How to Train ChatGPT on Your Business Information Safely

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Train ChatGPT on Your Business Information Safely.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Train ChatGPT on Your Business Information Safely.

You don't retrain ChatGPT; you give it your business context. Put current reference files and written rules into a ChatGPT Project (up to 25 files on Plus, 40 on Business), use ChatGPT Business, which doesn't train on your content by default, or switch off model training on a personal plan, and never upload customer personal data.

"Safely" covers three separate risks, and most guides only mention the first. Your content could be used to improve OpenAI's models, which a setting or the right plan prevents. The wrong people could see your files, through sharing or connected apps. And ChatGPT could answer confidently from an out-of-date price list, which only good housekeeping prevents. Each needs its own fix.

Follow me on Instagram@sagnikteaches

One route to skip: building a custom GPT. OpenAI is retiring them, and they stop running on 11 December 2026. Projects do the same job of pairing files with instructions, can be shared with colleagues, and aren't going away.

Connect on LinkedInSagnik Bhattacharya

Context, not training: what you're really doing

When people say they want to "train ChatGPT on the business", they almost always mean they want it to know their prices, services, policies and tone without pasting them in every time. That's context: documents and instructions ChatGPT reads alongside each question. Real training (fine-tuning a model's weights) is a developer job, costs far more and is rarely what a small business needs; RAG vs custom GPT vs fine-tuning explains the difference if a supplier is pitching it to you.

Subscribe on YouTube@codingliquids

Context has two big advantages. You can change it in a minute when a price changes, and you can see exactly what ChatGPT is working from when an answer looks wrong. Both matter more than any clever setting.

Choose where the files will live: Plus or Business

The plan decides the training default and the file limits, as of September 2026:

Free or GoPlus ($20 a month)Business ($25 a seat monthly, $20 annually)
Your content used to improve modelsOn unless you switch it offOn unless you switch it offNot by default
Files per project5 on Free, 25 on Go2540
Who's in controlEach personEach personAn admin, with company accounts and single sign-on
Minimumn/aOne personTwo seats
Other notesOpenAI says ads may appearNo adsNo ads; connected apps are on by default and admins can switch them off; no self-service data export

On a personal plan, switch training off before uploading anything: open your account menu, select Settings, then Data controls, and turn off Improve the model for everyone. OpenAI says it then stops using your new conversations for training, and the setting follows your account across devices. For a one-off sensitive question, Temporary Chat keeps the conversation out of your history, although OpenAI may keep it for up to 30 days.

A quick sum for a small team: the owner alone on Plus is $20 a month. Four staff on Business are $80 a month billed annually or $100 monthly, and even a sole trader who wants Business pays for two seats (about $40-$50 a month). What the extra buys is central control and a training-off default that doesn't depend on every person remembering a switch; ChatGPT Plus vs ChatGPT Business weighs it for teams, and stopping AI tools training on your business data covers the same switch in other tools.

Sort the salon's information before uploading

Follow one business through the setup, a dog grooming salon (illustrative), where the owner runs the front desk and books, three groomers do the work, and everyone answers customer messages at some point. They want ChatGPT to help with price enquiries, booking confirmations, aftercare advice and the questions new groomers ask about salon procedures. Before creating anything, the owner sorted the salon's documents into three groups:

Green: uploadAmber: upload after editingRed: never upload
Price list by breed and coat, datedStaff handbook (remove home addresses and pay rates first)Customer export from the booking system (names, phone numbers, addresses)
Booking, deposit and cancellation policyNotes on difficult cases, with no names or identifying detailsVet records and vaccination certificates
Matting and de-matting policySupplier price lists (check the supplier's terms allow it)Card details, bank details, passwords
Aftercare guides for common servicesAnything about a staff member's health or discipline
Message templates in the salon's own tone

The red column is where the one real mistake happened. Early on, the owner uploaded the booking system's customer export "so it knows our regulars". It contained 640 names, phone numbers and a notes column with remarks like "nervous with dryers, owner has a heart condition" about a customer. The file was deleted within the hour, and the owner replaced it with a one-line summary in the instructions: "Around 60% of bookings are repeat customers on 6-8 week cycles." That's all ChatGPT needed. If you do need to share documents with personal details, redacting personal data before sharing shows how, and classifying business data before using AI tools turns the traffic lights above into a policy.

