs to the controller at every level.
One process at three levels.
Illustrative example with invented data. The Monday receivables review at an invented architecture firm.
| Question | Better assistance | Redesigned workflow | Deeper automation |
|---|---|---|---|
| How the work runs | The same Monday, in the same order. The specialist runs a capability on each step. | Exceptions come first. A review pack is prepared before the meeting, and the controller starts with the decisions. | The pack is built at 7:00 a.m. on a schedule, from read access to the accounting system. Routine first reminders go out under a written rule. |
| What AI prepares | A follow-up draft for each overdue invoice and a first-pass list of receipts that do not match. | Provisional matches, the refreshed aging, the exception list with evidence, every follow-up draft and the partners’ summary. | The same pack, plus a log of each reminder sent and each invoice held, with the reason. |
| The $1,000 receipt | Flagged in the specialist’s list. She takes it to the controller. | At the top of the controller’s page, with the invoice, the remittance note and the $200 difference. | Held. No reminder is sent to that client, and the item waits for the controller. |
| What people keep | Everything they hold today. The specialist reviews and sends each follow-up and posts each payment. The controller decides exceptions. | The specialist confirms matches, posts payments and sends follow-ups. The controller decides every exception. Project leads approve messages to their clients. | The controller writes the rule, sets its limits, reads a weekly sample and can switch it off. People still post payments and decide every exception, dispute, credit and second follow-up. |
| What must be true first | A written follow-up standard and one person willing to use it weekly. | Agreed exception types, a named owner and a test on past weeks. | Weeks of review records showing routine reminders approved unchanged, reliable live data and IT’s approval of the access. |
The rule at the third level is narrow on purpose: a first reminder, for an undisputed invoice under $2,500 that is 8 to 30 days overdue, on an account no project lead has marked sensitive. Everything outside it waits for a person.
What each level is right for.
- Better assistance.Right when the time goes into producing something: a draft, a summary, a first-pass check. It suits work where judgment varies case by case, where one person or a small group does the job, and where you are still learning what good looks like. Overdue-invoice follow-up preparation is a capability at this level.
- A redesigned workflow.Right when the time goes into waiting, rework and handoffs between people. If the controller’s Monday is spent finding the exceptions instead of deciding them, drafting faster will not fix it. Reordering the work will. This level needs an owner, written rules and a role that changes. Receivables exception review and the weekly receivables summary are the pieces this redesign is built from.
- Deeper automation.Right for a class of work that is high in volume, stable in its rules, low in cost when wrong and easy to put right. Your own review records are the evidence: when a reviewer has approved the routine cases unchanged for weeks, that review is ready to become a sample. Our article on scheduling AI work into repeatable business routines covers the inputs, timing and missed-run planning this level depends on.
One process can sit at more than one level. This firm automates the first reminder, runs the review as a redesigned workflow and keeps a difficult client conversation at better assistance, where the project lead writes with a prepared draft beside them.
Six questions that choose the level.
- Where does the time go?Into producing the work points to better assistance. Into the gaps between steps and people points to a redesigned workflow.
- Is the standard written down?If good work lives in one person’s head, start with assistance and write the standard as you go. Redesign and automation both need it on paper.
- How often does the work run, and how alike are the cases?Weekly work across a role supports a redesign. Daily, near-identical cases support automation for that class of case.
- What does a wrong result cost, and who would notice?A late courtesy reminder is cheap to correct. A message to a client in dispute is not. Automate only where an error is small, visible and reversible.
- Can the inputs arrive without a person?Automation needs current, trustworthy records and access IT has approved. Files and exports are enough for the first two levels.
- What has the current level proved?Move up on evidence from your own review records and exception log. Stay where you are while the reviewer is still changing what she is given.
The checks that hold at every level.
The level changes how much is prepared and what starts the work. It does not change who decides. At this firm, people post every payment, decide every exception and approve every credit, write-off and payment plan, at all three levels. AI prepares. People decide. sets out that division of work.
Before a level goes live, run it on past weeks with a test set that includes the awkward cases, the $1,000 receipt among them. Keep one exception log. Give the owner the authority to move a step back down a level the week it stops earning its place.
A redesign or an automated step also returns hours to the team. Decide where they go before the rollout, with the worksheet in What will your team do with the capacity it gains?
Your next step.
Start from the outcome you want, map the work across the teams that touch it, and choose a level for each piece. Then prove it on real work and roll it out. All three levels run on the platform your team already has, whether that is ChatGPT, Claude, Copilot or Gemini. You set the pace, and your organization keeps control of how the work gets done.
Bring one process, where it runs today and the result you want from it. We will work out with you how far to take it.