Using AI prototypes to discover when automation is the wrong answer
Two AI prototypes looked promising in isolation. Testing them inside real workflows revealed a different question: could the workflow absorb the intervention without degrading judgment, attention, or trust?
Live testing revealed that the challenge was not whether AI could provide useful guidance. It was whether the workflow could absorb that guidance at the moment it appeared.
Can AI do this?
Can the workflow absorb it?
The prototypes worked. The workflows could not absorb them.
That distinction changed the product strategy from validating model output to diagnosing timing, criteria, trust, and attention costs.
AI creates pressure to build before the workflow is understood
AI concepts can appear viable early because the technology performs the task in isolation. But model capability does not prove that an intervention belongs in the workflow.
I used two lightweight LLM assistants as diagnostic instruments: Anne-bot in enterprise PRD coaching and Capture Caddy in a live creative workflow. Rather than testing only whether the assistants could generate useful output, I tested what happened when people tried to use them inside real work.
Build prototype
Build a lightweight LLM assistant to make the concept testable with users.
Observe use
Watch how people incorporate the assistant into real work and where interaction creates friction.
Reveal constraint
Identify whether the blocker is timing, criteria, attention, trust, or ownership.
Make a product decision
Decide whether to proceed, narrow scope, change the interaction, pause, or stop.
AI could extend coaching, but not resolve unclear approval criteria
The hypothesis was that limited coaching access was slowing PRD approval. Ford’s Digital Cabin PRD reviews were rigorous and Socratic, often stretching across several rounds. The early opportunity appeared to be preparation: could an assistant help PMs clarify their thinking before review?
The prototype tested whether AI could reproduce a director’s coaching patterns and guide PMs through stronger problem framing and clarification.
AI could surface pose plans, but not protect live-session attention
The hypothesis was that hands-free access to a prepared pose plan could help photographers stay oriented during live sessions. Photographers often plan poses, prompts, or shot sequences before a session, but accessing that plan in the moment can mean checking printed notes or pulling out a phone while still trying to direct and connect with the subject.
I tested whether voice commands through earbuds and a phone could make that plan accessible without requiring the photographer to stop, search, or visibly shift attention away from the subject.
Prototype progression. Desk testing showed that AI could generate useful direction. Live testing revealed that the photographer had to manage the subject, camera, environment, and assistant at the same time.
The prototypes revealed what each workflow needed protected
Anne-bot and Capture Caddy were not just tests of whether AI could help. Each prototype entered the workflow as support, then revealed a hidden condition the product needed to preserve: shared evaluation criteria in the PRD process, and presence, timing, and subject connection in the photography process.
The pattern was not bad AI output. It was that useful output was not enough when the surrounding workflow depended on conditions the assistant could not protect.
The real constraint was not coaching support.
It was shared criteria for what “approved” meant.
Anne-bot helped PMs clarify the problem before review, but approval still depended on a common understanding of what made the problem statement strong. Research revealed that problem-statement alignment made asynchronous director approval possible.
Shared evaluation criteria
The better-fit role was practice, not live-session assistance.
The assistant helped when the goal was to rehearse posing language before a live session. In a live shoot, the same support became disruptive because every delay or missed command split attention away from the subject.
Presence, timing, and subject connection
Capability wasn’t the constraint. Workflow fit was.
The prototypes did not fail because AI was incapable. They succeeded enough to reveal that capability was the wrong test. By the time both prototypes were tested in context, the product question had changed: not whether the assistants could produce useful coaching or direction, but whether the work could tolerate an AI intervention at that point in the process.
Anne-bot exposed a criteria problem: coaching could scale, but agreement could not be automated before reviewers shared what “good” meant. Capture Caddy exposed an attention problem: the assistant could provide useful guidance, but live photography required presence that the interaction interrupted.
That shifted the product decision from building around AI capability to deciding where AI belonged, what had to be clarified first, and what parts of the workflow should remain protected from automation.
Evaluate whether AI can produce useful output.
Evaluate whether the workflow can absorb the intervention.
Workflow fit became the real test
I used to ask whether AI could perform the task. Now I ask whether the workflow can absorb the intervention.
Both prototypes looked promising in isolation. The more important evidence appeared when each assistant entered the workflow.
Anne-bot exposed criteria that still needed to be aligned before coaching could scale. Capture Caddy exposed an attention cost the live session could not absorb.
That changed how I evaluate early AI product ideas. A working assistant is not the same as a viable product. The prototype also needs to reveal what should stay human, what must be clarified before automation, and where assistance can enter without weakening the work.