Build with forethought, not just AI-everywhere

Everyone worries their AI pilot will fail. The bigger surprise is what happens if it succeeds.

Will the model work? Will people use it? Will it scale? Will it ever make it into production? Reasonable questions. Also the easy ones. The harder question, the one that gets skipped because no one wants to spend six months over-analysing a project that might never ship, is what happens after the pilot starts to fly.

Think of an AI business case as an iceberg. The costs that get pilots approved sit above the waterline. The ownership commitment cost — what shows up if the pilot works — sits below it, and it is much larger than the bit you can see.

What gets approved

Most AI business cases focus on the visible costs:

• Build cost
• Software licences
• Pilot compute
• Vendor fees and initial integrations

Easy to estimate. Easy to defend. A fraction of what AI actually costs to own.

What ownership actually costs

If the pilot works, you have an AI ownership commitment cost. That commitment carries a set of costs most business cases barely touch.

Data preparation and integration. Often the largest part of any AI build, and rarely scoped properly up front. Clean inputs are not a one-time job; they are a permanent operating cost.

Production API costs. Usage at scale moves token spend by an order of magnitude. The pilot ran on a small sample. Production runs on everyone.

Retraining, evaluations and drift monitoring. Models do not stay useful by accident. They need someone measuring them and updating them on a schedule.

Human-in-the-loop oversight. Quality checks, exception handling, and escalation paths are recurring labour costs, not edge cases. They are the cost of taking responsibility for what the model says.

Compliance and audit trail. Model registers, risk reviews, and policy controls need real ownership. Auditors will ask. Customers and regulators will ask.

Change management and training. Adoption does not happen because a tool is available. It happens because someone made it happen, repeatedly, across teams.

Vendor lock-in and exit planning. The switching cost is invisible until you need to switch. By then it is also enormous.

Forethought beats AI-everywhere

The default playbook right now is to put AI into everything and see what sticks. It is a bad playbook. Every place AI sticks is a place you have just signed up for an ownership commitment cost — usually without scoping it, often without budgeting it, and almost always without deciding who owns it.

Forethought is the alternative. Before you spend a pound on a pilot, do the work to understand what owning the result actually looks like. Map the data preparation. Estimate production usage, not pilot usage. Name the human in the loop. Decide who owns drift, evaluations and exit.

Then ask one question.

If this works, is it still worth owning?

If the answer is yes, build it. If the answer is no, design something else — or don't build it at all. The worst outcome is discovering the true cost after the pilot has caught on, the users love it, and turning it off is the most painful option on the table.

AI everywhere is a fashion. Building with forethought is a strategy.

If you're sizing up an AI pilot and want a second opinion on the ownership cost before you commit, get in touch.