Service

AI Pilot to Production

One workflow, end to end: designed with your team, measured against the old way, and rolled out so it sticks.

Book a 30-minute scoping callTake the 2-minute scorecard

No pitch. If I can’t find a workflow worth automating, I’ll tell you on the call.

Who this is for

Operations and support leaders who already know the target, or who have a pilot that stalled after the demo. If a vendor tool is live for six people and nobody trusts the output, this is the engagement that fixes it.

How the engagement runs

Weeks one and two establish the baseline: current volume, handle time, error rate, and cost for the workflow, plus the edge cases the team already knows about. Weeks three to six build the automation, with human review on every output until the numbers earn trust. The remaining weeks are rollout: training, documentation, escalation paths, and a feedback channel the team actually uses.

  • Working automation in your stack, with human review where it matters
  • Vendor selection or a build, with the tradeoffs documented
  • QA, escalation paths, and a rollout plan the team helped shape
  • Training sessions and a playbook written for the people doing the work
  • Before-and-after numbers leadership can trust

Why rollouts fail, and what I do differently

Most AI pilots do not fail on the technology. They fail because nobody measured the baseline, the team was handed a tool instead of a changed process, and the first wrong answer had no path to a human. I started my career as a teacher, and I treat adoption as a teaching problem: clear expectations, practice, and a way to raise your hand.

Questions

What people ask first.

What kinds of workflows are good candidates?

High-volume, well-defined work with a clear right answer most of the time: ticket triage and routing, verification and data entry, document review, first-response drafting, QA sampling, and internal request handling.

Will you work with a vendor we have already bought?

Yes. If the tool fits the workflow I will make it work. If it does not, I will show you the numbers and we decide together.

How do you measure success?

Against the baseline we set in week one: volume handled, cycle time, error rate, cost per case, and adoption by the team. The readout shows what was automated, what it cost, what it saved, and where it was wrong.

Which AI tools and vendors do you work with?

I am vendor neutral. I have bought and implemented platforms such as Zendesk, Twilio Flex, and 8x8 at Lyft scale, and I build with Claude, OpenAI, and current agent tooling every week. The recommendation follows the workflow, not the other way around.

Do we need engineers for this to work?

Usually not for the first wins. Most operations automation comes from well-defined workflows and off-the-shelf tooling configured carefully. When a build is needed, I write the first version or a spec your engineers can act on.

We are in a regulated industry. Is that a problem?

It is the normal case for me. I run operations under FCRA today, and at Lyft I worked through regulatory programs that required careful tradeoffs. Compliance is designed into the workflow with audit trails and human review, not bolted on afterward.

Further reading
AI & Operations

Eliminate, Automate, Too Late: How to Sort Support Work Before You Automate It

The EAT framework (Eliminate, Automate, Too Late) for deciding which customer support work to remove, which to automate with AI, and which to leave to people, plus how to classify customer pain at scale with an LLM.

AI & Operations

Building a Lead Signal Engine with LLM Classification

A working case study: designing, prototyping, and iterating an AI classification system that surfaces high-signal opportunities, including what broke and what finally made it useful.

Let’s find the first workflow.

Thirty minutes. Bring the process that eats the most time. You leave with a view on whether AI can help and what it would take.

Book a 30-minute scoping call