How to Use ChatGPT as a Business Adviser Without Being Misled

Give ChatGPT your full context, ask it for options and trade-offs rather than a verdict, make it argue against its own first answer, and check every number and fact it gives you. Take anything with legal, tax or regulatory consequences to a qualified adviser. Treat it as a sparring partner that's often right, never as an authority.

The danger isn't that ChatGPT is foolish. It's that it's agreeable, confident and fluent all at once, which is exactly the combination that makes a weak plan sound strong. The steps below counter each of those habits, with prompts to copy and a worked example of a decision that looked very different after the second round of questions.

Follow me on Instagram@sagnikteaches

Five ways ChatGPT misleads business owners

1. It leans towards agreeing with you. In April 2025 OpenAI rolled back an update to GPT-4o in ChatGPT because it had become noticeably more sycophantic: overly flattering and agreeable. OpenAI has worked on this since, but the pull remains. If your question says "I think this is a great idea", the answer tends to find reasons it is.
2. It states guesses as facts. Market sizes, "typical" uptake rates, rules and competitor details can all be invented with complete confidence. The explainer on AI hallucinations covers why.
3. It fills missing context with averages. Ask about "a small clinic" and you get advice for an average clinic, which may be wrong for yours in every way that matters.
4. It can be out of date. Prices, rules and product features change faster than models are updated, and a confident answer may describe how things were. See catching outdated information in AI answers.
5. It slips on money maths. Margins, percentages of percentages and multi-month projections are where small arithmetic errors grow into wrong decisions.

There's a sixth, less obvious one: ask the same question twice and you may get two different recommendations. That isn't a fault you can fix with a setting, and why AI gives different answers each time explains what to do about it. For decisions, it's actually a useful test. If the advice flips between runs, the question is closer than it looks.

Connect on LinkedInSagnik Bhattacharya

What it's good for, and where it's weakest

Used well, ChatGPT is a genuinely useful thinking partner for an owner who has nobody else to talk a decision through with. It's strong at structuring a messy problem, listing options you hadn't considered, explaining a financial or legal term in plain words, drafting a difficult letter such as a price rise, role-playing a customer's objection so you can practise your answer, and turning a vague goal into a 90-day plan with steps.

Subscribe on YouTube@codingliquids

It's weakest exactly where owners most want help: predicting demand for your particular business, telling you what your own customers will pay, and giving you the specific rules that apply to you. Those answers depend on facts it doesn't have. It will still produce them, fluently, if you ask. Keep that split in mind and most of the risk goes away: use it to think, and use your own evidence to decide.

Step 1: Give it your business, once, in a project

ChatGPT's Projects keep related chats, files and instructions together, and chats inside a project inherit its instructions and files. That makes a project the natural home for "my business adviser". Set it up once with a context document and an instruction on how you want to be advised.

ABOUT MY BUSINESS
What we do:          [2 sentences]
Customers:           [who, how many, how they find us]
Numbers:             [revenue, margin, main costs, team size]
Constraints:         [cash, time, premises, skills we lack]
Goals this year:     [2-3, with numbers]
What we've tried:    [and what happened]
My appetite for risk: [low / medium / high, and why]

HOW TO ADVISE ME (project instructions)
- Be direct. If the evidence points against my idea, say so first.
- Separate what you know from what you're assuming. Label assumptions.
- Never give a statistic without saying where it comes from. If you
  don't have a source, say "I don't have a reliable figure".
- Ask me clarifying questions before recommending anything.
- Point out when a question needs an accountant, solicitor or
  regulator rather than you.

The context document is what stops the "average business" answer. A two-van removals firm asking "How do I get more customers?" with no context got the usual list back: a better website, social media posts, asking for reviews, a referral scheme. The same question inside a project whose numbers line read "70% of jobs come from two estate agents; Saturdays fully booked six weeks ahead; weekday vans used about 40% of the time" produced a different conversation (illustrative). The reply said the firm didn't have a customer shortage, it had a weekday shortage, and that relying on two agents for most of its work was the bigger risk. It suggested a weekday discount for flexible move dates, approaching a third agent, and courting office and storage clearances, which usually happen on weekdays. None of that could come from the first version, because none of the facts it rests on were in the question.

