Somewhere in the past year, a vendor or an eager consultant has probably offered to “build a custom AI model” for your business. It sounds impressive. It sounds expensive. And in most cases it’s more than you need.
This is an honest guide to what “custom AI” actually means for a small or mid-size business: three levels of customization, what each costs, and how to tell which one your problem calls for. Spoiler: fine-tuning — training an AI model on your own data — is usually the last step, not the first.
“Custom AI” almost never means training a model from scratch
Training a large language model from the ground up takes tens of millions of dollars in compute and a research team most enterprises can’t assemble. When someone offers your business “custom AI,” that is not what they mean.
What they mean — or should mean — is taking an existing foundation model (GPT, Claude, Llama) and adapting the way you use it. That adaptation happens on three rungs of a ladder, and the most expensive mistake in small-business AI is being sold the top rung when your problem lives on the bottom one.
Rung 1: Better prompts
The cheapest customization is no customization: a carefully engineered prompt with clear instructions, your output format spelled out, and a few good examples pasted in.
Modern models are startlingly good at drafting, summarizing, classifying, and extracting when you tell them — precisely — what you want. Most teams under-invest here, because “write a better prompt” doesn’t sound like an invoiceable deliverable. It is still the highest-leverage hour you can spend on AI.
Cost: essentially zero — a few hours to a few days of iteration by someone who knows your business.
The limit: the model knows nothing about your company. It can’t quote your refund policy or last quarter’s numbers, because none of that is in its training data.
Rung 2: Give the model your knowledge (RAG)
When the missing piece is your information — policies, product catalogs, documentation, past tickets — you don’t retrain the model. You connect it to your data. That’s retrieval-augmented generation (RAG): when a question comes in, the system searches your documents, feeds the relevant pieces to the model, and the model answers using them.
This is what powers the AI features most SMBs actually want: a support assistant that answers from your real help center and cites the article, an internal knowledge base that finds the right SOP, a sales tool that quotes your current pricing — not what the model half-remembers from the internet.
Cost: a production setup typically runs from a few thousand dollars to the low tens of thousands, depending on how many systems it has to connect to.
The limit: RAG adds knowledge, but it doesn’t change behavior. If the tone is wrong or the output format is inconsistent at volume, better retrieval won’t fix that.
Rung 3: Fine-tuning changes behavior, not knowledge
Here’s the part most articles bury: fine-tuning does not teach a model facts about your business. Fine-tune it on your product catalog and it still won’t reliably memorize prices — and you shouldn’t want it to, because prices change. What fine-tuning teaches is consistent behavior: an output format, a tone of voice, a domain’s way of reasoning.
What a real fine-tuning project takes:
- A dataset of roughly 100 to 10,000 examples of the behavior you want
- Data preparation — usually the real cost, not the training compute
- Evaluation, iteration, and ongoing maintenance, because fine-tuned models drift as your business changes
Cost: done by an external team, a business-grade project typically lands between $2,000 and $25,000 — and that’s with modern parameter-efficient methods that keep the training itself cheap.
It’s genuinely worth it when:
- You run high, stable volumes — think 100,000+ requests a month — where per-request savings compound
- You need a strict machine-readable output format: classified tickets, standardized reports, parsed invoices
- You need a consistent brand voice that prompting alone can’t hold
- Compliance requires self-hosted, open-weight models
For most SMBs, none of these apply yet. That’s not a limitation — it’s money not spent.
The whole ladder in one view:
| Level | What you get | Typical cost | Timeline |
|---|---|---|---|
| Prompt engineering | Better outputs from a general model | $0–$500 | Days |
| RAG | Answers grounded in your own data | $3K–$20K | 2–6 weeks |
| Fine-tuning | Consistent behavior, format, voice | $2K–$25K+ | 4–8+ weeks |
Three scenarios, mapped honestly
Customer support
Start with a well-prompted assistant over your help center (RAG). It answers from your real articles, cites them, and gets better when your docs do. Fine-tuning enters only when volumes grow enough that response length and consistency start costing real money.
Document processing
Invoices, contracts, applications: usually not fine-tuning either. Structured output modes plus a clean extraction prompt handle a surprising range of documents. Fine-tuning earns its cost when you process thousands of documents a month in formats base models consistently fumble.
Personalization and recommendations
For most SMB catalogs, rules plus retrieval beat a trained model: simpler, explainable, and updated in minutes. Custom recommendation models make sense when you have the transaction volume to feed them — a nice problem to have, later.
The expensive mistake: starting from the technology
Almost every failed AI project follows the same pattern: someone fell in love with a technique — an agent, a fine-tuned model, a vector database — and then went looking for a problem to justify it.
The order that works:
- Pick one task that costs real time or money today
- Measure that cost — hours, errors, missed follow-ups
- Apply the cheapest rung that could plausibly fix it
- Climb only when you hit a wall you can name
This is also how we approach custom AI development at Edgeware: start at the bottom rung, prove value in weeks, and reach for heavier tools when the problem demands them — not when they’d make a better pitch.
Where to start this week
Not sure which task deserves attention first? Take our free AI automation assessment — nine questions, about two minutes. You’ll get the one task an AI system should take over first, what it costs you today, and what it takes to build.
No custom model required — unless it actually is.
