AI Workflow Prototyping: Test Your Idea in 30 Days.

AI workflow prototyping means building a small, rough version of an AI system and running it on real work. You do this before you spend big. You learn fast what works, what breaks and what is worth building properly.

Think of it like a test kitchen. You cook one dish for a few people before you print the menu.

Here is how to do it. It is for founders, educators and consultants who want the work done without being in the middle of everything.

What an AI workflow prototype is (and is not)

A workflow is a set of steps that turns an input into an output. A client question comes in. A reply goes out. A call recording comes in. A lesson summary goes out.

An AI workflow prototype does those steps with AI. It is rough on purpose.

It is not:

• A pretty app demo.
• A slide deck of ideas.
• A full launch.

You may see search results for "AI prototyping" that mean app screens and design mock-ups. Those tools help product teams sketch interfaces. This guide is about something else. It is about prototyping the work itself: the decisions, the steps and the checks.

If you run a consulting or education business, this is the one that matters. Your value sits in how you think. A screen mock-up does not test that.

Why prototype first

Big AI projects fail in a quiet way. They look fine in a demo. Then they fall over on real work.

The numbers back this up. McKinsey's 2025 State of AI report found that 62% of organisations are experimenting with AI agents. In any single business function, no more than 10% say they are scaling them. The main barrier is not the tech. It is that people did not redesign the work around it.

A prototype fixes that early. You find the gaps while they are cheap.

A prototype also helps you answer three plain questions:

1. Can AI do this job at the standard I accept?
2. How much does each run cost?
3. Where must a person still step in?

Pick the right first job

Do not start with your hardest job. Start with one that is:

• It happens weekly or daily.
• You can say what a good result looks like.
• A mistake will not cost you a client or your name.
• It needs your way of thinking, not just generic text.

Good first jobs for a $1M+ educator or consultant:

• Turning a call recording into a client summary and next steps.
• Answering the same ten student questions in your tone.
• Checking a client's draft against your framework (your method) and giving feedback.
• Sorting new leads by how well they fit your offer.

Bad first jobs:

• Anything that moves money.
• Anything that gives legal or medical advice.
• A job you cannot describe in a few sentences.

The 30-day prototype, step by step

Here is a simple plan. It fits in about a month. Adjust it to your pace.

Week 1: Write down how you decide

This is the step most people skip. It is the most important one.

Take your chosen job. Do it five times yourself. Talk through each choice out loud. Record it.

Then write down:

• What comes in.
• What you look at first.
• What you ask yourself.
• What makes you say yes, no or "it depends".
• What the finished result looks like.

This is like writing a recipe from a meal you cook by feel. The AI cannot copy what you never wrote down.

Collect 20 to 30 real past examples too. Good ones and bad ones. If your data lives in messy sheets, an AI spreadsheet tool for cleaning and summarising data can help you tidy it before you use it.

Week 2: Set the pass mark

Before you build anything, decide what "good enough" means.

The Codenovai team makes the same point in its 90-day agent runbook. It says most pilots fail because no phase has an exit rule, such as passing a set of known-good test cases.

Keep yours simple:

• Take 20 of your past examples. Keep them hidden from the build.
• Write the answer you would have given for each.
• Decide the pass mark. For example: 17 of 20 match your standard.
• Decide what counts as a hard fail. For example: a wrong claim, or a rude tone.

This set of test cases is sometimes called a "golden set" (a fixed list of examples with the right answers). You will use it again and again.

Week 3: Build the smallest version

Now build. Keep it small.

Start with one AI model, a clear set of instructions and the few tools it needs. That is all. We build this in Claude Code, because a non-technical founder can get a first version running in days.

This matches advice from Anthropic's engineering team. In a talk on building effective agents, the advice is not to chase autonomy for its own sake. Use an agent (an AI that decides its own next step) only when the task is open-ended and valuable. Otherwise, a fixed workflow gives better cost control and predictability.

Here is the order we suggest:

1. Start with a fixed chain. Step one, then step two, then step three. No surprises.
2. Put your decision notes from week 1 into the instructions.
3. Show the AI two or three of your best past answers.
4. Add a person at the risky step. The AI drafts. You approve.
5. Only then think about more agents. If one is not enough, read our test for multi-agent versus single agent.

Complex frameworks feel clever. They are also hard to fix. Anthropic's advice, as summarised in this write-up of Barry Zhang's talk, is to start with the smallest setup that works.

Week 4: Test, score and decide

Run your hidden 20 examples. Score each one.

Then ask:

• Did it hit the pass mark?
• Where did it fail? Is there a pattern?
• What did each run cost?
• How long did your review take?

Fix the biggest pattern first. Run the test again. Repeat.