Keep customer and staff personal data out of it; totals are enough. For sensitive figures, use ChatGPT Business, which doesn't use business data for training by default, or switch off the model-training setting in a personal account's privacy settings. Projects can also use project-only memory, which keeps what it learns in one project from leaking into your other chats; the tutorial on ChatGPT's memory settings explains the options.

Step 2: Ask for options, not a verdict

The single biggest improvement is how you frame the question. "Should I do X? I think it would work" invites agreement. Asking for a comparison of options invites analysis.

I'm deciding between these options:
A) [option]
B) [option]
C) Do nothing for now.

Before you answer, ask me up to 5 questions you need answered.
Then, for each option, tell me:
- what would have to be true for it to work
- the main risks and how likely you think each is
- what it would cost me in money and time
- what I should measure in the first 90 days to know if it's working
Do not recommend one yet.

"Do nothing" belongs on every list. It's often the right answer, and a model asked to choose between two actions will never suggest it.

What the reframing looks like for a two-groomer dog-grooming salon that's booked up most weeks. The leading version: "We're turning people away, so I'm thinking of hiring a third groomer. Makes sense, right?" The options version:

I'm deciding between these options:
A) Hire a third groomer, 30 hours a week, from January.
B) Raise prices by $8 a groom and open two extra hours on Saturdays.
C) Do nothing for now.

In an illustrative run, the questions that came back before any analysis included "How many enquiries do you turn away a week, and how do you know?" and "Is your limit groomers, or tables and drying space?" The second one mattered. The salon had two drying stations, so a third groomer would have spent much of the day waiting for one. The leading version never got near that, because it was answering a staffing question it had been handed.

Step 3: Make it argue with itself

Once it has laid out the options, push back with a set of follow-ups. These work best in a fresh chat, or in a different assistant altogether, so the model isn't defending what it just wrote.

1. Make the strongest possible case for the option you rated worst.
2. Assume I chose [option] and it failed within a year. What are the
   three most likely reasons, given what you know about my business?
3. What evidence would change your view? What should I go and find out
   before deciding?
4. Rate your confidence from 1 to 5 in each claim you made above.
   Mark anything you were guessing.

Question 3 turns the conversation from opinion into homework. Question 4 is revealing: the confident-sounding claims are often the ones it marks as guesses.

For the grooming salon, part of the answer to question 4 read like this (illustrative):

"Option B would lose fewer than 1 in 10 regular clients"   2/5 - a general
                                                            pattern, not your data
"A new groomer takes 3-4 months to build a full diary"      2/5 - estimate
"Drying space, not staff, is your current limit"            4/5 - from your answers

The 4/5 claim came from the owner's own answers, so it stood. The two 2/5 claims had been stated flatly earlier in the chat, with no hint of doubt. Each became a piece of homework. The owner checked what happened to rebookings after the last price rise in the booking system's client report, and asked the groomer who had joined two years earlier how long her diary took to fill. Those two answers, not the model's estimates, went into the decision.

Step 4: Check what it tells you

Turning on search helps, because the answer then cites web pages. It doesn't remove the need to open them. A citation shows where the model looked, not that the page says what the model claims.

A typical case: an owner comparing booking systems asked with search on and was told one option "costs $29 a month and includes SMS reminders", with a link to the vendor's pricing page. Both parts were on the page, but not together. The $29 was the annual-billing price of the smallest plan, and SMS reminders started two plans up. The model had merged two true lines into one false one. Reading the pricing table took two minutes and changed the comparison, because reminders were the reason for switching in the first place.

The calculation row catches more owners than any other, and one slip comes up again and again. Ask "I pay $6 for a product and want a 40% margin. What should I charge?" and you can get back $8.40, explained with confidence. That's a 40% mark-up on cost. A 40% margin means the profit is 40% of the selling price, which needs $10: $4 profit on a $10 sale. If the owner of a gift shop applied the $8.40 logic across a whole price list, every item would be priced 16% below target, and the answer would still look perfectly reasonable on screen. The test is to put the answer back into your own spreadsheet: (price minus cost) divided by price should give the margin you asked for.

