Service
AI Product Development
Design and delivery of AI-enabled product features that are tested against real data, costed before they are built, and shipped with human review in the loop.
- Who it is for
- Founders and business leaders who have identified a language- or document-heavy workflow and need to know whether AI can carry it reliably.
- Engagement shape
- Fixed-scope evaluation, then phased build
- Typical duration
- 2 to 4 weeks for evaluation, 6 to 12 weeks for the first production feature
- Starts with
- A one-hour session on the workflow and the data available
Problems this addresses
Situations that lead here
If one of these is a statement you could make about your own business, this is the relevant service.
- “We are exploring AI but do not know which use case to start with.”
- “A demo worked well but we cannot tell whether it holds up on our real data.”
- “We need to understand running cost and failure modes before committing.”
- “Our team needs review and audit steps around any automated decision.”
What it includes
The work, in order
Each part produces something reviewable rather than ending in a document nobody reads.
Use case selection
Candidate workflows are ranked by measurable staff hours consumed, tolerance for error, and how cleanly the result can be checked. Work starts on the highest-value case that can be verified.
Evaluation against real data
The smallest useful version is tested on your own documents and records, with accuracy, latency and cost recorded per run rather than estimated.
Human review design
Where output affects a customer, a payment or a compliance record, the review step, escalation path and audit trail are designed alongside the feature.
Production delivery
Model calls, prompts, retries, cost controls and logging are built into the product rather than kept in a separate prototype.
What you keep
Deliverables
Everything below remains yours and is usable by another team if the engagement ends.
- Ranked use case shortlist with expected effort and value
- Evaluation results on your data, including failure cases
- Running cost model per transaction and per month
- Working feature in production with review and audit steps
- Documentation of prompts, guardrails and known limits
Questions
Answered directly
How do you decide whether AI is the right tool?
By checking three things: the task is repetitive and language- or document-driven, the correct answer can be verified by a person or a rule, and the current manual effort is large enough to justify the build. If any of the three fails, a conventional workflow is usually cheaper and more reliable.
What happens if the evaluation shows AI is not viable?
You get the evidence and stop. That is a successful outcome — the alternative is discovering the same thing after a full build.
Do you build on a specific model or provider?
No. Provider choice follows the evaluation, and the integration is kept behind a boundary so a model can be replaced without rewriting the product.
Considering ai product development?
A one-hour session on the workflow and the data available. Send the business context and the constraint behind it, and I will reply with a direct read on the approach I would take.