AI evaluation

Find what slows delivery before buying another AI tool.

I evaluate the path from a useful AI output to a shipped decision: who owns it, where it stalls, what requires human judgment, and what proof is missing.

See the scorecard

Evaluation lens

Cost

Find the handoffs and rework you keep paying for.

Time

Name the longest gap between a decision and production.

Risk

Put review and escalation where consequences change.

Headcount

Keep judgment with people. Automate repeatable work.

The read

Boring evals stop at a grade. Useful ones produce a decision.

You leave with one constraint, one first move, the person who owns it, and the proof that will show whether it worked.

Staged scorecard

Leave with a decision your team can run.

Four common failure modes, rewritten as clear ownership, reusable systems, production proof, and review gates.

Area	What stalls	What good looks like
Strategy	A broad AI roadmap with no named owner	One valuable problem, owned end to end
System	Prompts and process scattered across teams	One reusable workflow with review built in
Shipping	Demos accumulate while production stays the same	A working change ships behind your existing checks
Governance	Every output needs another manual inspection	Review and escalation match the actual risk

The method

Three steps to one clear decision.

Map the drag

Trace where time, money, risk, and attention leave the workflow.

Find the constraint

Separate a tooling problem from an ownership, process, or review problem.

Define the first proof

Name what should change first, who owns it, and how you will know it worked.

The deliverables

Four artifacts turn the diagnosis into action.

If the answer is not a retainer, I will say so. If the first move is deletion, I will name it.

A decision record naming the one problem worth solving.

A risk register tied to review gates.

A near-term action map with first proof defined.

A build-versus-buy read on tooling, process, or delivery.

Start with the constraint