Custom AI for a Small Business: What You’re Actually Buying

At small-business scale, custom AI development almost never means building or training a model. It means paying someone to wire an existing one — the same models behind ChatGPT and Claude — into the tools you already run, so a specific piece of work happens without a person grinding through it. The model is rented, and it is the cheapest line on the bill. What you are actually buying is the wiring, and then the thing nobody puts in the quote: someone who notices when the wiring comes loose.

What does “custom AI development” mean when you’re not a tech company?

Start with what it isn’t. Nobody at this scale trains a model. You rent one from Anthropic, OpenAI, or Google for cents per task, the way you buy electricity without owning a power plant. The expensive research happened somewhere else, and you get it on a meter.

The custom part is everything around the model: connections into your accounting system, your field app, your project board, and the spreadsheet somebody maintains by hand; written rules that teach the agent your formats, your job codes, and your exceptions; a review step so nothing leaves the building without a person signing off; and logs you can actually read when you want to check its work. It is real software development. It is integration work, not research.

It is also narrower than the marketing suggests. In an April 2026 Census Bureau working paper on how AI spreads through firms, 57% of AI-using companies ran it in three or fewer business functions. Custom AI at this scale is not an everything-machine. It is two or three workflows, wired properly, and then left alone to run.

So what is the money actually buying?

Strip out the part that gets demoed and the invoice covers five things, in rough order of how much of the price they represent:

1. Access. Getting into every system the work touches, in accounts you own, with credentials that survive a password rotation.
2. Translation. Writing down the rules that currently live in one person’s head — which customers get the rounded number, which job codes roll up where, what to do when a field comes through blank.
3. Proof. Running the new version in parallel against the manual one until the numbers match twice, because nobody should trust it on day one.
4. Watching. Checking every cycle that the sources returned data, that nothing unrecognized showed up, and that the totals still reconcile — before anything sends.
5. Fixing. Being the person who deals with it in April when somebody renames a column.

The model sits underneath all of that and barely registers: for a single automation, usage usually costs less per month than one software seat. You are not buying intelligence. You are buying plumbing and attention, and attention is the part that keeps costing money after launch.

Where the surprise costs hide

Three things drive the build price, and none of them are mysterious: how many systems the work touches, how expensive a mistake would be — which sets how much testing and review it needs — and how messy your data is going in. Any developer worth hiring will quote against all three out loud.

The cost that ambushes people is upkeep, because on a fixed-scope build it never appears on an invoice at all. It gets paid in your controller’s Friday afternoons — the most expensive hour in the building and the one nobody bills. Source files change shape, credentials expire, and a new job code shows up that the logic has never seen, so the automation reports a confident wrong number instead of stopping. We wrote up the specific failure modes in what breaks a weekly report automation; almost none of them are failures of the AI, and every one still lands on somebody.

That is why we price this as a monthly service instead of a build fee. Standing Reports is $1,500 a month, flat, month to month: one recurring report or workflow, its source connections, the transformation, scheduled delivery, break monitoring with the fixes included, unlimited format changes, and a monthly note on what it actually saved you. Your first correct report lands within 14 days of kickoff or month one is free. A whole back office is a bigger build — $12,500 up front plus $3,500 a month — but most projects start at the $1,500 line.

We onboard four new clients a month. That is not scarcity marketing — it is the real capacity of a two-person shop, and the reason we turn work down rather than run somebody’s numbers badly.

Product or service: the question that actually sets the price

Two different things get sold under the same words. Software sells you a place to build the thing — a login, a connector list, a query builder, a free trial — and the assembly, the checking, and the fixing stay yours. A service sells you the finished thing and the person who notices when a source quietly changes shape. Comparing the two feature by feature is a category error: a feature list tells you what a tool can do and nothing about who is doing it in month seven.

Five questions sort any vendor faster than a demo. Ask them of software companies, of contract developers, and of us:

• When one of my systems changes next quarter, who fixes it, how fast, and at what rate?
• Who finds out first that the output is wrong — you or me?
• Do I get the logic, the prompts, and the connections in accounts I control, or does this only run inside your platform?
• If I stop paying, does the work keep running, and can I take it with me?
• What is explicitly outside your support boundary?