Setting up the project in five steps

  1. Create the project and choose its memory. When you create a project you can choose default memory or project-only memory. With project-only memory, chats in the project can draw on other chats in the same project but not on anything outside it, and your saved memories from elsewhere don't shape its answers. For business work, choose project-only. You can switch it on later in the project's settings, but OpenAI says the change takes a few hours to apply, so set it at the start.
  2. Upload the green files. Plus allows 25 files per project and Business 40, uploaded up to 10 at a time. The salon needed eight. Give every file a clear name with a date, such as "2026-09 Price list by breed.pdf".
  3. Write the project instructions (below). These apply to every chat in the project.
  4. Set short personal custom instructions for style that should apply everywhere, in the personalisation section of Settings, with customisation switched on. Plus and Business allow up to 5,000 characters; Free and Go 1,500. Keep business rules in the project, not here, so they don't leak into unrelated chats.
  5. Test it with real questions before anyone relies on it.

The whole setup took the owner about two hours, most of it spent editing the staff handbook and writing the instructions. Uploading took five minutes.

Instructions that keep answers inside your policies

You help a dog grooming salon's staff answer customer messages
and staff questions.

Prices: use only the dated price list. Prices depend on breed and
coat condition. If the breed or coat isn't listed, say "we'll
confirm the price when we see your dog" and don't estimate.

Matting: follow the matting policy exactly. Never promise that a
matted coat can be saved; say the groomer will assess on the day.

Health: don't give medical advice. For skin problems, lumps,
injuries or anything a vet should see, suggest contacting their vet.

Bookings: deposits, cancellations and lateness follow the booking
policy file. Don't make exceptions; say the owner will decide.

Tone: warm, short, plain. Customer messages under 100 words,
signed "[salon name] team". Use the templates where one fits.

If you can't find the answer in the files, say so.

The last line and the "don't estimate" rule do the heavy lifting. Without them, ChatGPT fills gaps with reasonable-sounding answers, which is exactly what a customer later quotes back at you.

Testing the salon's project on real questions

The owner took twelve messages from the previous fortnight and asked each one inside the project. Here's one that needed a fix. A customer wrote: "My cockapoo is quite matted behind the ears and under the legs. How much for a full groom, and can you keep her coat long?"

The first illustrative answer: "A full groom for a cockapoo is $75. Light matting can usually be brushed out, so we should be able to keep her coat long. We'll add a small de-matting fee of $10." Two problems. The price list gives cockapoos a range ($70-$90) by coat condition, not a single price, and the matting policy says a groomer decides on the day whether matting can be brushed out or has to be clipped short, for the dog's comfort. The answer had promised the one thing the policy forbids.

The fix was in the files, not the prompt. The price list was a scanned image inside a PDF, so ChatGPT had partly guessed; the owner re-saved it as text. And the matting rule moved into the instructions, as shown above. Re-asked, the answer became (illustrative): "Thanks for letting us know! A full groom for a cockapoo is between $70 and $90, depending on coat condition. For matting, our groomer will check her coat on the day: if it can be brushed out comfortably we'll keep it long, and if not we'll clip it shorter so she isn't uncomfortable, and we'll talk it through with you first. De-matting is charged by time, from $10." Every figure now came from the price list, and the promise had gone.

The difference from ChatGPT without the project was stark. Asked the same question in an ordinary chat, it gave generic advice about matting and a price range "typical for the industry", which is worse than useless when a customer holds you to it. Of the twelve test messages, nine were right first time, two needed file fixes and one showed a gap (the salon had no written policy on nail clipping for anxious dogs), so the owner wrote one.

A new groomer's question

Staff questions are the other half of the project's work. A groomer in her second week asked: "A dog's just arrived and I think it has fleas. What do I do?" The staff handbook, uploaded after the pay rates and addresses were removed, has a clear procedure: stop, don't start the groom, tell the owner at the front desk, offer to rebook after flea treatment, and clean and disinfect the table and tools before the next dog.

The first illustrative answer covered all of that and added a step of its own: "You can use our flea shampoo to treat the dog before continuing." The salon doesn't do that, because treating fleas is the owner's and vet's job and the next booking can't be delayed. ChatGPT had filled a gap with something many salons do. The owner added one sentence to the handbook ("We never treat fleas in the salon; we rebook") and re-asked; the answer then matched the procedure exactly. This is the pattern to expect: most wrong answers point to a missing sentence in a file, and each fix makes every later answer better.

