Before You Buy AI: What Exactly Are You Paying to Solve?

AI investment decisions are often made in the wrong order: technology first, business problem second. That can become an expensive way to discover that an impressive solution was solving the wrong thing.

There is a familiar moment in technology projects.

Someone presents a new system.

The demonstration works beautifully.

The slides contain phrases such as automation, productivity, competitive advantage and, increasingly, AI.

There is usually a graph pointing in the correct direction.

And eventually somebody says something close to:

We cannot afford to be left behind.

Perhaps.

But there is another question worth asking first:

What exactly are we paying to solve?

That question is less exciting than the demonstration.

It is also considerably cheaper.

Start with the problem, not the product

Artificial intelligence can be extraordinarily useful.

That is precisely why it deserves more serious evaluation than either enthusiasm or scepticism provides.

The mistake is not using AI.

The mistake is deciding that AI is the answer before establishing what the question is.

A business problem may be caused by poor software, a bad workflow, fragmented data, unnecessary manual steps, inadequate training, unclear responsibility or simply a process that should have been abandoned years ago.

Adding artificial intelligence to any of those does not automatically remove the problem.

Sometimes it merely allows the problem to operate more quickly.

The useful sequence is much less glamorous:

What is the problem?

What evidence shows that it exists?

What is causing it?

What would improvement look like?

What alternatives exist?

And only then:

Does this actually need AI?

That sequence sounds obvious.

I have written elsewhere about the mythology that AI can remove the need for engineering, analysis and operational reality.

In practice, technology procurement has spent decades finding ingenious ways to avoid it.

The demo is not your business

A product demonstration takes place in a controlled environment.

Your company does not.

Your environment contains customers who behave strangely, employees who have discovered seventeen unofficial ways to complete the same task, spreadsheets whose original creators left in 2019, databases that disagree with each other and software integrations held together by institutional memory.

Then there are permissions, security, privacy, exceptions, regulation, training, adoption and the small matter of what happens when the system is wrong.

None of this means the proposed technology is bad.

It means the demonstration is the beginning of the investigation, not the end.

A system that performs brilliantly under demonstration conditions may still be uneconomic, operationally unsuitable or dependent on assumptions that do not survive contact with the organisation buying it.

Due diligence exists to find that out before the invoice becomes the most reliable part of the project.

Claims need evidence

AI is particularly vulnerable to impressive numbers.

A productivity improvement of 30 per cent sounds persuasive.

So does 95 per cent accuracy.

So does a claim that a system will eliminate half the manual work.

But every percentage creates more questions.

Thirty per cent improvement in what?

Measured against which baseline?

Among which users?

Under what conditions?

Does time saved become actual financial value?

What happens inside the remaining five per cent of errors?

Does the system remove work, or move it somewhere else?

A research result can be excellent evidence without being your business case.

A vendor’s result can be genuine without being transferable to your organisation.

The job of management is not merely to collect impressive evidence.

It is to determine whether that evidence supports this decision.

Enthusiasm is not due diligence.

Follow the money beyond the licence fee

Another common mistake is treating the quoted product price as the cost of the project.

The actual investment may also include:

• integration;
• data preparation;
• infrastructure;
• security;
• legal and compliance work;
• training;
• change management;
• human review;
• monitoring;
• quality assurance;
• error correction;
• support;
• maintenance;
• model or API usage;
• and, eventually, the cost of getting out.

The cheapest line on an AI proposal is sometimes the one labelled software.

That does not mean an AI investment cannot produce an excellent return.

It means the return has to survive the complete arithmetic.

What happens when reality disagrees?

One reason I wrote Before You Buy AI was the gap between the way technology projects are discussed before implementation and the way organisations experience them afterwards.

Real-world AI projects rarely divide neatly into success and failure.

Some work extremely well.

Some fail.

Some succeed technically while creating unexpected operational problems.

Some automate work but damage another part of the customer experience.

Some begin with spectacular claims and become considerably more nuanced once deployed at scale.

The guide examines real cases involving McDonald’s, IBM, Klarna, GitHub Copilot and others, alongside research behind some of the productivity claims now routinely repeated in discussions about AI.

The point is not to compile a museum of technology disasters.

It is to ask what management should have known, measured or challenged before reaching the decision.

A kit for making the decision

Before You Buy AI — The Executive Due-Diligence Kit is not an introduction to artificial intelligence.

It contains no collection of prompts and no prediction about machines taking over the world.

It is designed for the much less glamorous moment when somebody has to decide whether an organisation should actually spend the money.

The guide includes twelve practical due-diligence tools covering areas such as:

• defining the business problem;
• testing whether AI is actually necessary;
• interrogating vendor claims;
• evaluating evidence;
• assessing data readiness;
• examining workflows;
• calculating hidden and three-year costs;
• testing ROI assumptions;
• identifying operational and management risk;
• structuring a meaningful pilot;
• and reaching a defensible decision.

The final outcome is deliberately simple:

GO.

REWORK.

NO-GO.

The difficulty is earning the confidence to choose one.

Sometimes NO-GO is the profitable answer

Technology projects create momentum.

Once enough time has been invested in meetings, demonstrations, proposals and internal discussion, saying no begins to feel like failure.

It is not.

If proper due diligence prevents an organisation from spending £100,000, £500,000 or considerably more on a project whose assumptions do not survive examination, the analysis has created value without deploying a single model.

Likewise, REWORK is not indecision.

It means the problem may be worth solving, but the evidence is not yet strong enough, the economics do not yet work or the implementation needs to change.

And GO should mean more than enthusiasm.

It should mean that the organisation understands what it is buying, why it is buying it, what evidence supports the decision, what it should cost, what can go wrong and how success will be measured.

That is considerably less exciting than an AI demonstration.

It is also considerably closer to management.

Why I wrote it

My professional career began in 1987, and much of it has involved the awkward territory between what technology is supposed to do and what happens when organisations actually try to use it.

The technology changes.

The management problems are remarkably persistent.

A solution still has to solve something worth solving.

Evidence still has to justify expenditure.

Costs still have to be counted properly.

People still have to use the system.

Failures still need owners.

And somebody still has to decide whether the proposal deserves a yes.

Artificial intelligence does not make those disciplines obsolete.

It makes them more important.

Because AI can be extraordinarily useful.

But enthusiasm is not due diligence.