AI for accountants

January is not brutal because clients are difficult. It is brutal because nothing prompts anybody until the deadline does, so a year's work arrives in a fortnight. That, rather than posting transactions, is where AI for accountants earns its keep in a practice — bank feeds and rules already handle most of the posting.

A practice's problem is the client base, not the ledger

Chasing is a full-time job nobody has. Records, receipts, queries, approvals. Each one is a small polite message and there are hundreds of them, and the practice's deadline risk is entirely built out of them.
The same query, forty times. Every quarter, the same handful of questions to different clients. Individually trivial, collectively a week.
The work arrives in a lump. Not because clients are difficult but because nothing prompts them until the deadline does. The peak is manufactured, and it is manufactured by the absence of a system.
Nobody sees a client drifting until renewal. Fewer transactions, later responses, a missed VAT quarter. All visible months before the client leaves, in data the practice already holds.

What gets built across the client base

Chasing that runs itself: who owes what, contacted on a schedule in the practice's tone, escalating to a person only when a client goes quiet or replies with something real.
Query batching: the routine questions asked once per client, answers collected in structured form, and only the exceptions landing on a manager's desk.
Document intake that handles what clients actually send — photographs, PDFs, a forwarded email, a carrier bag of receipts scanned badly — and returns what is missing rather than what arrived.
A practice-level view of risk: deadlines at risk, clients slowing down, work not started. One short list, daily, not a dashboard.

What stays with a person

The filing itself. A return goes out under someone's professional judgement and that does not move to a machine.
Anything that looks like advice. The line between prompting a client and advising a client matters, and the system stays firmly on the prompting side.
Treatment decisions on anything unusual: a transaction that does not fit, a related-party question, anything where the answer depends on facts the machine cannot see.
The decision to disengage a client. It can show you who is drifting. It should not send that letter.

Hold off if

Client records are not centralised enough to know who owes what. That is the first build and it is where most of the value turns out to be.
Under about twenty clients. At that size the practice manager is the system, and a build will not beat them.
You are mid-migration between practice management systems. Wait, then build once.

Where these parts already run

Advertising manager — the many-at-once shape, in production: roughly 600 campaigns pulled, merged and judged every morning, each with a plain verdict on it. A client base is the same problem with different objects.
Data platform — the part that makes any of it trustworthy: several systems reduced to one true record per item, so the chasing is not built on two contradictory answers about what a client has already sent.

Questions from practice managers

Does AI for accountants mean automated bookkeeping?

Partly, and that part is already well served by bank feeds and rules. The bigger win in a practice is upstream: chasing records, batching queries, handling whatever clients actually send, and seeing deadline risk early.

Will it talk to clients directly?

For chasing and routine queries, yes, in the practice's tone and with an easy way to reach a person. Anything that reads as a question about tax treatment stops and goes to a manager. Keeping that line clean is a design decision, not a setting.

Can it read the receipts clients send?

Yes, including bad photographs, which is most of them. The useful output is not the extracted data — it is the list of what is still missing, sent back to the client the same day rather than found in week three.

How does this change January?

It flattens it, by making the prompting continuous instead of leaving the deadline to do the prompting. Practices that get this working report the peak moving earlier and getting shallower, not disappearing.

Do our clients need to use anything new?

No, and insisting they do is why most practice automation fails. It works with email, messages and photographs because that is what clients actually send.

Count one week of chasing, then decide

Every message whose entire purpose was asking a client for something they had already agreed to send. Count them for a week and multiply by fifty. Most practice managers find that number uncomfortable, and it is the whole business case, made without me. If you want a second opinion on what to do about it, the first hour is free.

Book that hour →