Models
Which AI model should a small business use?
You have a free hour, three AI apps open in three tabs, and a dozen articles telling you each one is the best. You just want to know which one to pay for.
Does the AI model you pick really matter?
Less than the headlines suggest. ChatGPT, Claude and Gemini can all draft a reply, summarise a long email thread, tidy a quote or turn rough notes into a job sheet. For everyday business work, the gap between them is small, and it moves every few months as each one is updated.
What decides whether AI saves you time is everything around the model.
The leading three are close for everyday jobs
A plumber asking for a polite reply to a customer who wants a cheaper quote will get a usable draft from any of them. Benchmarks measure maths puzzles and coding tests. Your business runs on enquiries, invoices and reminders, where all three do the job.
The research says the same
Stanford's 2025 AI Index report found the score gap between the top model and the tenth-ranked model fell from 11.9% to 5.4% in a year, with the top two separated by 0.7%. When the field is that tight, a league table is a poor way to choose a tool for your till, your diary and your inbox.
The gap is in the set-up, not the model
One café types a question into a chat window when it remembers to. Another has written down how it answers booking enquiries, gives the AI its menu and opening hours, and has a person read anything before it goes out. Same model, very different results. We call that second part the harness, and our post on what harness engineering is explains it in plain terms.
Every model changes every few months
Whatever is top this month will be overtaken. If your choice rests on a league table, you will be tempted to switch every time a new one is announced. If it rests on your tools and your job, you can ignore most launches.
- 11.9%score gap between the top and tenth-ranked model, a year earlier
- 5.4%the same gap a year later
- 0.7%between the top two models
Source Stanford HAI, 2025 AI Index report.
How to choose an AI model in four steps
- 01
Start from where your work already lives
If your email, calendar and files are in Google, Gemini is built into the tools you open every day; Google's help page on Gemini in Gmail shows it summarising threads and drafting replies. If they are in Microsoft 365, Copilot is. Microsoft's Copilot overview explains which licences include Copilot Chat and which add Copilot inside Word, Excel and PowerPoint. Using the AI that sits next to your email saves copying and pasting, and keeps your data in the accounts you already pay for.
- 02
Name the one job you want help with
"Use AI" is not a job. "Draft replies to booking enquiries" is. A bookkeeper might pick chasing late paperwork from clients; a venue might pick answering the same ten questions about parking and access. If you are not sure which job to hand over first, read where to start with AI in a small business.
- 03
Test two models on the same real task
Take five real examples of that job from last month, with names removed. Give the same instructions to two models and compare the answers side by side. Judge them on what you would actually send, not on which one sounds cleverer. Our guide to trialling AI on one job sets the pass or fail bar.
- 04
Check the plan as well as the model
The same model behaves differently on a free personal plan and a business plan: who owns the account, what happens to your conversations and how much you can use it. Our post on free or paid AI for business covers what changes.
Which AI to try first, by the tools you already use
| If your business mostly runs on | Try first | Why |
|---|---|---|
| Gmail, Google Calendar and Google Drive | Gemini | It is built into Google Workspace, so it can work next to your email and files |
| Outlook, Teams and Word | Microsoft Copilot | It is built into Microsoft 365, next to the documents your team already shares |
| A mix of tools, or mostly writing and documents | ChatGPT or Claude | Both are strong general assistants; test the two on your job and keep the one you prefer |
| Your own website, booking system or database | Whichever connects to it | The connection matters more than the model; both ChatGPT and Claude can use outside tools through the open MCP standard |
Three businesses, three different answers
The same four steps give different answers in different businesses. Here is how they play out for three typical owner-run firms. The details are illustrations, not client records.
A heating engineer on Google Workspace
Two vans, one office manager, everything in Gmail and Google Drive. The job that eats the week is replying to quote requests that arrive at 9pm. Gemini is already next to the inbox, so the test is simple: five real enquiries, the price list and the call-out policy pasted into the instructions, and a draft for each. The office manager reads and sends. Nothing has to be copied between apps, which is half the time saved.
A dental practice on Microsoft 365
Reception lives in Outlook and Teams. The job is answering the same questions about appointments, fees and parking. Copilot sits inside Outlook, so it is the first to try. Patient details are health data, which UK GDPR treats as special category data, so the practice keeps the AI on its business accounts and away from clinical records. The GDPR Compliance Kit is where that paperwork gets done properly.
A three-person design agency on a mix of tools
Slack, a project tool, Google Docs for some clients and Word for others. No single suite wins, so the agency tests ChatGPT and Claude on its real job: turning a messy client call transcript into a tidy brief. It keeps whichever needs fewer edits, and writes its house style into one shared document so the choice can change later without starting again.
