Everyone asks what AI costs. Almost nobody answers with a number, because the honest answer depends on what you are actually doing with it, and most of the people quoting figures are selling a subscription rather than building something.

So we measured one real project, start to finish, and are publishing the receipt.

What we built

A local chamber of commerce needed a way to find and recruit new members. They had a spreadsheet, a membership roster, and a president with a full calendar. What they did not have was a list of who to call, why that particular business, and what to say.

We built them a recruitment CRM. Three days, start to handover, no invoice. It was a give-first project to build a relationship in our own community.

The pieces:

18,285 active registered businesses pulled from the state's open data portal and filtered to the chamber's territory
Cross-referenced against the president's own LinkedIn connections to find the businesses where he already knew somebody
Suppressed against the current 342-member roster, so no existing member could ever be sent a recruitment pitch
Cut to 49 target companies, each with a named decision maker and a researched profile explaining what the business does and why the chamber would matter to them
Loaded into a free HubSpot account with a seven-stage recruitment pipeline, 27 scheduled outreach tasks, and message copy written for each contact
Documented with a user guide, a method document, and handover instructions, so the chamber owns the system rather than renting us

Then we handed over the keys.

The number: about $285

We priced the entire three days against published API rates for the model we used. Not an estimate. Measured from the actual session log.

$285.

That covers every question asked, every file written, every piece of research, and every correction across three days of work.

For context, we asked the same question a second way: what would this have cost as billable consulting hours? Reconstructed honestly, the same deliverable is somewhere between 29 and 52 hours of skilled work. At a normal consulting rate that is several thousand dollars.

We want to be careful here, because this is where most AI cost claims go wrong. That comparison is real but it is not the whole story, and the rest of this post is the part that matters more.

Where the money actually went

This surprised us, and it changes how we work.

Of that $285, less than $23 was the cost of everything the AI actually wrote. All the research, the documents, the message copy, the code. Twenty three dollars.

The other $262 was the cost of the AI re-reading the conversation.

Here is why. The system has no memory between requests. Every single time you ask it something, the entire conversation up to that point gets sent again so it knows where things stand. By the end of a long project, each new question is dragging three days of history behind it.

Over those three days that came to roughly 200 million units of text processed, and about 91 percent of it was the same conversation being read over and over.

The practical consequence: cost grows with the square of how long your conversation is. The four hundredth question in a thread costs far more than the fourth, even if it is the identical question.

If you take one operational thing from this post, take that one. Long sessions feel productive and are expensive. Finishing a piece of work, writing down where you landed, and starting fresh from that summary is not just tidy. It is materially cheaper.

About half the work never needed AI at all

This is the part we would rather you hear from us than discover on your own.

| The work | Did it need AI? | | --- | --- | | Pulling and filtering 18,285 businesses | No | | Matching connections and suppressing members | Partly | | Researching the target companies | Yes | | Building the CRM and importing records | No | | Writing the outreach copy | Yes | | Writing the guides and documentation | Yes |

The business list came from a single free public request to the state's open data portal. No vendor, no subscription, no scraping. Most states publish their business registry this way. Anyone with a spreadsheet and an afternoon can pull it.

Building the CRM itself was ordinary administrative work. Clicking through settings, defining fields, importing a file.

The genuine compression was in research and writing. Reading about two dozen companies and understanding what each one does, then producing documentation someone else can actually operate from. That is where three days replaced several weeks.

We say this because "AI did the whole thing" is the claim you will hear from people selling AI, and it was not true here. Roughly half of this project was labor that a competent person does with ordinary tools.

Where it was wrong

Four companies made it onto the ready-to-contact list that should never have been there.

The system matched a company name against the state business registry, found a business with that name inside the chamber's territory, and treated that as confirmation. The registry did confirm a business by that name existed locally. It did not confirm it was the same business the contact actually worked for.

One of those four was headquartered in Miami. One in Buffalo. One in San Antonio. One in Redondo Beach.

Every one of them would have reached the chamber president as a live prospect, and he would have spent a first impression on a meeting that never should have been booked.

They were caught because we asked for deeper research on records that had already been checked and accepted. Nothing flagged them. No error appeared. The automated verification said the list was clean.

That same pattern showed up three more times inside the build itself. A pipeline that reported a successful rename while quietly keeping its old settings underneath. A column that verified as populated but rendered empty on screen because it was reading a different field. A task list that existed in the database and displayed as nothing at all.

Each one passed its check and failed the moment a human looked at the screen.

What this means if you are considering it

Three honest conclusions.

The compute is cheap. Cheaper than most people expect. If you are weighing whether AI is worth it for a real project, the token cost is very unlikely to be the deciding factor. On a project of this size it rounded to a couple of hundred dollars.

Not everything you are paying for needs AI. Some of the most valuable moves here, like pulling a free public dataset instead of buying a list, had nothing to do with AI at all. Be suspicious of anyone who wraps ordinary work in an AI story.

The checking is the job. The build was fast. What made it deliverable was somebody looking at the screen and noticing that four confidently-listed prospects were in the wrong states. AI is very good at producing something that looks finished. Deciding whether it actually is finished is still human work, and it is the part we would never hand off.

That is the honest version. If you want to talk about whether a project of yours fits this shape, we offer a free consult and no pitch attached.

What does it actually cost to build something with AI?