Worked example: a hearing-aid practice weighing an aftercare plan

An illustration: the owner of an independent hearing-aid practice with about 1,100 active patients is considering a $15 a month aftercare plan covering batteries, cleaning and an annual check. The practice is hypothetical; the pattern is a common one.

First attempt. The owner asks: "My patients love our service, so I think most would sign up to a $15 monthly aftercare plan. Good idea?" The reply is enthusiastic, suggests a majority might join, and projects a healthy monthly income. Nothing in it is sourced. The question had supplied the conclusion.

Second attempt, using the steps. The project context includes a detail the first question didn't: about 40% of the practice's patients wear rechargeable aids. The options prompt compares a subscription, pay-as-you-go aftercare bundles, and doing nothing. The challenge questions bring up three things the first answer missed:

• Patients with rechargeable aids get little from battery cover, so the plan is worth much less to 40% of the list.
• The plan would replace some battery and cleaning income the practice already earns, so the gain is smaller than the headline figure.
• Recurring billing adds admin: failed payments, cancellations and questions every month.

Checking. The uptake estimate had no source, so the owner asks 50 patients at appointments over a fortnight. Fourteen say they'd be interested, 28%. At 28% of 1,100 patients, that's about 308 plans and $4,620 a month before costs. But after subtracting the battery and cleaning income those patients already pay, and the cost of supplies, the owner's spreadsheet shows the real gain is well under half that.

Decision. A six-month pilot of the plan for non-rechargeable wearers only, with take-up and cancellations tracked monthly. That's a smaller, safer decision than the one the first chat encouraged, and it rests on the owner's own numbers. For the next step, scenario planning with AI shows how to build best, worst and likely versions of a pilot like this.

Keep a decision log so you don't get talked round

A long chat has a way of moving you. By the fortieth message, you can end up agreeing to something you'd have rejected at the start, simply because each step seemed reasonable. A short written log, kept outside the chat, protects against that. It also gives you something to review later, which is how you learn whether the AI's advice has been any good for your business.

DECISION LOG
Decision:           [one line]            Date: [date]
Options considered: A / B / C (do nothing)
What AI suggested:  [one line, and how confident it said it was]
What I checked:     [sources opened, people asked, sums redone]
What I decided:     [and why, in my own words]
What would change my mind: [the evidence or number]
Review on:          [date]

Here is the log the hearing-aid practice owner from the worked example might have written, filled in:

DECISION LOG
Decision:           Aftercare plan: launch or not        Date: 6 Oct
Options considered: A subscription / B pay-as-you-go bundles / C nothing
What AI suggested:  first chat: launch to everyone (no confidence given);
                    second round: pilot, uptake unknown (rated 2/5)
What I checked:     asked 50 patients (14 interested); battery and
                    cleaning income per patient from the till report;
                    supply costs from our distributor's price list
What I decided:     pilot for non-rechargeable wearers only, because
                    the rechargeable 40% get too little from it and
                    I'd be selling them something they don't need
What would change my mind: pilot take-up under 10% by month 3, or
                    cancellations above 3 a month
Review on:          6 Jan, then 6 Apr

The line "in my own words" matters. If you can't write the reason without copying the chat, you haven't decided yet; the chat has. Review the log on the date you set, and note where the advice held up and where it didn't. After a handful of decisions, you'll know which kinds of question to trust it with.

Where ChatGPT stops and a professional starts

Use ChatGPT to prepare for these conversations, not to replace them:

• Tax, company structure and anything that changes how profit is taxed.
• Employment problems: disciplinary issues, contracts, redundancies.
• Leases, contracts and personal guarantees.
• Regulated activity, such as health, finance or anything licensed.
• Insurance, and what your cover actually includes.

What it's good at here is making the paid hour count. Ask it to summarise your situation in a page, list the questions you should ask, and explain the terms you'll hear. You'll arrive better prepared and leave with clearer answers.

