Skills that matter when AI does the processing

Advice to accountants about AI usually arrives as "learn data skills", which is vague enough to be useless and slightly wrong. The skills that actually appreciate are not technical. They are things the profession has always valued informally and never taught systematically.

Six of them, in rough order of how much they matter.

1. Knowing immediately when something is wrong

This is the load-bearing skill, and it is the hardest to fake.

When you review rather than produce, your entire value rests on catching what the system got wrong. A proposal that is plausible and incorrect is the dangerous case, and the only defence is an internal sense of what the number should be.

It comes from calibration — knowing what this supplier normally bills, what this cost normally runs at, what this margin normally is. Previously that arrived as a by-product of doing the work by hand. Now it has to be built deliberately: look at whole reports rather than just your queue, look at comparatives, and make yourself predict a figure before you look at it.

How to build it: before opening the month's numbers, write down what you expect gross margin to be. Then look. The gap is your calibration error, and it shrinks with practice.

2. Describing a process precisely

The constraint on automation is no longer technical capability. It is that most processes exist only in people's heads, in a form too vague to automate.

"We check it before paying" is not a description. Who checks, against what, with what tolerance, and what happens when it fails — that is a description. The accountant who can produce the second version can get a process automated correctly. The one who produces the first gets an automation that does not match reality.

This is now a practical skill rather than a documentation chore, and it is covered properly in why accountants should learn to describe processes.

3. Investigating rather than clearing

An exception is a small mystery. The skill is in resolving the cause, not the item.

A payment short by RM 340 can be cleared with a write-off in ten seconds. Or you can find out that a customer has been applying an unauthorised settlement discount for four months across every invoice, which is a different-sized problem with a different-sized recovery.

The habit that separates the two: before resolving anything, ask why did this happen and has it happened before. That question turns exception handling from clerical work into something worth paying for.

4. Explaining figures to people who do not read accounts

When the pack assembles itself, producing it stops being the contribution. Interpreting it starts being the contribution — and interpretation only has value if the person receiving it understands.

"Margin fell 2.1 points" is a fact. "We sold more of the low-margin range because of the promotion, so revenue looks good but we made less money on it than last month" is an explanation somebody can act on.

Most accountants are better at this than they think and do it less than they should, because producing the numbers used to consume the time.

5. Judging what to automate and what not to

Someone has to decide which processes go, where thresholds sit, and what stays human. Those are accounting judgements, and if an accountant does not make them, someone without accounting judgement will.

The specific competence is knowing which processes carry judgement that looks mechanical from outside. Period cut-off looks like a date field. Anyone who has closed a set of books knows it is not.

6. Working out whether the automation is still right

Automated processes drift — not because the software changes, but because the business does. A threshold set when your average invoice was RM 3,000 is wrong when it becomes RM 12,000. A category that made sense before you added a product line no longer does.

Nobody is assigned this and it is nobody's obvious job, which is exactly why it goes undone. The accountant who periodically checks whether the automation still reflects the business is doing something quietly valuable.

What does not appear on this list

Programming. Not required, and increasingly beside the point as systems are configured in plain language.

Data science. Interesting, rarely necessary. You need to interpret output, not build models.

Tool-specific certification. Useful for a job application, worth little beyond it. Products change; the six skills above transfer.

The honest framing

None of this is new. Good accountants have always caught wrong numbers, explained results and known which processes were fragile. What changed is that these used to be the part of the job that got squeezed by processing, and now they are the job.

That is good news for anyone who found the processing the least interesting part, and genuinely difficult for anyone whose value was throughput.

Common questions

What skills do accountants need for AI?

The ones that appreciate are recognising immediately when a figure is wrong, describing a process precisely enough to automate it, investigating the cause behind an exception rather than clearing the item, explaining results to non-financial colleagues, judging which processes should stay human, and periodically checking that automated processes still match the business. None require programming or data science.

Do accountants need to learn Python or data science?

Generally no. Modern systems are configured by describing what you want rather than by writing code, and the accountant's contribution is judging whether output is correct and deciding what should be automated — neither of which requires building models. Programming is a reasonable interest but it is not the constraint on an accounting career.

How do I build judgement about whether a number is right?

Deliberately, since it no longer arrives as a by-product of manual processing. The practical exercise is predicting figures before looking at them — write down what you expect margin or a cost line to be, then check — and reviewing whole reports with comparatives rather than only the items in your queue, so you see the distribution rather than isolated transactions.

Is exception handling a skilled job?

It is when done properly, because each exception is a question about why something did not fit the expected pattern, and the answer is frequently a real problem such as an unauthorised deduction, a price increase nobody communicated or a process being bypassed. Clearing items quickly is clerical; finding and fixing the cause is not, and the difference is largely a matter of habit.

Related: why accountants should learn to describe processes · when the accountant becomes the reviewer · will AI replace accountants