That last one is the honest question, and good software vendors answer it cleanly: their job ends at the edge of their product. The export somebody runs by hand, the login that expires, the job code nobody told the system about — all of that stays on your side of the line. That genuinely is the deal, and it is a fine deal if you have somebody to hold the other end of it.

Our answers, for the record: we fix source changes as part of the monthly price; we monitor before anything sends, so we find out first; everything runs in accounts you own; and month to month means month to month — if you leave, you keep the logic, the access, and the documentation.

When you don’t need any of this

A $20-a-month chatbot subscription is often the right answer, and we would rather say so than sell you something. Off-the-shelf is enough when:

• One person does the task and can paste in whatever context the AI needs.
• The work is drafting, summarizing, or answering questions, and that person reviews the output anyway.
• Nothing has to move between systems on a schedule — somebody starting it by hand is fine.
• The output is for internal orientation, so a stale number costs you an awkward meeting rather than a customer.

You are into custom territory when the work pulls from more than one system, has to run whether or not the person who normally does it is on vacation, leaves the building for a customer or a lender, or gets repeated by several people who each do it slightly differently. There are also jobs we turn down outright no matter how well they pay, and we keep that list public on what we won’t automate.

What does the build actually look like?

Kickoff is one call where you walk us through the work as it happens today and where each number comes from. We do the digging from there — you don’t assemble anything for us. When the parallel run matches your manual version, you stop doing it by hand and it becomes ours to run. Nothing ships in a mode where it can send or change a record without a person signing off.

What’s a realistic first result?

One workflow, measured in hours returned. Our clearest example is a client-services team whose weekly project-update emails now draft themselves — a person still reviews and sends — and who got roughly two days a month back. That is the honest scale of a first win. Anyone promising more than that from a first project is selling you the second one early.

The gap in that third number is the whole reason this work exists. Individuals already have the tools; the business hasn’t captured them. Buying access is easy, and plenty of people already have it — turning access into a process that runs whether or not anybody remembers to run it is the part that requires building something.

The same pattern shows up by size. Census Bureau survey data through spring 2026 has AI use climbing at firms with 20 or more employees and flat below that line, and the technology is identical on both sides. What the bigger company has is slack: a role with enough room in it to catch a number that looks wrong before anyone acts on it. Buying that slack is most of what this kind of service is really for.

Frequently asked questions

Does custom AI development mean training a model on our data?

No. The agent uses a commercial model and reads your data at the moment it does the task, the way an employee looks something up. Nothing is trained on your numbers, and your contract should say they can’t be used to train anyone else’s product either. Ask software vendors for the same thing in writing.

How long before something actually works?

Our commitment is the first correct report within 14 days of kickoff, or month one is free. Work that spans several systems or carries compliance requirements takes longer to settle down after that. If someone quotes six months before anything works at all, the scope is wrong, not the technology.

What if our data is a mess?

That’s the normal case, and it’s the part we’re actually good at — cleaning and connecting messy small-company data is what we did for six years before we touched AI. You don’t tidy up first. If your sources are genuinely too broken to automate, we say so up front rather than after you’ve paid.

How is this different from paying a developer to build it once?

A one-time build gives you a working system and the maintenance that comes with it. That is the right shape when you have an internal owner who will take that on and enough of their week to do it. If you don’t, the build quietly degrades — and by the time it does, the person who wrote it has moved on to another client.

Will a custom AI agent replace people on my team?

The evidence says that isn’t the pattern. In the Census Bureau’s April 2026 working paper, only 2% of firms reported AI-related employment decreases, and 66% of users applied it solely to augment existing work; the U.S. Chamber of Commerce found 82% of AI-using small businesses grew their workforce over the past year. The realistic outcome looks like that reporting project: same team, minus the grind.

We’re two brothers — Ian works out of Midland, Texas, Cal is in Denver, and the clients are anywhere. We’ve been doing this since 2020, starting with data foundations and dashboards years before anyone was buying AI. That unglamorous background is most of the reason the agents hold up.