How to run a fair side-by-side test in one afternoon
- 01
Write the instructions once, outside either app
Put your instructions in a plain document: who you are, what the job is, what good looks like, what never to say. Paste the same text into both models. If you tweak it for one, tweak it for both. Our guide to writing AI instructions it will follow has a template.
- 02
Use real examples, with names taken out
Made-up examples are too tidy. Use last month's awkward enquiry, the customer who wanted three things in one email, the quote with a missing measurement. That is where the models differ.
- 03
Score what you would have to change
For each answer, note three things: did it get a fact wrong, would you change the tone, and how many minutes did it take you to make it sendable. Facts wrong is a fail whatever else it got right.
- 04
Pick the one that needs less of you
Add up the minutes. As an example, if one model needs 4 minutes of editing per reply and the other needs 2, on 20 replies a week that is 40 minutes back every week (example figures, not a measurement). That is a better reason to choose than any headline.
Common mistakes when choosing an AI model
Most bad choices come from choosing for the wrong reason, not from choosing the wrong model. These are the ones we see most often.
Choosing on a demo instead of your own work
A launch video shows the model at its best on a task picked to impress. Your job has your prices, your customers and your awkward cases. Only a test on those tells you anything.
Paying for three subscriptions at once
Owners often keep ChatGPT, Claude and Gemini running side by side "to compare" for months. Run the test for an afternoon, keep one, cancel the rest. One business plan set up properly beats three used half-heartedly.
Pasting customer details into a personal account
The model you pick matters less than the account you use it on. A personal free account is not the place for customer emails. Check what your provider says about training on your data in its own words, for example Microsoft's page on Copilot data, privacy and security.
Expecting the model to know your business
No model knows your call-out charge or your returns policy until you give it them. When it guesses, that is not a weak model; it is a missing fact. Our post on why AI makes things up shows how to close that gap.
When should you switch AI models?
Switch when your job keeps failing in the same way after you have fixed the instructions, not when a new model is announced. Most problems that look like a weak model turn out to be missing instructions or missing information.
Fix the set-up before you blame the model
If the AI keeps getting your prices wrong, it probably cannot see your price list. If its replies sound wrong, it probably has no examples of how you write. Give it both and try again. Switching models would carry the same gaps across. Each fix you make is worth writing down as a rule, which is the habit in every AI mistake should become a rule.
Switch for a reason you can name
Good reasons: your business moved from Microsoft to Google, a model will not connect to a tool you need, or the same written test keeps failing on one model and passing on another. Keep your instructions in your own document, not inside the app, so moving is an afternoon's work.
Connections can decide it for you
If you want the AI to read your booking system or stock list, check what it can connect to. The Model Context Protocol is an open standard that lets assistants such as Claude and ChatGPT connect to outside tools and data, so one connection can serve both. That is how an AI Assistant Connector lets customers look up your services from inside either assistant.
Ignore most launch news
A new model is worth a look when it fixes a problem you actually have. If your current set-up does the job, a better score on a test you will never run is not a reason to spend a week moving. Our post on keeping up with AI without chasing launches sets out a 30-minute monthly routine instead.
Questions
08Asked and answered
None is best for every business; the right one is usually the one closest to the tools you already use. Test two of them on the same real job before paying for either. Our post on where to start with AI helps you pick that first job.
No, one is enough for most small businesses. Pay for one business plan, set it up properly for one job, and add another only when a job needs something it cannot do. See free or paid AI for business for what a paid plan changes.
No. Switch AI models only when a job you rely on keeps failing for a reason a new model would fix, and your own saved test proves the new one does better. Moving costs a fresh round of checking, so a higher score on someone else's test is not enough. Our post on keeping up with AI without chasing launches sets out a simple routine.
For most small businesses, the set-up matters more than the model: the instructions, the facts it can read, the tools it may use and the checks on its work. The leading models are close on everyday jobs, so two businesses on the same model can get very different results. That set-up is what we mean by harness engineering.
Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train the foundation models behind Copilot. That promise covers staff signed in with their work accounts, so make sure nobody uses a personal Microsoft account for business data. Microsoft's page on Copilot data, privacy and security has the detail.
Yes, if those assistants can read accurate, current facts about you. A connector lets them look up your services, prices and availability live; that is what the AI Assistant Connector sets up, and getting found in AI answers covers the rest.
We set up whichever AI model fits the job and the tools you already pay for, inside accounts you own, so you are never tied to a tool of ours. If you run on Google Workspace that is usually Gemini; on Microsoft 365, usually Copilot. Read more about who sets it up and how it works.
This is general guidance, not legal or financial advice. For your own situation, speak to a qualified adviser.