Practical AI product strategy: five questions before you build

The short version

• Start with the job to be done, not the model's capability.
• Check that the data, the process and the people are ready before you build.
• Ship the smallest coherent product that can earn trust, and say "not yet" when the conditions aren't there.

Most AI conversations start with capability: what the latest model can do. Useful product conversations start somewhere else, with the job, the workflow that holds it, and the organizational conditions needed for adoption. Having led AI strategy at Al Jazeera, and overseen the integration of AI and ML into the product at Maqsam before that, these are the five questions I ask before any AI idea earns a place on the roadmap.

1. What decision, task or workflow becomes meaningfully better?

Name the job. Who does it today, how long does it take, and what does it cost when it goes wrong? Then define what "better" means: faster, more accurate, less risky, or something people couldn't do before.

If you can't say it in one sentence, you don't have a use case yet; you have a technology looking for a problem.

2. Is the data and the process ready?

AI is only as useful as the data it can reach and the workflow it can slot into. I look at both:

• Data readiness: Is the data accessible, reasonably clean, permitted for this use, and current enough to matter?
• Process readiness: Does a stable workflow exist? AI can't fix a process nobody agrees on.

When the data isn't ready, the first product is often the work that makes it ready. That's less exciting than a demo, and much more valuable.

3. Where does human judgment sit?

Decide, deliberately, where a person reviews, approves or overrides. The cost of an error should set the design: when it is low, automate more; when it is high, let the model draft and the human decide.

Make the review step part of the interface. If it relies on goodwill, it will disappear under deadline pressure.

4. Build, buy, or assemble?

There is rarely one right answer. Hosted models, open-source stacks and vendor products trade off control, cost, privacy, speed and lock-in differently. I've worked with both LLMs and open-source model stacks, and the practical rule I follow is to decide with evidence: build a small evaluation set and compare the options on your own tasks, not on someone else's benchmark.

5. What is the smallest coherent product that can earn trust?

Pick one workflow. Define how you will evaluate it. Put a human review loop around it, choose one metric, and ship that. Trust is the scarce resource in AI products; you earn it with something small that works reliably, then expand.

Delivering AI prototypes into newsroom workflows is a good example of the target: less manual work and better publishing efficiency inside the workflow people already follow, not a demo that impresses a room.

Knowing when to wait

Part of strategy is saying "not yet." I'd hold an AI idea when:

• nobody owns the workflow it would change,
• the data isn't accessible or permitted,
• the cost of an error is too high for the current level of reliability, or
• the problem isn't painful enough to justify the change.

Novelty is easy to demo. Value is harder to operationalize. The work of product leadership is bridging that gap with clarity, sequencing and restraint.