How much AI capability you actually need is usually less

5 days ago
3 min read

Updated: 4 days ago

Introduction

The products are sold on capability, which is reasonable, and businesses buy on capability, which is not. A firm with a specific problem — extracting data from four hundred invoices a month — evaluates platforms that could do a great deal more, chooses on breadth, and pays for a capability it will never configure, let alone use.

Overbuying is the more common error by a considerable margin, and it is expensive in two ways. There is the direct cost, and there is the implementation cost, because a capable general platform requires configuration, integration and maintenance that a narrow product does not. A small business is far more likely to succeed with something limited that works than with something powerful that is never finished.

1. How much AI capability you actually need starts from the specific problem

Not from the category.

Write the problem in one sentence with a number in it. Then evaluate against that sentence. Products that solve it are candidates; products that could also do six other things are not thereby better.

2. Resist the platform in favour of the point solution

For a first implementation.

A narrow product that does one thing well can be running in a fortnight. A platform that could do everything requires a project, and small businesses do not reliably finish projects.

3. Count the configuration burden as part of the cost

Where capability becomes expensive.

Flexible systems require decisions: rules, workflows, fields, integrations. Each is somebody's time, and the more capable the product the more of those decisions it demands before it does anything.

4. Check what proportion of the features you will use

An honest estimate.

If the answer is a tenth, you are paying for the other nine tenths in both money and complexity. Sometimes that is a reasonable bet on future needs; more often it is a bet that does not pay.

5. Be sceptical about future requirements

They rarely arrive as predicted.

"We might need this later" justifies most overbuying, and later usually brings a different requirement and a better product. Buying for a need you do not yet have is a poor trade in a fast-moving market.

6. Prefer something your team can operate

Capability that requires a specialist is not capability you have.

A product your existing people can configure and maintain is worth more than a superior one that depends on external help for every change. This is the constraint that decides most outcomes in small businesses.

7. Watch the integration requirement

Where the real cost hides.

The more a product does, the more it needs to connect to. Integration is consistently the largest and least-estimated part of implementation, and a narrow tool with one connection is a much smaller undertaking.

8. Start below what you think you need

The safer error.

Underbuying is correctable by upgrading, at a known cost. Overbuying leaves you with an unfinished implementation and a contract, which is considerably harder to unwind.

9. Revisit after twelve months of real use

The decision improves with information.

After a year you know what you actually do, what you never used, and what you genuinely lack. That is a far better basis for a larger commitment than a specification written before any experience.

Be careful about the opposite error in one specific case: where a requirement is genuinely regulated, safety-related or subject to audit, buying something too limited to meet the obligation is a false economy.

Conclusion

Buy for the problem you have, because overbuying costs money and, more importantly, time.

Write the requirement in one sentence with a number and evaluate against that, prefer a narrow product to a platform for a first implementation, count configuration and integration as part of the cost, estimate honestly what proportion of the features you will use, discount future requirements heavily because they arrive differently, choose something your own team can operate, start below what you think you need because underbuying is easier to correct, and revisit the decision after a year of real use.