How to Choose an LLM for Your Small Business
Pick the model that can do one real job inside your tools, on your data rules, at a cost you will keep paying.
TL;DR
Choose an LLM by the job it will do, not by the brand on the box. Test two or three models on your own examples, score them on facts and format, and only then consider a bigger model or fine-tuning. Privacy rules, daily cost, and whether the model connects to software you already use matter more than a clever demo.
Match the model to one repeatable job. Then test it on your files, your messages, and your constraints. Brand matters less than whether it runs inside your tools, stays inside your privacy rules, and stays cheap enough that people will use it every day. Start with the work, not the model list.
What is an LLM, in plain terms?
A large language model predicts the next words in a sequence. You give it a prompt. It generates text that looks like a reply, a summary, a draft, or a classification.
It does not know your business unless you give it context. It does not understand the way a person does. It is a pattern engine. That makes it useful for drafting, sorting, and extracting. It is a bad substitute for judgment on money, safety, or legal questions.
What job should the model actually do?
Write down one job that already exists. Common fits: turn a voicemail into a work-order draft. Answer a question customers ask every week. Summarize a long email thread. Pull fields off an invoice. Rewrite a service page.
If you cannot name the input, the output, and who checks the result, you are not ready to pick a model.
The job tells you what good means. For a booking confirmation, the time, name, and address have to be right. For a blog draft, a human editor should finish it in a few minutes. Accuracy and tone are different tests.
This is where AI Integration starts. The model has to sit next to the software you already use. A separate chat tab you forget to open does not count.
Do you need the largest model on the market?
No. Bigger models handle messy, open-ended writing and long instructions better. Smaller models cover classification, extraction, short replies, and templated documents.
A smaller model that can see your real records beats a famous model you are not allowed to feed customer data. If a reply is slow or costs more than the job is worth, people skip it.
Try a capable general model first on a sample of real work. If it fails the same way every time, tighten the instructions and examples. Move up in size only after that. Do not start at the top of the price list.
How should you handle private customer data?
Treat every prompt as a document you are handing to another company, unless a contract says otherwise. Customer names, job addresses, health notes, financial figures, and employee records do not belong in a consumer chatbot.
Ask any vendor four questions. Does your data train their public models. Where is it stored, and for how long. Can you run the model in your own cloud or on a machine you control. Who can see the logs.
Vague answers mean you do not send production data. Test with fake but realistic samples. Work that must stay in-house needs a model you can host, or a vendor that signs a data-processing agreement you can actually read.
Custom Software is the path when the model has to sit behind your login, your database, and your audit trail.
Want this working on your numbers?
Viewmedia makes marketing you can prove, matched to real, closed sales.
What will it cost once people use it every day?
Most vendors charge for tokens in and tokens out. A token is a chunk of text, roughly a short word. Long prompts cost more than short ones. Paste the same policy packet into every question and that packet will dominate the bill.
Estimate from volume, not from a demo. How many jobs per day. How long is a typical input. How long is a typical output. Does a person review every result.
Add the human review time. A cheap model that needs heavy editing is not cheap. A slightly pricier model that produces a usable first draft can cost less in labor.
Watch extra fees. File storage. Extra tools. Phone minutes if you add voice. Per-seat licenses for a chat wrapper. The model API is often the small line item. The product around it is not.
How do you test a model before you commit?
Build a small, ugly test. Take a few dozen real examples of the job. Strip or fake the sensitive details. Write the instruction once. Run that same instruction on two or three models.
Score each result on what the job actually needs: facts correct, format usable, tone acceptable, nothing invented. Pass or fail. Do not grade how impressive the prose sounds.
Watch for invented facts. A price, a date, a policy that was not in the input. For customer-facing text, that is a fail even if the rest reads well.
If staff will use it, have one of them run the test. Not only the owner. Tools fail when they fight the way people already work.
AI Training is the next step: teaching your team what to paste in, what to check, and when to ignore the output.
Should you fine-tune, or just write better instructions?
Most small businesses should start with instructions, examples, and retrieval of their own documents. Fine-tuning means further training a model on your examples. It costs more, needs a clean dataset, and is hard to undo if your process changes.
Write better prompts and keep a set of approved examples first. If the model still misses your format, then consider fine-tuning. Or build a system that always fills a template and only lets the model write the variable parts.
How does the model fit the rest of your software?
A model in a browser tab is a trial. A model that writes into your CRM, your inbox, or your scheduling tool is work.
Ask whether you can call it from software you already pay for. Whether output can be a form instead of an essay. Who owns the history. What happens when the vendor retires the model name.
Prefer a setup that lets you swap the underlying model. The instruction layer and the integration are the assets. The model is a replaceable part.
Do not buy a new AI platform for every task. One integration pattern, a few jobs, and a review habit beat a pile of unused subscriptions.
Founder, Viewmedia
Brian Wroblewski is the founder of Viewmedia. For more than two decades he has helped local and regional businesses turn marketing spend into provable, closed sales.
