David Klein
For Fractional AI

I turn ambiguous
business problems into
solutions that ship.

That has been the through-line of my career: from a fintech founder and CEO, to McKinsey and American Express before that, and now leading hands-on AI transformation for companies of all sizes.

View résumé
The match

The six things you are looking for

I’ve done it.

Customer-facing, high-stakes, ambiguous You lead when stakeholders are not aligned and the path is not obvious.
High stakes, low certainty: steering a lender through the COVID crash without tripping a covenant, earning and keeping the world’s most demanding enterprises as partners (IBM, BlackRock, Adobe, Humana), and putting products into the hands of a million consumers and nearly one in five U.S. high-school seniors.
Real product and delivery You have shipped with engineers and designers across many cycles.
A decade shipping with engineers: built and launched products alongside engineering and design teams, and ran the day-to-day on an Agile cadence – stand-ups, sprint planning, retros, and clear delivery and value KPIs – without turning it into process theater.
Structured problem-solving and communication You turn chaos into a plan, then a crisp doc for engineers.
McKinsey-trained, founder-tested: my superpower is turning the complex into the simple and the meandering into the concise and clear, and I am maniacal about documentation – doing it clearly, concisely, and comprehensively.
Bias for action and ownership You roll up your sleeves and make things happen.
A founder’s reflex: built a national fintech from scratch; stood up a new solar-financing business in eight months and grew it to a top-five U.S. lender; removed a third of the cost structure in weeks when COVID hit.
Comfort with ambiguity and intensity You like messy, “we have not done this before” problems.
Ambiguity is where I thrive: built a company that scaled in a space that didn’t exist before; got investors, employees, and partners on board; led a 23-company coalition that changed U.S. federal law.
AI curiosity, not AI dogma You want to overuse AI in your own workflow.
AI-first operator: I live, breathe, and work in AI every day. I build with frontier models, APIs, connectors, agents – to produce client-ready work.
Case Study: A real engagement

From a fuzzy “what about AI?” to a new AI-infused operating model.

This comes from some of my recent work. Client details are kept anonymous.

Software & services · AI-era transformation

The company’s core business was being disrupted by AI. They did not know what to do about it. They brought me in.

The situation
A software services firm with a few hundred clients watched AI start to erode the work it sold. The brief was wide open: figure out what AI means for us, and what we should do.
What I did
Embedded with the CEO and leadership team. Mapped where AI created advantage versus exposure, and turned the ambiguity into a three-part plan that included how to: fix the non-AI growth engine, use AI to cut the cost of delivery, and use AI to grow the value delivered to clients.
The Organizational Unlock
Designed a repeatable “innovation loop”: an operating model that turns delivery data and frontline sales signal into new AI-enabled offerings the team can design, test, and ship – continuously.
The hidden problem
Re-cut the client’s CRM data across thousands of implementations and found their real bottleneck was not win rate but lead-to-opportunity conversion – reframing the number-one growth problem and redirecting where they invest.
Forward-deployed teardown

Pick a problem. Here is how I would ship it.

A few typical client problems. For each, how I would approach it.

Make content moderation faster and more accurate

“Our review queue is slow and inconsistent. Reviewers burn out, and bad content still slips through.”

The real problem
Improve precision and recall on the highest-risk categories while cutting human review time – without raising false removals that punish good users.
Approach
Start with the policy, not the model. Turn it into a labeled rubric and a golden set from real cases. Use an LLM to triage: auto-clear the obvious, escalate the ambiguous to humans, with confidence thresholds. Tighten with evals each cycle.
How we measure success
Share of cases auto-resolved at target precision, reviewer time per case, false-removal rate, and time-to-decision on high-risk content.
Why it works
Evals turn a black box into something you can trust and improve – the same discipline behind your Zapier hallucination work.
Path to production
0–30 days
Policy to rubric; build the golden set and baseline evals.
30–60 days
Run in shadow mode against the live queue; tune thresholds.
60–90 days
Phase in on lowest-risk categories first; expand as scores hold.

Make customer service better and cheaper

“Support is expensive and slow. We want AI to handle more without making the experience worse.”

The real problem
Resolve and accelerate the contacts that can be resolved, at quality – and route the rest to humans with better context. Measured by resolution and satisfaction, not deflection alone.
Approach
Mine ticket history for the top resolvable intents. Build a retrieval-grounded assistant that answers from your own help content and policies, with guardrails and a clean, summarized handoff to agents. Prove it against real tickets before it goes live.
How we measure success
First-contact resolution, average handle time, CSAT, and cost per contact – with containment that does not dent CSAT.
De-risk first
Launch internally as agent-assist (suggested replies) before anything is customer-facing, so quality is proven where it is safe.
Path to production
0–30 days
Intent analysis; assemble the grounding corpus and eval set.
30–60 days
Agent-assist live internally; measure quality and lift.
60–90 days
Customer-facing on top intents; widen coverage by eval score.

Build third-party integrations far faster

“Every new integration eats weeks of engineering time. We want to ship them in a fraction of that.”

The real problem
Turn the manual, doc-reading, trial-and-error connector build into an AI-assisted pipeline that produces a working, tested connector quickly – with engineers verifying, not hand-coding from scratch.
Approach
Crawl the partner’s API docs; extract auth, endpoints and schemas; generate a connector scaffold; then test it against real calls and let it fix itself against a test harness. An engineer reviews and ships.
How we measure success
Build time per connector, share of generated connectors that pass tests unaided, engineer hours saved, and defect rate in production.
Why it works
It mirrors the connector-builder pattern you proved with Airbyte – integrations in minutes, not hours.
Path to production
0–30 days
Pick three representative APIs; build docs-to-scaffold and the test harness.
30–60 days
Close the generate-test-fix loop; measure the pass rate.
60–90 days
Roll across the integration backlog; track time-to-ship.

Turn a decades-old dataset into a domain agent

“We have a deep, proprietary dataset. Can AI turn it into an agent that does the expert work?”

The real problem
Encode the expert workflow and the data into an agent whose outputs an expert would sign off on – where correctness matters and mistakes are costly.
Approach
Sit with the experts and map the real decision workflow. Use the dataset for grounding and as an eval set of known-correct cases. Build the agent to show its work, so experts can audit, with human sign-off where the stakes are high.
How we measure success
Accuracy against expert ground truth, share of cases needing human correction, expert time saved, and auditability of every output.
Why it works
Grounding plus a real eval set is how a high-stakes agent – a medical coder, say – earns expert trust instead of guessing.
Path to production
0–30 days
Map the workflow; build a golden eval set from the dataset.
30–60 days
First agent on a narrow, high-value slice; measure against experts.
60–90 days
Expand coverage as accuracy clears the bar for sign-off.

Illustrative approaches, written to show how I scope and ship – not finished projects.

Why this work, why now

The work I love, at the moment it matters most.

This work sits at the cutting edge of technology and hard business problems, and that intersection is where I’ve operated my whole career. I love the work. It’s what I am trained for and good at. Above all, it energizes me, and that is the metric I care about most.

AI has hit its inflection point. It is now very much a real tool that real companies can use to ship real solutions. I see it every day in my own work, and the chance to do it inside a growing firm, at this exact moment, is exciting to me.

I am based in the Triangle, near your Raleigh-Durham office.

Prepared independently by David Klein. This is a personal candidacy page and is not affiliated with, endorsed by, or sponsored by Fractional AI.

© 2026 David Klein · Durham, NC