Five administrative jobs AI can do on its own

A company’s administration has never had so much software. The ERP, the expense program, the bank’s platform, the electronic invoicing portal, the spreadsheets holding everything together. Yet ask anyone who keeps the books what the month-end close is like, and the answer is the same as ten years ago: long, tiring, full of reconciliations by hand, approvals to chase and the same mismatches to fix every time.

The software does the big pieces. It keeps the journal, brings invoices into a digital flow, shows a dashboard. In between remain hundreds of small tasks that ask a person to copy a value, compare two numbers, write an email when they do not match, chase when nobody replies. People are the glue between the systems. That is the gap AI has to fill: not a new ERP, not a finance brain of the future, but the connective tissue that today is done by hand. Here are five places where it works.

1. Exceptions

It is the job with the highest return, and the one nobody names because it does not sound good. An invoice arrives with no order reference. A transfer comes in and matches no open invoice: it is a few euros short, or the customer paid two invoices in one go. A credit note arrives and nobody knows what it refers to. An approval has been sitting for six days in a manager’s inbox. Each case asks a person to open three programs, investigate and resolve.

With an agent, the exception is picked up the moment it is born. The invoice without an order is searched by supplier, amount and date; if the match is clear, it is attached and moves on; if it is ambiguous, the person receives the two most likely candidates and chooses. The bulk transfer is split over the invoices it covers. The stalled approval gets a reminder, then passes to the deputy. It is not one more flag: it is an investigation already done, bringing the person only the decision.

2. The month-end close

The close is slow not because closing is slow. It is slow because everything upstream is dirty. The first days go on clearing the month’s backlog of exceptions, things that should have been fixed when they happened. Then the reconciliations start, the recurring entries, the variance checks, in sequence, because every step waits for the one before.

The biggest gain is not in automating the close. It is in resolving exceptions every day, so the close starts with clean data. On top of that, an agent prepares the reconciliations, flags the variances, sets up the entries that repeat and writes a first draft of commentary comparing the month with the one before. The job of whoever keeps the books becomes reviewing, instead of doing.

3. Data between one system and another

Bank reconciliation means downloading the transactions, opening the ledger and matching line by line. With several accounts, or several companies, it is hours. The tie-outs between the ERP and the bank, between invoices and payments, between one company of the group and another, are the same work repeated in different places.

An agent takes the transactions from the bank, takes the entries, and matches. Exact matches close on their own. Partial ones, a difference in dates, a receipt covering several invoices, a fee in between, reach the person with the proposal and the reason for the proposal. The person reviews the exceptions, not the whole reconciliation. It is the case where the word “glue” is most literal: almost all the work is moving data and comparing it, and the part that needs judgement is a fraction. That fraction is where people should spend their time.

4. Communication with the outside

In every admin office someone spends hours a week writing the same emails. The reminder to the customer who is late. The request to the supplier for the missing document. The reply to whoever asks “when are we getting paid?”. They are necessary communications, with a precise timing, and they always follow the same pattern, but they are done by hand because no ERP has a “chase this supplier” button.

An agent keeps the schedule and sends the reminders on a rising scale, answers questions about the status of a payment by reading the ledger, asks for the specific document that is missing and waits for the reply. The person steps in only when there is a dispute or when the amount crosses a threshold. In many companies there is someone who does only this all day; it is the perfect case for having nobody do it.

5. Accountant, auditor, bank

Every quarter, every set of accounts, every credit request starts with a list of documents to deliver. The accountant wants the invoices and the transactions of the period, the auditor wants a sample of operations with proof of approvals, the bank wants the accounts and the updated position. Usually it falls to the most experienced person in the office, who is the person who should least be spending days collecting files.

An agent reads the list, works out which system each item lives in and pulls out what it can on its own: registers, lists, samples. For what needs judgement, it prepares a draft and passes it to the right person. The person checks and signs, instead of building the file from scratch.

How you put it into practice

The rules are few and hold for all five jobs. Do not add programs: one layer, on top of and between the ones that are there. Sit next to the person doing the work and map the real process, with its shortcuts, not the manual’s. Build inside the tools the office already uses. Build agents that do the work, not dashboards that show it. Bring a person only the decisions that need judgement, with the context ready.

And the mistakes that kill projects are just as few. Automating one piece at a time without a design of the whole. Building a chat that answers questions about the data instead of something that does the work. Treating it as an IT project instead of an administration project, which is who knows the process. And automating the manual instead of reality.

Where to start

From the process the office complains about most. You map it as it is today, shortcuts included. You look at how much is pattern recognition and how much is judgement: if the former dominates, and it usually does, that is the starting point. Then you measure after a month, with only three questions: am I saving time or money, am I collecting more or sooner, am I reducing a risk.

What we build

The five jobs in this article are the trade of Ditta’s virtual employee. It reads invoices from the SDI exchange and transactions from the bank, matches, records, prepares the journal and keeps the schedule; it writes the reminders and keeps the documents ready for the accountant. When it is not sure, it asks, on WhatsApp or in the app, and learns from the answer.

The starting point was an article by Varick Agents, written for large groups with dozens of people in administration. The five jobs are the same in an office of two.