The salon also gave staff a simple rule for when an answer looks wrong or incomplete: don't argue with it in the chat, where the correction disappears with the conversation. Screenshot it, send it to the owner, and carry on with the procedure you know. The owner fixes the file on Friday, so the correction reaches everyone instead of living in one person's chat history.

Memory, custom instructions and what ChatGPT keeps about you

Three places can carry information between chats, and it helps to know which is which:

  • Project files and instructions: what you deliberately put there. The right home for business facts, because you can read and change them.
  • Custom instructions: your personal preferences, applied to every chat. Good for style; wrong for prices or policies.
  • Memory: what ChatGPT picks up and references from past conversations. OpenAI changed how memory works significantly in June 2026, so check what your account is set to remember in its memory settings, and review or delete saved memories periodically.

The salon's rule is simple: nothing about a customer, a staff member or money goes into memory. Anything ChatGPT should know goes into a file in the project, where the owner can see and correct it. Project-only memory helps here, because conversations inside the project don't mingle with the owner's personal chats about, say, a family holiday.

Sharing the project with staff safely

Project sharing is available on every ChatGPT plan, and a project in a Business workspace can have up to 100 collaborators. Shared projects use project-only memory, and collaborators can see the project's files, so the traffic-light sort matters even more once a project is shared. The salon shares its project with the three groomers, who can ask questions and draft messages; only the owner uploads files.

Connected apps (what ChatGPT used to call connectors) need a decision too. On Business they're switched on by default, and an admin can turn them off or limit them; on Enterprise they start switched off. A connected shared drive can be convenient, but it lets ChatGPT search wherever that account can reach, which is harder to control than eight chosen files, so the salon's admin switched off every app it didn't need. For a small team, uploaded project files are the safer starting point. Setting up separate projects so different clients' or areas' work never mixes is covered in using ChatGPT Projects to keep client work separate.

A monthly routine that keeps it current

A project is only as safe as its files are current, and it should never hold the only copy of anything: keep the master version of every file and of the instructions in your own shared drive, not least because ChatGPT Business workspaces have no self-service data export. The salon's routine takes about twenty minutes on the first Monday of each month:

  1. Replace any file that changed (prices, policies, templates), delete the old version the same day, and keep the date in the filename.
  2. Re-ask five of the original test questions and check the answers against the files.
  3. Add a line or a file for any question staff said ChatGPT couldn't answer that month.
  4. Check memory settings and the sharing list, removing anyone who has left.

After three months the salon estimated it saved about three hours a week on customer messages and new-starter questions, with no wrong prices sent to customers since the price list was re-saved as text. The safety came less from any setting than from the habit of knowing exactly what the project contains.

Questions about giving ChatGPT your business information

Can I upload our customer list so ChatGPT knows our regulars?

It's rarely worth the risk. A customer list is personal data, and the useful part (who is a regular, what they usually book) rarely needs names or phone numbers. Keep customer records in your booking system, and if ChatGPT needs patterns, give it an anonymised summary. Where personal data is involved, check your obligations under data-protection law such as the GDPR with your adviser.

Does switching off model training delete what I've already shared?

No. OpenAI says turning off Improve the model for everyone stops new conversations being used to train its models; existing chats stay in your history, archived chats stay archived and project chats stay in their projects. Delete chats or files you no longer want kept, and switch the setting off before uploading anything sensitive.

Should I build a custom GPT instead of a project?

No. OpenAI is retiring custom GPTs: they stop running on 11 December 2026 and are being moved to plugins. A project gives you the same files-plus-instructions setup, can be shared, and isn't going away. For procedures you repeat inside ChatGPT, OpenAI's plugins and skills are the replacement it recommends for GPTs.

Further reads

Sources: OpenAI help pages on projects, data controls and the Improve the model for everyone setting, custom instructions and memory (via search results); ChatGPT plan, privacy and upload facts from the verified fact sheet. All checked September 2026.

Want your ChatGPT project set up safely?

On a 1:1 call we'll sort which of your documents belong in ChatGPT, choose the right plan and settings, and write project instructions that keep answers inside your policies.

Book a 1:1 call with me