Technology is not the first consideration in an AI strategy because it is the part that changes fastest and the part that least distinguishes you. Models, platforms and vendors are refreshed every quarter, and most of them are available to your competitors on the same terms. The customer problem you choose to solve, the data only you hold, and the judgment you apply to a model's output are what create commercial value. Settle those first and the technology choice becomes narrow, cheap and reversible. Lead with the technology and you end up with a capability nobody asked for and a bill that keeps arriving.
In short
- Start with the customer friction and the commercial value of removing it. If that value is unclear, no model will make it clear.
- Off-the-shelf models run on public data and return the same generic recommendations to everyone. Your first-party data is the advantage.
- AI delivers speed, not judgment. Define where computational scale stops and a person takes the decision.
- Test demand at low fidelity before funding a build. Building fast is no longer a moat.
What should you settle before choosing a model or vendor?
Four things, in this order. Each one narrows the next, and each one makes the eventual technology decision smaller than it looked at the start.
1. The customer problem and the commercial goal
Before looking at a single model, vendor or API wrapper, define the business objective. What specific friction are customers experiencing, and what measurable value does solving it unlock for your margins, retention or growth? If there is no clear commercial impact, no amount of AI will fix it.
Write the problem as a sentence that names the customer, the friction and the outcome. "New customers abandon setup because they cannot find the right configuration" is a problem. "We should use AI" is not. The first gives you a test, an owner and a measure. The second gives you a project.
- Name the segment experiencing the friction rather than all customers
- Estimate what removing it is worth in margin, retention or growth
- Say how you will know the friction has reduced
- Note what you would stop doing once it is solved
2. Your proprietary customer data
Off-the-shelf AI models run on public datasets. They hand you the exact same generic recommendations they hand your competitors. Your true moat lies in the unique first-party behavioural telemetry, usage logs and customer feedback sitting inside your own systems. Strategy starts with gathering and auditing the data only you hold.
That audit is usually uncomfortable and usually quick: what exists, who owns it, how current it is, and whether you are permitted to use it the way you intend. Where your own data is thin, that is the first thing to fix. Buying a model does not compensate for not knowing your customers.
- Behavioural telemetry that shows what customers actually do, not what surveys say they do
- Support and sales conversations that record friction, objections and workarounds
- Usage logs that show where customers stop, stall or leave
- Feedback and cancellation reasons, grouped by segment rather than in aggregate
3. Human interpretation and judgment
AI offers speed, not judgment. Models excel at processing datasets, finding surface patterns and accelerating discovery, but they regularly hallucinate, predict probabilities instead of reality, and lack commercial context. Data does not make decisions. People do.
Your strategy must define where computational scale stops and human strategic direction begins. That boundary is a written decision, not a culture. It states which outputs can act automatically, which must be reviewed by a named person, and who is accountable when a decision goes wrong.
- Outputs that can run automatically because the cost of error is low and reversible
- Outputs that require a named person to review before anything reaches a customer
- Outputs that inform a decision but never make it, such as pricing or proposition changes
- The evidence a reviewer needs to challenge the model rather than accept it
4. Rapid validation over a heavy build
Building fast is no longer a moat. Do not burn capital or engineering time deploying complex AI pipelines on unproven assumptions. Start with low-fidelity tests, prototypes or direct customer validation to de-risk demand before committing budget.
A conversation, a paper prototype or a waitlist page tells you whether the problem is real and whether anyone would change their behaviour for the fix. That is the information a build cannot give you cheaply, and it arrives in days rather than quarters.
- Talk to a handful of customers in the affected segment before writing a line of code
- Prototype the outcome the customer needs, not the infrastructure behind it
- Agree in advance what result justifies investment and what result stops it
- Keep the first test cheap enough that being wrong costs a week, not a quarter
The order matters more than the tools
The sequence is the strategy. Each stage answers a question the next one depends on, and the technology is only named once all four are settled.
| Order | Question to settle | What a good answer looks like |
|---|---|---|
| 1. Customer problem | What friction are customers experiencing, and what is it worth to remove it? | A named segment, a stated friction, and a value estimate tied to margin, retention or growth |
| 2. Proprietary data | What first-party evidence do we hold that no competitor can buy? | An audited inventory of telemetry, logs and feedback with owners, currency and usage rights |
| 3. Human judgment | Where must a person own the decision? | A written boundary between automated output, reviewed output and human-only decisions |
| 4. Rapid validation | How will we know demand is real before we build? | A low-fidelity test with a defined result that triggers investment or stops it |
Signs a technology-first AI strategy is going wrong
- A vendor or platform was chosen before the customer problem was written down
- The business case rests on efficiency savings nobody has measured
- Model output is passed straight to customers with no review boundary
- Engineering time is committed before a single customer conversation has taken place
- Success is described as deployment rather than a commercial result
What this looks like in practice
A problem-first AI strategy is usually smaller than the technology-first version. It names one segment, one friction, one value estimate and one test. It uses the cheapest tool that does the job, often a standard model pointed at your own data plus the judgment of people who already know your customers. What it produces is a decision you can defend and a result you can measure.
Where grapl fits
grapl works upstream of the tooling. We help leadership teams gather the signals they already hold, refine them into opportunities worth pursuing, align propositions and pricing around customer value, prioritise the moves that are validated rather than assumed, and leverage them into focused commercial plans. If you are writing an AI strategy and cannot yet name the customer friction it removes or the value that creates, that is the right place to start.