Advisory Services / AI-Assisted Product Development
I use coding assistants every day to build Lorenz and this site. I also brought AI-assisted development tools into a platform modernization across a global technology organization. In both cases, the model bill was the smallest cost. Review time, discarded attempts and rework cost more, and they are rarely tracked.
The direct answer
Teams make AI-assisted development reliable by starting from product intent and acceptance criteria, putting repeated work into versioned instructions, scripts and skills, checking outputs independently and tracking model spend alongside review effort and rework.
A person stays accountable for every decision the model proposes.
Interactive example
One fictional feature built with a coding assistant. Move the slider to add the costs that usually go untracked.
This is the number finance usually sees.
After two discarded attempts on one task, stop. Rewrite the task and its acceptance check, then run it again.
Illustrative. Token cost at September 2026 rates. Hours are typical of a mid-size team’s first quarter with coding assistants, not a measurement.
When to address it
Code arrives fast, and reviewers spend the next three days working out what it was meant to do.
Useful context lives in one person’s chat history instead of shared skills, scripts and tests.
Model spend is visible. Review hours, discarded attempts and rework are not.
The working model
The useful level of abstraction is the smallest durable practice that makes the next piece of work easier to direct, check and improve.
Every task connects to a product goal, a decision owner, its constraints and explicit acceptance criteria before an assistant touches it.
Recurring patterns become reviewed instructions, skills, scripts and tool permissions, so nobody depends on a perfect prompt.
Tests, evaluations, design review and human acceptance verify behavior. Model output is a proposal until it passes.
What worked goes back into the shared practice, and model spend is measured together with review and rework.
Illustrative example
Task. A team asks an assistant for a customer-facing billing rule. The code is fast, clean and passes its tests.
Review. The reviewer spots the problem: the rule optimizes last quarter’s request, not the current pricing goal. Two runs are discarded.
Template. The task template now carries the goal, the approved scope, a design constraint and the acceptance check. The third run lands.
Reuse. The review finding is now part of the shared skill, so the same mistake is caught earlier next time.
A composite of common situations. Names and figures are illustrative.
What you leave with
Tools come after the work is clear. Claude Code, Cursor or whatever your team already trusts can fit, as long as the output is reviewable.
Lorenz, our planning software, prices every backlog item in tokens and dollars before the work starts, so the cost view begins at the plan.
Common starting points
How an engagement starts
Which workflow is fast but not better, and what reviewers keep catching.
A Two-Day Intensive ($2,500) takes one workflow end to end with review criteria. A Sprint Advisory ($6,500) adds the skills, checks and cost view for a team.
As a Fractional CPO ($8,000 a month) I keep the feedback loop running across teams until your own owners run it.
Two-Day Intensive and Sprint Advisory fees are credited toward the first month of Fractional CPO. Onsite in the New York City area, remote everywhere else. See pricing. Specialists join only when the work needs them: meet the bench.
Common questions
No. Tool selection starts with the work and constraints. Existing tools such as Claude Code or Cursor can fit into the operating model when they are useful and reviewable.
Vibe coding accepts whatever the model produces if it seems to work. AI-assisted development starts from product intent and acceptance criteria, checks the output independently and feeds what reviewers learn back into shared skills. Vibe coding can work for small personal projects. Teams shipping production software need the second approach.
Compare accepted output against the full cost: model spend plus review time, discarded attempts and rework. Track it per workflow, not per seat, and watch whether the ratio falls as the shared skills improve.
Decisions. Agents can draft, implement, test and summarize, but choosing what to build, accepting a tradeoff and signing off on quality stay with a named person.
Yes. Figma and AI-assisted design workflows can carry design intent, component rules and review criteria into prototypes. A person still judges usability, coherence and product fit.
Define task boundaries, usage visibility and stopping rules. Include the time spent reviewing and repairing outputs when comparing approaches, not only the model bill.
Start with a real workflow
It might be a review queue that keeps growing, a prompt only one person trusts or a model bill nobody can explain. We can start there.
Book a discovery call