There is a strange gap in the market right now.
On one end you have enterprise consultancies who will happily scope a six-figure “AI transformation engagement” for a company with forty employees. On the other you have a wave of “AI automation agencies” that are, structurally, a prompt pasted into a no-code tool with a retainer attached.
In between sits almost everybody: the dealer group with an inventory feed, the roofing company with eight years of job history, the restaurant group drowning in review responses. These businesses have real data, real repetitive work, and no credible path to getting software built against it.
This post is about what actually lives in that gap. What “custom AI” means when you strip the marketing off it, what it costs in effort rather than in license fees, how to tell a builder from a reseller, and — the part nobody selling this will tell you — when you should not do it at all.
The three tiers buyers confuse
Nearly every bad AI purchase we’ve seen comes from conflating these.
Tier 1 — Off-the-shelf AI products. ChatGPT, Copilot, Gemini, Jasper, the AI features baked into your CRM. You are buying a general-purpose tool and a seat license. It knows nothing about your business except what someone types into it each time.
Tier 2 — Automation with a model in the middle. A workflow that already had structure — a form submission, an inbound call, a nightly feed sync — now has a language model doing the judgment step in the middle. Classify this lead. Summarize this call. Draft this reply in our voice. Flag this job for follow-up. The plumbing is conventional software; the model is one component.
Tier 3 — Genuinely custom tooling. Software built specifically for your operation, running against your systems, with the model as one part of a larger application. An internal tool that answers questions across your inventory, your CRM, and your job history at once. An agent that takes actions in your stack, not just drafts text for a human to paste.
Here is the uncomfortable pattern: most businesses need tier 2, and get sold tier 1 at tier 3 prices.
| What you buy | Who owns it | Typical shape | |
|---|---|---|---|
| Tier 1 | Seats on a general tool | The vendor | Monthly per-user fee |
| Tier 2 | Workflow integration | You, mostly | Build cost + small run cost |
| Tier 3 | An application | You | Real project + ongoing maintenance |
Tier 2 is where the leverage is for a business under a few hundred people. It’s also the tier that’s hardest to sell, because there’s nothing to demo on a sales call — it’s invisible infrastructure that quietly stops a category of work from existing.
What actually makes an AI tool “custom”
The model is not the custom part. Everyone has access to the same handful of frontier models through the same APIs. If a vendor’s differentiation is which model they use, they have no differentiation.
Custom means it’s connected to your data. Your live inventory feed. Your CRM records. Your job history going back years. Your call recordings. Your pricing rules and your compliance constraints. That connection is the entire product.
Concretely, “custom” looks like:
- A tool that knows what is actually in stock right now, because it reads the feed, not because someone pasted a spreadsheet into a chat window last Tuesday.
- A drafting tool that knows the brand’s pricing disclaimer requirements and refuses to publish copy that violates them.
- A classifier that has seen how your company labels a lead, not how the internet in general labels leads.
And the work — the part that takes actual engineering time — is almost never the prompt. It’s:
- Getting authenticated access to the source system, and keeping that auth alive.
- Normalizing messy real-world data into something a model can reason about.
- Deciding what the tool is allowed to do on its own versus what needs a human.
- Handling the case where the data is wrong, missing, or contradicts itself.
- Making the whole thing observable so you find out it broke before your customer does.
The prompt is maybe five percent of it. Anyone who talks about prompting as the craft is selling tier 1.
The honest cost structure
This is where most buyers are miscalibrated in both directions.
The API bill is usually the smallest line item. For the volumes a local or mid-market business actually runs — classifying a few hundred leads a day, drafting a few dozen replies, summarizing calls — token costs are typically the least significant recurring expense in the system. People fixate on it because it’s the number with a dollar sign attached on the pricing page.
The real cost is everything around the model:
- Integration. Every system you touch has its own API, its own quirks, its own rate limits. Some have no API at all and need a browser automation layer, which is meaningfully more fragile.
- Auth and secrets. OAuth tokens expire and rotate. Refresh flows fail silently. Credentials need to live somewhere that isn’t a config file in a repo.
- Error handling. What happens when the upstream feed returns half the records? When the model returns malformed output? When a rate limit hits mid-run? Silent partial success is the most expensive failure mode in automation, because nothing alerts and the bad data propagates.
