Case 04 · Selected Work
Three AI proposals, and the two I advised against
Three AI feasibility engagements in three unrelated domains, each turned into a solution design with a costed build estimate: two were correctly declined on capital cost, and the third was built.
- Role
- Independent AI Solutions Engineer
- Capabilities
- Discovery · Solution framing · Commercial judgement
- Domains
- Computer-vision monitoring · unmanned retail · automated website generation
- Outcome
- Two declined on capital cost, one built
Context
Six months working independently on applied AI, during which small and mid-sized businesses asked me to tell them whether an AI project they had in mind was worth doing. Two of them found me at a trade show. None of them had bought an AI system before, which is the normal situation and the one most proposals are written badly for.
Problem and ambiguity
In all three cases the customer described an outcome, not a system: run the shed without staff, run the shop without staff, stop paying for backend engineering on every website project. Nobody had established what was physically possible on their site, what it would cost, or what would have to be true for it to pay back.
My role
All three were mine end to end — the discovery, the solution design, the costing and the proposal. For the first two I scoped and costed; the deliverable was a solution design and a concept walkthrough, not a running system, and neither went ahead. The third I built.
Discovery
For the livestock-monitoring engagement I never visited the site. I worked from footage of their existing operation, which turned out to be enough — and better than a visit for the specific question of what a camera can actually see there. For the unmanned-retail engagement, the discovery was mostly about what happens when the system is wrong, not when it is right. For the third, the constraint was commercial: they had explicitly refused a template-CMS approach and wanted natural-language input producing real sites they could build or resell.
Architecture and trade-offs
The decision I am proudest of in this period is not a model choice. The livestock system had to count animals per zone on a schedule and flag when a person entered frame, under two hard constraints: the on-site machine could not stay online continuously, and detection still had to be timely. Instead of reaching for a bigger model to count a whole flock, I advised changing the physical setup — camera grade, and above all camera placement, sited at the movement choke points the animals have to pass. That turns an intractable vision problem into a tractable one. In enterprise AI the cheapest lever is often not in the software.
Implementation and integration
The third engagement was built: a multi-tenant website generation platform taking natural-language input and producing single- or multi-page sites, removing per-project backend work from the agency’s delivery model.
Production controls
The platform is in closed beta with a small number of tenants — internal testing, not general availability. I do not describe it as launched or in production, because it is not.
Impact
Two costed proposals, two clients who declined on capital cost, and one platform built and paid for. The valuable outcome here is the judgement, not a win rate: being able to tell a customer early and with a real number that their idea does not pay back is worth more to them than a proposal that survives to a contract and dies in delivery.
Reusable learning
An AI feasibility assessment is a commercial artefact wearing a technical costume. The technical answer is usually “yes, this is possible”; the useful answer is what it costs, what has to change physically, and at what volume it stops being cheaper to keep doing it by hand.
Evidence pending
The method could be published as a scoping questionnaire, a costed-estimate template and one worked example on synthetic data — project P7. The third engagement’s deliverable is paid client work and is not mine to republish.