Small experiments.Real momentum.

A working lab for practical AI. Pick one manual step a team hates, ship it in days, and measure the win.

Start with the test See the Win Card

Just have it do a thing.

Three moves that beat the big AI roadmap.

  1. Ship narrow

    Pick one manual step one team hates. Build it in days, not months or quarters.

  2. Define "working" first

    Write the baseline, the target, and the owner before anyone writes a prompt.

  3. Spend the win

    Use the measured result to pick, prioritize, and fund the next thing.

Scope for finishing, not for applause.

The project that sounds best in a steering committee is rarely the one you can finish. Cut it down to one step inside a process.

Impressive enough to stall

  • Automate decisions end to end
  • An AI-powered collections platform
  • Replace the shared inbox
  • Enterprise data intelligence

Small enough to ship

  • Draft the memo narrative from fields you already capture
  • A one-page call prep brief for tomorrow's queue
  • Tag and route incoming email, draft the reply for review
  • Normalize messy descriptions into clean, structured fields

The Thing Test

Six yes answers keep a first project small enough to ship. Tap each one that is true for your idea.

0of 6

Tap every answer that is true.

The loop is the strategy.

Design thinking with a scoreboard. No roadmap up front: empathize, define one step, prototype fast, test against a number, then go again.

EmpathizeDefineIdeatePrototypeTestIterate

1 2 3 4 Harvest ideas Empathize and define

  1. Harvest ideas

    EmpathizeDefine

    Sit with the people doing the work and watch the real process. Name one narrow, manual step and the person who owns it.

    Ideas come from the people doing the work.

  2. Build and demo

    IdeatePrototype

    Sketch a few ways AI could take the step, pick the simplest, and build it in days. Show it live and let the team react to something real.

    Real users clicking in week one.

  3. Win

    Test

    Test against the baseline you wrote first. Hit the number on the Win Card, or stop and keep the lesson.

    Success is a number, not a launch.

  4. Fund the next

    Iterate

    Feed what you learned into the next idea. Bring the result, the next Win Card, and the ask.

    Results, not a roadmap, pick the next build.

Define working before you build.

The Win Card

Process step
Call prep
Owner
Team lead
Baseline
22 minutes of prep per account
Target
Cut prep time in half
Measure
Timed sample, 30 days
Exit criteria
Under 25% saved, or people stop using it

A vibe is not a result.

A prototype that "seems to be helping" is a vibe, not a result. Write the Win Card first, then pick one or two metrics.

Adoption is the honest metric.

If people stop using it, the other numbers do not matter.

Translate the win into money.

Finance wants hours returned times what those hours cost. Set the frequency and minutes to price your own Win Card.

Counts working days only, about 21 a month.

Keep a human in the loop and count the review time.

Measured from that app's own API key. Small document tasks often land in cents.

Hours returned each month

0hrs

Net value each month

$0

Person, per run

$0

AI plus review, per run

$0

Defaults: a $100,000 salary, 10 times a day, 22 minutes by hand, and 5 minutes with AI including review. Even at cost parity, the hours returned go to higher-judgment work.

Things learned the useful way

Accuracy is an input problem.

Gold in, gold out. Clean data and a clear prompt decide the quality of the answer long before the model does.

Say the quiet part out loud.

Tell people the goal is augmentation, not replacement. Then fund the training that proves you meant it.

Eighty percent can be done.

Ship the useful version. Polish only when the numbers say it is worth it.

Tech debt, eyes wide open.

Move fast, write it down, and refactor into standards once a pattern earns it.

New process, new app.

Think beads on a bracelet. When the next step belongs to a different process, it becomes a different tool.

On the bench

Personal experiments that test the playbook outside the office. Some will ship, some will be retired, and each one leaves a lesson.

  • Self-hosted meeting transcription

    Running

    Open-source transcription on an always-free cloud server turns recordings into transcripts, summaries, and action items, with no monthly subscription.

  • Secure agent integrations

    Building

    Zero trust by default: every agent sits behind identity-aware access and a private tunnel, with scoped API keys per app and nothing exposed to the open internet.

  • Agent usage and cost dashboard

    Building

    One key per app and one private view of usage, uptime, and cost per run, so every experiment has a number attached.

  • Open agent experiments

    Exploring

    Testing open-source agents across hosted and local models to learn where each one earns its keep.

Questions, answered

The concepts on this page, in short answers.

Hands-on workshops AI fluency workshops and advisory for leadership teams. Visit kimbark.org
How do you pick a first AI project?

Scope for finishing, not for applause. Run the Thing Test: one team and owner, one manual step, clear in and out, days to test, a human in the loop, and no core rewrite.

What does "it's working" look like in numbers?

Write a baseline, a target, an owner, and exit criteria before you build. Track one primary metric and one guardrail, usually time saved and adoption.

Where does design thinking fit?

It is the loop. Empathize and define while harvesting ideas, ideate and prototype while building, test against the Win Card, and iterate by funding the next step.

How do you turn one win into momentum?

Demo the tool live. When people see a colleague use something real, they bring their own narrow pain point. That becomes the next Win Card, and the last result funds it.

How should teams address the fear of job loss?

Directly. Say the goal is to augment people, fund training for new skills, and make subject matter experts part of the build. Their expertise is the value.

How do you know when an app is done?

When the next step belongs to a different business process. That step becomes its own app instead of scope creep in this one.

What about junior staff who learned on these tasks?

It is a real tradeoff. Redesign junior roles and training so people still build judgment, even when the tedious version of the work is automated.

Can my team get a hands-on workshop?

Yes. Workshops are available through Kimbark.

Built by a practitioner.

Steven Siler is a technology executive with 20 years in financial services, leading data, product, engineering, and IT teams through growth and multiple M&A integrations.

A former college swimmer, he still thinks in laps: steady rhythm, small gains, repeated until they compound. siler.ai is where those laps happen.

Connect on LinkedIn

Private apps

Sign in with an approved account to open the app launcher.