How I Would Build a One-Person AI Consulting Business with ChatGPT Codex or Claude Code

A one-person AI consulting business is realistic today, but only if the work is framed correctly. Plenty of people can build an automation. Far fewer can walk into a company, find the part of the operation that actually holds it back, and fix that part in a way the owner can measure. This note is my working plan for doing it with ChatGPT Codex or Claude Code as the daily workspace.

Sell outcomes, not builds

The label you choose changes the conversation. "I build automations" puts you in the same bucket as every freelancer with a workflow tool. "I help your business get a specific result with AI" puts you next to the owner, looking at the same problem. Builders sell features. Partners sell outcomes, and businesses pay more for outcomes.

ChatGPT Codex and Claude Code make this practical for one person because the interface is plain language. You do not need a computer science background to describe a workflow clearly, and a build that used to take half a day can often be drafted in under an hour. That lowers the barrier to starting. It does not remove the need to understand the business you are working in.

Three buckets every project must fit into

Almost every business wants one of three things, and every project I take should be tied to exactly one of them:

• Get more customers. Lead qualification, faster follow-up, more booked appointments from an existing source.
• Make each customer worth more. Better onboarding, CRM automation, higher retention, more repeat purchases.
• Cut costs. Fewer hours per task, fewer errors, fewer support tickets, faster turnaround.

If a project cannot be placed in one of these buckets, it will be hard to prove its value later, so it is probably not the right project. Some work touches two buckets, such as onboarding. Choose the main one anyway. A single clear target keeps the whole engagement honest.

Build it for yourself first

I would not sell something I have not used on my own work. Before looking for clients, I would spend a couple of weeks building a personal AI operating system in ChatGPT Codex or Claude Code and running my own business through it.

A good first project is a morning brief: a scheduled routine that reads my calendar, task list and inbox, then summarizes what matters today. It looks simple, but building it teaches how the tool handles schedules, context, instructions and recurring work, and it teaches this on real data instead of a tutorial. Afterwards it is a template. The data sources change from person to person, while the pattern stays the same. It also gives me something concrete to show when someone asks what I have actually built.

Start broad, then let evidence choose the niche

The usual advice is to pick a niche on day one. If you already worked in an industry, know many people in it, or are truly fascinated by it, that advice is good and you should follow it. Without any of those, a niche chosen from zero data is a guess, and defending a guess for months is a good way to avoid real work.

My alternative for the first month is simple: have five real conversations with five business owners, in any industry. The goal is not revenue yet. The goal is repetitions and pattern recognition. After a handful of engagements, the same problem tends to appear in very different companies, and that repeated problem is what tells me where to focus.

Where the first conversations come from

I would go in this order:

1. Warm contacts. Former colleagues, classmates, people I have met at events. Most people ignore this network, yet it is the fastest route to a first paid project. The message is not "hire me". It is closer to: "I built this for my own business. Can I show you and hear what you think?"
2. Marketplaces such as Upwork. Buyers there have already posted a job and set aside budget, so there is nothing to convince them of. Search for terms that map to the service: AI workflow integration, automation setup, custom workflow design.
3. Building in public. Share what I build, what it does and what I learned, on whichever platform my audience uses. Over time this works as a portfolio. Anyone who searches my name finds real work.

A mindset shift helps here: aim for ten "no" answers rather than the first "yes". If each conversation is data, a refusal stops feeling personal, the pitch improves quickly, and the first yes often arrives before the tenth no.

Look for the constraint, not the annoyance

When a business owner says "I want an AI agent for this", the stated task is not always where the leverage is. I would ask them to walk me through the operation in order, from the first lead to the delivered work, and listen for friction: where they sigh, where the team complains, where something is done again and again by hand. The first point of friction from the front of the process is usually the real constraint.

Sometimes the right fix is not AI at all. A plain workflow can be cheaper, more reliable and less risky than an agent. Recommending that is part of being a consultant.

Four blanks to fill before building anything

Before writing a single line, I need to complete this sentence:

This project is in the [bucket] bucket. The KPI is [metric]. Today it stands at [baseline]. After 60 days I expect [target].

If any blank stays empty, there is no project yet, only a demo. Filling all four means the client can later see whether the number moved. A project with a measured before and after becomes a real case study, and the client's industry, size and stage become the profile of the next ideal client. That is how a niche emerges from evidence rather than guessing.

A service ladder that earns trust step by step

Trust is built in steps, and trying to jump straight to a monthly retainer with no proof rarely works. I would use four steps:

• Education or consulting hour. A single session where I teach the owner or team to use AI on real tasks, or set up their first project. Low commitment for them, and a quiet form of discovery for me.
• Audit. Paid scoping. I map the workflows, separate what is automatable from what is politically sensitive or simply nice to have, and deliver a short document with a proposal for the first project and the result I expect.
• Project. One focused build with the four blanks filled in, delivered end to end.
• Retainer. Ongoing improvements and support, which turn one-off income into predictable monthly income.

Each step is earned by the one below it. The first one can be skipped when I am already working with a client.

Pricing: where I would start

These are starting ranges to test and adjust to the local market, not guarantees:

• Education or consulting hour: roughly $100 to $500 per hour.
• Audit: roughly $500 to $3,000 as a flat fee.
• Project: roughly $2,500 to $10,000 per focused build.
• Retainer: roughly $3,000 to $10,000 per month.

For a project, I would anchor the price against what the manual version costs the client today. If a task takes fifteen hours a week and the automation reduces it to one, the arithmetic is on my side and I should show it. I would also prepare for the follow-up question every buyer eventually asks: how did you arrive at this number?

Ask what you actually want

This model asks for conversations with strangers, public sharing of results and a willingness to sell. If that sounds unpleasant, a different role may fit better: partnering with someone who enjoys sales, or becoming the AI lead inside a company. Running your own practice is one way to benefit from this shift, not the only one.

The first client is the hardest. After that, the case study and the referral start to compound. So the practical sequence is: use it on my own business, have conversations with real owners, fill in the four blanks, prove one result, and let that result define who I work with next.

This plan was inspired by a public video about building a one-person AI consulting business with Claude Code.