AI Prototyping for Business: A 90-Day Plan for Founders

AI prototyping for business means building a working AI system that runs one real workflow, with a human checking the output. Not a slide deck. For a founder-led consulting or education business, 90 days is enough to get a live prototype, real numbers and a scale decision.

Here is the plan we use, step by step.

AI prototyping for business: the 90-day roadmap

Think of a kitchen before opening night. You test each station before the doors open. A prototype works the same way. You test one workflow, with real tickets, before you scale.

We build this with Claude and Claude Code, because they let a non-technical founder get a first agent system running in days, not months. The goal is not a clever one-off. It is the first piece of an AI Operating System: AI employees that carry your method, so the work stops waiting on you.

Four phases:

1. Weeks 1 to 2, extract. Pick one workflow. Write down your decision rules. You sign off the rulebook.
2. Weeks 3 to 6, build. Build two agents. One drafts, one checks. Your ops lead signs off the test outputs.
3. Weeks 7 to 10, validate. Run it on real client work. A human checks every output. Your delivery lead signs off quality.
4. Weeks 11 to 13, decide. Review the numbers. Widen the rollout, or don't.

Each phase ends with something you can hold:

• A written rulebook of your expert judgement.
• A working two-agent prototype on sample tasks.
• Real output, checked and scored.
• A go or no-go call backed by data.

Schedules slip for boring reasons: no clear sign-off owner, missing sample data, or scope creep. Name one sign-off owner per phase. Freeze scope before week 1.

Start with your most repeatable, highest-volume workflow. Leave the messy, judgement-heavy ones for later. The 90-day prototype breakdown goes task by task.

What to build: the core components

A prototype needs ingredients, a recipe, staff and a pass. The pass is where a dish gets checked before it leaves the kitchen.

• Data. Collect 20 to 50 real examples of the task done well. Strip out personal details. Label the good ones and the bad ones.
• Prompts. Write instructions the way you would brief a new hire. Keep them portable so you are not locked to one tool.
• Agent roles. One agent drafts. One agent critiques. That is a generator and a critic.
• Orchestration. Keep it simple. A pipeline, where one agent hands off to the next, beats a hub with many agents reporting to one coordinator in early rounds.
• Human checkpoints. A real person reviews every output until errors drop and stay low.

A Harvard Business School field experiment looked at AI as a teammate in real work. The lesson holds for your build: AI drafts, people decide.

Don't fine-tune a model on day one. Test with prompts and a general model first. Fine-tuning comes later, once you know the workflow works. Claude Code supports this well, because you can split the drafter and the checker into separate sub-agents with their own instructions. Our guide to designing your first agent system has role templates.

Governance, IP and compliance before you pilot

Before a prototype touches client work, lock down the basics.

Fully AI-generated output usually isn't protected by US copyright. You need documented human input to claim ownership. Founders Legal explains that a record of human edits and decisions is what supports a claim. The US Copyright Office keeps its guidance on AI works in one place. For a wider risk checklist, the NIST AI Risk Management Framework is the standard reference.

Build these guardrails in from day one:

• Keep a version log of every prompt and every human edit.
• Set an approval gate. No output reaches a client unchecked.
• Write a short test plan before each pilot run.
• Get client consent where their data runs through the system.
• Don't promise clients what the system does until you have tested it.

Measuring success: KPIs that prove the value

You can't manage what you don't measure. Track four numbers from week one.

Metric	What it shows	How to collect it
Throughput	Capacity gained	Task logs, timestamps
Quality index	Share of outputs that pass review first time	Human review scorecards
Founder hours saved	Time freed for you	Weekly time tracking
Revenue impact	Client capacity freed up	Client intake records

Run a short pilot of two to four weeks. Compare the AI-assisted workflow with the old manual one. Set the pass mark before you start. For example: same quality, half the time.

Put the four numbers side by side on one dashboard, updated weekly. That is enough to make a scale call.

Week-1 checklist: assets, costs and first actions

Gather three things before you start:

1. A written version of your framework or method.
2. Ten to twenty real client examples, good and bad.
3. A list of the decision rules you use that aren't written down yet.

Cost drivers to watch: how many workflows you prototype at once, how messy your source data is, and how many review rounds it takes before quality settles. Keep scope to one workflow to control spend.

Write your rulebook in week 1. Don't pick a tool first. The clarity of your rules matters far more than the tool.