- Monitoring. If a nightly job stops running, how long until someone notices? If the answer is “when a customer complains,” you don’t have a system, you have a demo.
- Maintenance. APIs deprecate. Models get retired and replaced. Your own business process changes. A tool built and abandoned degrades on a timescale of months.
A useful framing: the build is maybe half the lifetime cost. Budget for the tool to be owned, not delivered. Anyone quoting you a one-time build with no ongoing story is either underpricing or planning to disappear.
How to vet an AI development company
Five questions. They’re cheap to ask and they separate builders from resellers quickly.
1. “What happens when the model API is down?” A real answer describes retries, fallbacks, queuing, and what the user sees in the meantime. A reseller has never thought about it, because their tool has never had to survive a bad day.
2. “Where do the credentials live?”
You want to hear a secrets manager, a platform secret binding, an OS keychain — something with a
name. If the answer is vague, or involves a .env file emailed around, you now know how they’ll
treat your customer data. Ask specifically whether secrets ever get written to logs.
3. “Show me something you run on your own business.” This is the strongest filter. Anyone can produce a client logo wall. Very few AI vendors run custom AI on their own operations, because it’s harder than selling it. We run an in-house agent against our own stack and a multi-client portal we built ourselves — not because that’s remarkable, but because a vendor who won’t eat their own cooking is telling you something.
4. “Who owns the code, and who owns the data?” Ask whether the repository is transferable to you. Ask whether your data trains anyone’s model. Ask what happens if you fire them next quarter — do you get a working system or a dead login? Get it in writing. “We host it” and “you own it” are not the same sentence.
5. “What would you talk me out of?” If a vendor can’t name a scenario where you shouldn’t buy from them, they’re not evaluating your business, they’re closing. Which is a good transition.
When you should not build custom
Genuinely: most requests we get for custom AI should not become custom AI. Four disqualifiers.
An off-the-shelf tool already covers 80% of it. If your CRM has a decent AI summarizer built in, use it. Building your own to capture the last 20% means owning maintenance forever to close a gap you’ll stop noticing in a month. Buy the boring thing. Spend the build budget where nothing off-the-shelf exists — usually somewhere specific to your industry.
Your data is a mess. This is the most common one. If your CRM has four spellings of the same lead source, if half your job records have no close date, if nobody agrees on what “qualified” means — an AI layer on top will produce confident, well-formatted, wrong answers. Faster wrong is worse than slow right. Fix the data first. That project is less exciting and more valuable.
Nobody internally will own it. Every automated system needs a human who notices when it stops. Not an engineer necessarily, but someone whose job visibly gets harder when the tool goes dark. If you can’t name that person before the project starts, the tool will run for a while, break quietly, and become a line item nobody can explain.
The process isn’t stable yet. If you’ve changed how you handle a workflow twice this quarter, don’t encode it in software. Automation makes a process rigid and cheap. You only want that once you’re confident the process is right. Run it manually until it’s boring, then automate the boring version.
There’s a fifth, softer one: if the honest goal is to look modern rather than to remove work, the project will be judged on how it demos and abandoned when it’s time to maintain it.
Why owning a tool compounds
Here’s the argument for doing it anyway, when it does make sense.
Advertising is a rental. You spend in June and you get June’s leads. Stop spending in July and July looks like you never advertised. It’s a valuable rental — we run a lot of it — but it resets every month and the price goes up as competition does.
A tool you own behaves differently. Build a system that classifies inbound leads correctly and it does that in month one and month thirty, at effectively the same cost. Better than that, it accumulates: every lead it processes is another labeled record you own, which makes the next version better. The asset appreciates while the ad account depreciates.
That’s the real case for custom, and it has nothing to do with AI being new. It’s the same reason owning your building beats leasing it, applied to operations. The businesses that will be hard to compete with in five years are the ones that spent this year turning their repetitive work into systems they own, instead of spending it entirely on media.
Most of them should start at tier 2, with one workflow, that one person actually owns.
We build custom AI and automation for local and mid-market businesses out of Mandeville, Louisiana — and we run the same stack on our own operation. If you want a blunt read on whether your idea should be built or bought, tell us what you’re trying to fix.