What AI accounting costs you in effort

Proposals for accounting automation account for the subscription and, if you are lucky, implementation. The effort costs sit outside both, land on people who did not choose the project, and are the reason a well-priced implementation still feels expensive.

None are reasons not to proceed. All are better planned than discovered.

Before: data preparation

When: before anything is configured. Who: your finance team. Roughly: two to three days for a small finance function.

Deduplicating suppliers, pruning the chart of accounts, clearing the historical backlog, resolving inconsistent coding. See preparing your data before automating.

Skippable, and the cost reappears larger during the first two months as a permanently long exception queue.

Before: describing the process

When: before configuration. Who: whoever owns the process. Roughly: half a day per process.

Writing down how it actually works, including the exceptions and the branches nobody has articulated. Frequently surfaces decisions nobody has made, which is the valuable part.

During: the parallel run

When: one full cycle. Who: the whole team. Roughly: the process done twice, plus comparison time.

The largest single effort cost, and the one most often cut. Cutting it removes your ability to know whether the automation agrees with your existing treatment — which is the entire point of the exercise.

During: months one and two

When: the first two months live. Who: whoever handles exceptions. Roughly: more time than the manual process, temporarily.

A long queue containing historical backlog, frequent corrections, and a team learning a new way of working. Real, temporary, and the reason people conclude in week three that it does not work — see what month two of automation looks like.

Ongoing: the controls

When: forever. Who: a named owner. Roughly: fifteen minutes monthly, two hours quarterly, half a day annually.

Sampling, exception trend monitoring, threshold review, keeping documentation current. Small, permanent, and the first thing dropped when busy — which is precisely how automated functions drift. See keeping automated books healthy.

Ongoing: the knowledge

When: continuous, spiking when someone leaves. Who: the process owner.

Someone has to understand why the thresholds are what they are. When that person leaves without documenting the reasoning, their successor either changes nothing out of caution or changes things without understanding — both bad.

The mitigation costs almost nothing: write down the reasoning, not just the setting.

Which ones go away

Do: data preparation, process description, the parallel run, the months one and two transition. One-off, front-loaded, finite.

Do not: the controls and the knowledge. Permanent, small, and they determine whether this still works in three years.

Most planning accounts for the first group and none of the second, which is why implementations succeed and then quietly degrade.

The honest shape of the first year

Months −1 to 0: preparation. More work, no benefit yet. Month 1: parallel run. Significantly more work. Month 2: live, still more work than before. Month 3: roughly break-even. This is where it stops feeling like a mistake. Months 4–12: the benefit, plus a small ongoing control overhead.

Roughly three months before it feels worthwhile. Telling people that in advance is the single cheapest thing an implementation can do, because a team expecting three months interprets month two correctly and a team expecting immediate benefit does not.

On subscription cost

Pricing models vary too much to generalise usefully, and what any business pays depends on volume, scope and how much implementation support it takes.

The point worth making is structural: the subscription is usually not the largest cost in year one, and it is usually the only one in the proposal. A business that budgets only for it will find the project more expensive than expected, not because the price was wrong but because the effort was never counted.

Common questions

What does AI accounting cost beyond the subscription?

Data preparation of roughly two to three days before anything is configured, half a day per process to document how it actually works, a parallel run where the process is done twice for a full cycle, a transition period of one to two months where the team does more work rather than less, and permanent ongoing controls of about fifteen minutes monthly and two hours quarterly.

Which of these costs are permanent?

The ongoing controls — sampling, exception monitoring, threshold review, keeping documentation current — and the knowledge cost of someone understanding why the configuration is as it is. Data preparation, process description, the parallel run and the transition are one-off. Most planning accounts for the one-off costs and none of the permanent ones, which is why implementations succeed and then gradually degrade.

How long before automation feels worthwhile?

About three months. Preparation happens before any benefit, month one involves a parallel run with the work done twice, month two is live but still heavier than before as the backlog clears, and month three is roughly break-even. Setting that expectation in advance is the cheapest thing an implementation can do.

Is the subscription the main cost?

Usually not in the first year, and it is usually the only cost in the proposal. Effort costs — preparation, parallel running and the transition — typically exceed it, which is why a well-priced implementation can still feel expensive to a business that budgeted only for the licence.

Related: what month two of automation looks like · keeping automated books healthy · proving the case to your finance director