A sole trader thinking about incorporating the business might ask: "Summarise my situation for my accountant in under a page, then list the questions I should ask. Don't give me any tax figures." Part of an illustrative reply:

QUESTIONS FOR YOUR ACCOUNTANT
1. At my current profit of about $58,000, would a company change
   what I pay overall, once your fees and the extra filing are included?
2. How would I pay myself from the company, and what does that mean
   for my personal borrowing, such as a mortgage application?
3. What happens to the van and tools I already own?
4. You'll save on tax once profits pass a certain level, so ask
   when that point is reached.

Questions 1 to 3 are good, and the second is one owners often forget until a lender asks for two years of proof of personal income. Question 4 is the model slipping back into advice: it asserts a saving as fact and then turns it into a question. Rewrite it as "Is there a profit level at which a company becomes worthwhile for me, and are we there?" The accountant's answer is the one that counts; the chat's job was to get you to the meeting with the right list.

Settings that make it a better adviser

• Custom instructions. Put "be direct; disagree with me when the evidence does; say when you're unsure" in your personalisation settings, so it applies to every chat, not just the adviser project.
• Memory. Memory can make it remember an assumption you've since dropped. A café owner who mentioned in spring that the café closed on Mondays kept getting staffing advice built around a six-day week in autumn, months after Monday opening had started, until the memory entry was deleted. For a fresh view on a decision, use a temporary chat or project-only memory, and update the project's context document whenever a fact in it changes.
• Plan. Plus at $20 a month is plenty for occasional decisions. If you'll share real financial figures regularly, the choice between Plus and Business matters more than the choice of model; ChatGPT Plus vs ChatGPT Business compares them.

Using ChatGPT for business advice: common questions

Is ChatGPT better than Claude or Gemini as a business adviser?

For talking through a decision, the leading assistants are close, and each has habits you'll learn. The bigger gain comes from using a second one to criticise the first one's answer, because a model defending its own earlier reply tends to hold its position. If you pay for one, try the free tier of another for the challenge step.

Can I trust ChatGPT's deep research reports?

Trust them as a reading list, not as conclusions. Research modes search the web and cite sources, which is a big improvement, but summaries can still overstate what a source says or lean on weak ones. Open the sources behind any claim you'll act on, and check the date of each one.

Do I need ChatGPT Pro for business advice?

Usually not. Plus at $20 a month is enough for most owners thinking through decisions. Pro, at $100 or $200 a month, buys more usage, which matters for heavy daily work rather than occasional advice. If you'll share sensitive figures, ChatGPT Business, which doesn't train on business data by default, matters more than a bigger personal plan.

Further reads

• How to Run a SWOT Analysis of Your Business With AI — A structured way to use AI on your whole business.
• How to Write a Business Plan With AI, and What to Check Yourself — The same discipline applied to a full plan.
• Is ChatGPT Safe for Business Use? Risks, Settings and Plan Choice — Risks and settings before you share real figures.
• Perplexity vs ChatGPT for Business Research — When a research tool beats a chat for facts.
• ChatGPT Prompts for Small Business Owners: 50 Tested Examples — More tested prompts for everyday decisions.
• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A fast routine for checking any AI answer.
• How to Build an Investor Pitch Deck With AI, and What to Check — Which parts of an investor deck to hand to AI and which to keep, slide by slide, plus the checks that catch invented market figures and mismatched numbers.
• How Consultants Use AI for Client Research Before Discovery Calls — A research routine for discovery calls: scale the prep to the deal, make AI cite everything, check what matters and turn findings into sharper questions.
• How Accurate Is ChatGPT? What Owners Should Expect by Task — Where ChatGPT is dependable, where it guesses, and a 20-case test that measures its accuracy on your own work before you trust it with customers.
• What Is the Cheapest Way to Get Expert AI Advice? — Free help pages, forums, adoption kits and discovery calls answer most AI questions. When paid advice is worth it, and how to compare quotes.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI, 'Sycophancy in GPT-4o: What happened and what we're doing about it' (April 2025); OpenAI help centre on Projects in ChatGPT.