By week 4 you should have a first draft of the two-agent system, tested on sample tasks. An AI automation audit gives you a template for this first pass.

Why you must redesign the workflow, not just add a tool

Bolting an AI tool onto your current process rarely pays. McKinsey's research on how organisations capture AI value points the same way: value shows up when work flows differently, not when a tool sits on top of the old flow.

It is like swapping the stove but keeping the same jammed layout. You cook faster. The kitchen still clogs.

Ask three questions:

• Which steps need your judgement, and which are repetitive checking?
• Where are you the bottleneck, approving things a documented rule could approve?
• Which handoffs between people could become handoffs between an agent and a reviewer?

MIT Sloan separates AI tools from AI solutions. Tools help one person. Solutions change the system. A prototype aimed at business value has to be a solution.

This is why the rulebook comes first. You are not writing it down to keep the process the same. You are writing it down so you can redesign it.

Connecting the prototype to the tools you already use

A prototype that lives in a new app gets ignored. Plug it into where your team already works: your CRM, scheduler, content platform and support inbox.

A sensible order:

• Connect the agents to where your source data already lives, like documents, call notes and past client work.
• Send outputs to the place your team checks daily, such as your project board.
• Add a notification so a reviewer knows when something needs a look.

If your team lives in Slack, build the review step there. Our Slack AI workflows playbook shows the pattern. If you have custom billing or a bespoke client portal, bring in an engineer for that connection and keep the prototype itself simple.

The prototype should cut clicks and tools, not add them.

Turn your IP into an AI Operating System, not a pile of prompts

Here is the thesis behind all of this. Your method is the asset. Most founders hold it in their head, so every client output waits on them.

A prototype is the first proof that the method can run without you. Scale it, and you end up with an AI Operating System: a set of AI employees, each carrying a piece of your judgement, working together under human checkpoints. That is the shift from supply-constrained to build-ready.

We build this with Claude Code, and the custom AI delivery systems guide shows what that looks like. If you're not technical, start with Claude Code for non-technical founders. For the bigger picture, Njin's piece on what an AI Operating System is for business is a good companion.

Change management: getting your team to use it

The best prototype fails if the team won't use it. Adoption is a people problem as much as a build problem.

• Involve the people doing the work in writing the rulebook, not just you.
• Run the first pilot with volunteers, not the whole team.
• Share early wins fast, such as time saved on one task.
• Train reviewers on what good output looks like before they check AI drafts.

Treat it like onboarding a new hire. You wouldn't hand someone a client file on day one with no guidance. Don't do it to an agent, and don't expect your team to trust one that hasn't earned it. Why the best systems keep humans in charge goes deeper on that trust loop.

Scaling from one workflow to many

One working prototype is proof of concept, not a finished system. Scaling means repeating the process on new workflows, one at a time.

Research on the jagged frontier, the uneven way AI helps some tasks and hurts others, published in Organization Science, is a good reason to go slowly. A win on workflow one doesn't guarantee a win on workflow two. Each new workflow gets its own short validation round.

A sensible order:

• Scale the highest-volume, most repeatable workflow first. It pays back fastest.
• Add a coordinator agent only once your first drafter and checker are stable.
• Keep human checkpoints on every new workflow until errors hold steady.
• Write each new rulebook the same way you wrote the first.

Build on one shared setup that can hold many workflows. Don't start from scratch each time.

Where to start

Most founders overbuild before they validate. The fix is small. One workflow, one rulebook, one human checkpoint, tested fast.

If you want help, the 90-day program takes you from a written rulebook to a tested agent system, built alongside your team. The Foundation Month covers extraction and the rulebook first, if you want a smaller step. After the build, Platform Access gives you the tools to keep running and extending your agents.

An assessment shows you which workflow to prototype first and what the 90 days would look like. Book your assessment and we'll map it.

Sources

• How to manage the two faces of GenAI, tools and solutions, for business success | MIT Sloan
• The state of AI: how organisations are rewiring to capture value | McKinsey
• The cybernetic teammate: a field experiment on generative AI reshaping teamwork and expertise | HBS
• Copyright guidelines for AI-generated material | Founders Legal
• Organization Science, jagged frontier research | INFORMS
• Sub-agents | Claude Code docs
• AI Risk Management Framework | NIST
• Copyright and artificial intelligence | US Copyright Office