Calculating payback period on an AI tool that you can defend

5 days ago
3 min read

Introduction

Payback period is the right measure for a small business. Return on investment over five years assumes a five-year horizon that most owners do not have and that the pace of change in this area does not support. How many months until this has paid for itself is the question that actually gets asked, and it is the one worth answering carefully.

Answering it carefully mostly means being honest about the numerator. The recurring subscription is easy and small; the setup cost, the time to learn it, the productivity dip during transition and the ongoing review burden are larger and routinely omitted. A payback calculated on subscription alone is not wrong so much as irrelevant.

1. Calculating payback period on an AI tool starts with the full cost

Not the licence fee.

Subscription, setup, data preparation, integration, training hours, and the productivity dip while people learn. For most implementations the one-off costs exceed the first year of subscription.

2. Include the ongoing cost of running it

The part that continues.

Reviewing output, handling exceptions, administering accounts, correcting errors. This is a permanent reduction in the benefit and it belongs in the recurring side of the calculation rather than being ignored.

3. Use the benefit you can actually evidence

Conservative wins arguments.

Time you can measure and money you can trace. Speculative benefits — better decisions, improved morale, future capability — are real and unquantifiable, and putting numbers on them makes the whole calculation dismissable.

4. Convert time to money honestly

The step that decides credibility.

Freed hours only count if they became revenue, or avoided a hire, or replaced overtime. If they became a less pressured week, they are worth something and not money, and the calculation should say so.

5. Expect the benefit to build

Not a step change.

The first month is negative, the second is neutral, and the benefit reaches its steady state after several months. Modelling it as immediate produces a payback figure that will be missed and then disbelieved.

6. Set the threshold before you calculate

Prevents rationalisation.

Decide what payback period would justify the investment — twelve months, eighteen — before doing the sums. Otherwise the acceptable period is whatever the calculation produces.

7. Treat a long payback as a warning about the horizon

Specific to this area.

A three-year payback assumes the tool, the vendor and the requirement are all still there in three years. In a market changing this quickly, that is a substantial assumption and it argues for shorter-payback projects.

8. Recalculate at six months against what happened

The check that matters.

Actual costs and actual benefits against the model. This tells you whether to continue and, more usefully, how wrong your estimating is, which improves every subsequent decision.

9. Keep the calculations and compare them over time

The compounding benefit.

After four or five of these you know your own bias, which is almost always optimistic on benefit and optimistic on setup cost. Correcting for a known bias is the cheapest improvement available to any decision process.

Be careful about comparing payback across very different projects. A short payback on a small saving may matter less than a longer payback on something that removes a constraint on growth, and the number alone does not capture that.

Conclusion

Count the whole cost, because payback calculated on the subscription alone tells you nothing.

Include setup, data preparation, integration, training and the transition dip, add the permanent cost of reviewing output and handling exceptions, use only benefits you can evidence, convert time to money only where it genuinely became money, model the benefit as building over several months rather than arriving immediately, set your acceptable payback period before doing the sums, treat a long payback as a warning given how fast this area moves, and recalculate at six months against what actually happened.