What happens after discovery?

Discovery ends with a decomposed map of the work, not a deck. Workflows are broken into tasks, each with its inputs, judgment, context sources, and owner. From there I pick a small set of candidates that have a named owner, a success measure, and a cost ceiling, and everything else waits. At Honeycomb that filter took 23 designed experiments down to 4 live pilots.

How long does this take?

Understanding the work is measured in weeks, not quarters, and the first pilots run while discovery is still finishing in other functions. Changing the operating model is the slow part, because shared context, integrations, governance, and ownership are organizational changes rather than builds. Expect early evidence quickly and a changed operating model over a longer horizon.

How is autonomy introduced safely?

Nothing is trusted before it is measured. Each task sits on a ladder from Human Only to Human in the Loop to Human on the Loop to Autonomous, and it only moves up when evaluation supports the move: accuracy, precision and recall, false positives, resolution, containment, and exception rate. Some work stays with people on purpose. At Honeycomb 89% of contracts reconcile automatically end to end and the remaining 11% go to human review by design.

AI Transformation

What happens when I am hired to change how an organization works with AI.

AI transformation is a product problem at organizational scale. The model is rarely the blocker. Context, ownership, connected systems, evaluation, and trust are. So the work is to understand how the work runs, then rebuild it so AI can carry a real part of it.

The method

01
Understand the work
I sit with the work as it runs today, function by function, before anyone talks about tools.
36 discovery sessions across go-to-market, finance, marketing, and engineering.
02
Decompose it
Work gets broken into tasks with their inputs, judgment, context sources, and handoffs, so AI decisions happen at task level rather than by job title.
100+ workflows broken into 400+ tasks.
03
Find leverage
Candidates get an owner, a success measure, and a cost ceiling before anything is built. Most ideas do not survive that.
23 experiments designed, 4 moved into live pilots.
04
Build
I build production workflows and agents myself, which is how I know what the work actually needs instead of guessing at it.
89% of contracts reconciled automatically end to end at Honeycomb.
05
Evaluate
Every system gets measurement before it gets trust: accuracy, precision and recall, false positives, resolution, containment, and exception rate.
700K tracks a day at 99.36% accuracy and 99.75% precision at SoundCloud.
06
Set autonomy
Each task is placed on a ladder from Human Only to Human in the Loop to Human on the Loop to Autonomous, and it only moves up when the evaluation says it can.
The remaining 11% of contracts go to human review on purpose.
07
Change the operating model
The parts that make AI stick are organizational: shared context, real integrations, governance, literacy, and named ownership.
One operating model applied across 6 functions.
08
Measure what changed
The result is reported as changed work, not model output: what is faster, what is cheaper, what stopped being a person's job, and what deliberately stayed one.
Support for 100M+ monthly users rebuilt as a product.

Where it has been done

Rebuilding How Finance Work Gets Done →
Honeycomb. 89% of contracts reconciled automatically end to end.
AI-Generated Content Detection at Platform Scale →
SoundCloud. 700K tracks a day, with the decisions around the model as the product.
The Evidence Changed the Question →
Honeycomb. Where the discovery method came from and why the thesis moved.
Turning Consumer Support Into a Product →
The Weather Channel. Confidence thresholds deciding what AI handles alone.

All case studies →

Questions I get asked first

What happens after discovery?: Discovery ends with a decomposed map of the work, not a deck. Workflows are broken into tasks, each with its inputs, judgment, context sources, and owner. From there I pick a small set of candidates that have a named owner, a success measure, and a cost ceiling, and everything else waits. At Honeycomb that filter took 23 designed experiments down to 4 live pilots.
How long does this take?: Understanding the work is measured in weeks, not quarters, and the first pilots run while discovery is still finishing in other functions. Changing the operating model is the slow part, because shared context, integrations, governance, and ownership are organizational changes rather than builds. Expect early evidence quickly and a changed operating model over a longer horizon.
How is autonomy introduced safely?: Nothing is trusted before it is measured. Each task sits on a ladder from Human Only to Human in the Loop to Human on the Loop to Autonomous, and it only moves up when evaluation supports the move: accuracy, precision and recall, false positives, resolution, containment, and exception rate. Some work stays with people on purpose. At Honeycomb 89% of contracts reconcile automatically end to end and the remaining 11% go to human review by design.

Working together

Open to AI transformation and product executive roles. Available for select advisory engagements.

If you are hiring for AI transformation leadership, or you have one function where the work has stopped scaling, that is the conversation worth having.

Contact

Start with the work.

Tell me which function is stuck and what you have already tried. I will tell you what I would look at first.

Book a call →Email me →View résumé →Other ways to reach me →

The method

Understand the work

Decompose it

Find leverage

Build

Evaluate

Set autonomy

Change the operating model

Measure what changed

Where it has been done

Questions I get asked first

Working together

Start with the work.