At the end you make one of three calls:

Result	What it means	Next move
Passes, cost is fine	The job suits AI	Build it properly and add monitoring
Close, but fails in one area	Instructions or data are weak	One more round of fixes
Fails, and you cannot say why	The job is not clear enough yet	Go back to week 1 or pick a new job

A "no" is a good result. You spent a month, not a year.

For a deeper test plan, see our guide to AI agent testing with 10 scenarios.

What it costs

There are three kinds of cost. Plan for all three.

1. Tools. Prices vary a lot. One 2026 roundup of agent platforms says managed options start around $39 a month. It also says that self-hosting a visual builder on a $5 to $10 monthly server can be the cheapest path for small teams. The catch: you own the upkeep and the security.

2. Model fees. This is the one that surprises people. The same source notes that a busy agent running 10,000 times a day can pass $1,500 to $3,000 a month in model fees alone. A prototype runs far less, but measure the cost per run now. It tells you if the idea can scale.

3. Your time. The biggest cost. Writing down how you decide takes real hours. It is also the part that makes the system yours.

Keep a human in the loop

A prototype should never act alone on something risky.

Put a pause before any step that:

• Sends a message to a client.
• Changes a record.
• Spends money.
• Gives advice a client may act on.

The Nari Media summary of Anthropic's guidance says that most production failures come from skipping this approval gate. They also come from not recording what the agent saw, decided and did at each step.

So log everything from day one. If something goes wrong, you can read the trail.

One more point from research on multi-agent systems. A CHI 2026 workshop paper on human oversight argues that execution is where oversight matters most. Watching the machine run is not enough. You need to see whether the work is meeting the goal. Our piece on agent observability covers how to set that up.

Five mistakes that waste the month

• Starting with the tool. Pick the job first. Tools change every few months.
• Testing on clean, invented examples. Real work is messy. Use real work.
• Having no pass mark. You will argue about "good" forever.
• Adding too many agents too soon. More parts means more places to break.
• Treating risk as a one-off check. The NIST Generative AI Profile says to treat AI risk as an ongoing task. Keep an inventory, watch for errors and review often. Our 90-day governance plan shows one way.

From prototype to a real system

A passing prototype is a start, not a finish.

The Codenovai runbook lays out a 90-day path: discovery in weeks 1 to 2, build in weeks 3 to 8, and hardening in weeks 9 to 12. In hardening, an agent that fails its test thresholds is blocked from going live. That is a sensible rule.

What changes between prototype and live system:

1. Grow your golden set from 20 to 100 or more test cases.
2. Run the tests automatically each time you change anything.
3. Track cost, errors and whether the goal is met.
4. Decide what the system keeps and forgets. See AI agent memory. For how the build side works, see custom AI delivery systems with Claude Code.
5. Name one owner. One person is responsible for it.

The same steps scale up into an AI Operating System: a set of AI employees that run your method. Njin explains the idea in what an AI operating system for business is. For the full path, read our guide to AI workflow design and your first agent system in 90 days. If your team builds this in code, the practical guide to agentic workflows is the better fit.

Our view: your judgment is the product

Many founders run businesses that sit on their own expertise. For them the tool is rarely the problem. The problem is that the best decisions live in the founder's head. So the team waits. Or the founder steps in. Or quality drops.

That is why we prototype the thinking first. Not the screens. Not the chat window.

We pull out how you decide. We write it down in plain steps. Then we test whether a small set of AI agents can follow it to your standard. An agent is just one AI worker with one job. Several of them, handing work to each other, is an orchestra. Hence our name.

Three things we hold to:

• Test on real work. Your own past cases are the exam.
• Keep it simple until it fails. Add parts only when a test shows you need them.
• You own it. The system, the notes and the tests are yours. No lock-in.

And one honest warning. A prototype will not fix a business with no clear method. If you cannot explain how you decide, no tool can copy it. Start there.

James Killick

Want help running this?

If your business earns $1M+ and relies on you for the key decisions, we can do this with you.

Our Foundation Month is $3,500 one-off. It is the place to start. Our 90-day program is $18,500 one-off. It builds a working network of AI agents around your method, with hands-on prototyping from the start. See both on our AI consulting services page.

Prefer to build yourself? Platform Access is $90 per month or $900 per year.

Sources

• McKinsey, The State of AI 2025
• Codenovai, Agentic Pilot-to-Production: A 90-Day Runbook
• AI Engineer, How We Build Effective Agents (Barry Zhang, Anthropic)
• Nari Media, Building Effective AI Agents: Lessons from Anthropic's Barry Zhang
• PickMySoft, Best AI Agent Orchestration Platforms
• CHI 2026 workshop paper, Managing Multi-Agent Research Systems
• NIST AI 600-1, Generative AI Profile