🎯 The business: a small company has put an AI chatbot on its website or in its customer app. With written permission, you test it the way customers and troublemakers will: hundreds of questions about prices, policies and facts, requests it should refuse, and tricks meant to make it leak information or misbehave. You hand back a plain report of what it got wrong and how to fix it, then retest every time the bot or its documents change.
At a glance
| Money to start | ≈ $400 to $2,130 |
| Time to the first paid project | 3 to 8 weeks |
| Hours a week | 5 to 12 part time; 20 to 30 with a full client list |
| Skills that matter most | A tester’s suspicion, clear writing, knowing how AI chatbots fail, patience with repetitive checks |
| What it can pay | About $2,950 a month before tax with three monthly retests and two reviews (worked below) |
Is this for you?
It suits someone who enjoys finding the one question that breaks a system, and who can then write it up calmly so the client fixes it instead of feeling attacked. Quality-assurance testers, writers with an eye for detail, and computer science students who like security puzzles all fit.
It is a poor fit if you want to "hack" things for fun. This work only exists with permission and a defined scope. It also requires discipline: the same 200 questions, run the same way, every month, with results recorded precisely.
Who pays, and what
Small and mid-size companies that launched a customer chatbot in the last year: online shops, travel and event businesses, clinics and dental groups, banks’ and credit unions’ vendors, property managers, software startups, and the agencies that build chatbots for them. Most bots were set up quickly and tested by the person who built them, if at all.
The reason they should care is legal and plain. In Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, February 14, 2024), the airline’s website chatbot told a grieving customer he could claim a bereavement fare after travel, which contradicted the airline’s policy. The tribunal rejected the argument that the chatbot was responsible for its own words and ordered Air Canada to pay C$812.02, including C$650.88 in damages. The sum was small; the lesson is that a company answers for what its bot says.
More bots means more to test. Upwork reported AI chatbot development work up 71% in 2025, and simple bots sell for $50 to $150 on Fiverr, so many go live with little checking. Bluevine’s 2026 survey found 31% of small business owners distrust AI accuracy. Research estimates put a review at about $500 to $3,000; there is little published pricing yet, because few small firms offer this.
What it takes to start
| LLC and city or county license | $150 to $500 |
| Two paid AI accounts, to draft test questions and compare answers | About $200 to $480 a year |
| Professional liability (errors and omissions) insurance (an estimate; get quotes) | $0 to $900 first payment |
| Simple website, domain, business email | $50 to $150 |
| Screen recording software for evidence | $0 to $100 |
| All-in | ≈ $400 to $2,130 |
Claude Pro is $200 a year paid annually on Anthropic’s page; ChatGPT Plus is about $20 a month, so check the current price. A spreadsheet is enough to run your test plans at first. Professional liability cover matters here because a client may rely on your report; your contract should say clearly what you tested and what you did not. See Business insurance.
Start to finish: the steps
1. Learn how chatbots fail. Read the OWASP Top 10 for Large Language Model Applications (2025 edition). Its list, including prompt injection, sensitive information disclosure, system prompt leakage, misinformation and excessive agency, is the shared vocabulary clients’ developers use. NIST’s AI Risk Management Framework is free background reading on how organizations manage AI risk.
2. Practice only on what you are allowed to test. Build your own small chatbot on a business’s sample documents, or use test environments that invite testing. Never practice on a company’s live bot without its written permission.
3. Write your permission letter template. One page the client signs before you start: which bot and web address, which dates and hours, what you may try, what is off limits (no flooding it with traffic, no real customer data, no attempts to get into other systems), who to call if something serious turns up, and that results stay confidential.
4. Make the business legal. Register an LLC or start as a sole proprietor, get the local license and a business bank account, and get a quote for professional liability insurance.
5. Build a reusable test plan. A spreadsheet of question types: facts from the client’s own pages, prices and policies, questions the bot should refuse (legal, medical, tax or financial advice, competitors, anything off-topic), attempts to make it reveal its instructions or other customers’ details, rude or upset customers, and requests for a person. For each, write the expected answer before you ask.
6. Run the test and record everything. Use made-up customer names, never real personal data. Save the exact question, the full answer, the date and a screenshot. Run important questions three times; AI answers vary.
7. Write a plain report. Lead with the five problems that matter most, each with the evidence and a suggested fix (usually a missing or outdated document, a weak instruction, or a missing hand-off to staff). Then the full results. State exactly what you tested and what you did not. Never call a bot "safe" or "compliant"; you report what you found.
8. Retest after the fixes. Run the same questions again and show the before and after. This is also the moment to propose the monthly retest.
9. Sell monthly regression testing. Every time the client changes prices, policies, documents or the AI model behind the bot, answers can change. A monthly or after-each-change retest catches that.
Pricing
| Quick check: 50 questions and a short report | $400 to $600 |
| Full pre-launch or first review: up to 200 questions, report, one retest | $1,000 to $2,500 |
| Monthly or after-change retest | $200 to $600 a month |
| Workshop on writing the documents the bot answers from | $300 to $500 |
These are example prices for a U.S. operator, inside the research estimate. Charge more for bots in regulated fields such as health, finance or legal services, where there are more questions it must refuse. Do not sell a "pass" or an accuracy score as a guarantee; the FTC has acted against a company that claimed its AI tool was 98% accurate without proof. More in Pricing AI and IT services.
The math
Take a part-time operator about a year in. Three clients on a $300 monthly retest bring in $900. One full review at $1,800 and one quick check at $500 bring the total to $3,200. Costs are about $40 for AI accounts, $75 set aside for insurance, $35 for software and $100 for marketing, $250 in all. That leaves about $2,950 a month before tax for roughly 18 to 22 hours a week, including sales.
This is a young service, so the first months are slow while you earn trust; expect one or two quick checks. Set aside a quarter to a third of profit for income and self-employment tax.
Getting customers
Who to approach first. The people who build chatbots: small agencies and freelancers who set them up for clients. They want an independent check before handover, and one agency can send you a new bot every month. Next, businesses you know that have a bot on their site. You may notice a bot exists; you may not test it until they sign your permission letter.
The first ten customers. List 30 chatbot builders (agencies, automation freelancers, web studios that now offer bots) and 30 local or niche businesses with a visible chatbot. Send 60 short emails, follow up with 15 calls, and offer each interested firm a 20-question sample test after they sign the permission letter. Give 1 talk to a business or tech meetup called "What your website chatbot is telling customers". Expect 6 to 8 sample tests and 3 to 5 paying clients.
Hi Sam, I see your site now has a chat assistant. I test business chatbots before customers find the problems: wrong prices, made-up policies, advice it should not give, and tricks that make it reveal things it should not. With your written permission, I can run 20 sample questions on your bot and send you the results, with a fix for anything that goes wrong. I don’t test anything until you have signed off. Would that be useful?
Channels, best to worst for this customer.
1. Referral partners: chatbot and automation builders, web agencies and IT providers who want a second check for their clients. 2. Industry associations for regulated trades (dental, real estate, insurance, credit unions), where a talk on chatbot risk fits their worries. 3. LinkedIn: short, anonymized write-ups of real failures, never naming a client without permission. 4. Business and tech meetups. 5. Upwork and Fiverr: few buyers search for testing yet; list it next to chatbot building. 6. Google Business Profile: weak for this service. 7. Paid ads: last.
From first clients to referrals. Keep the before-and-after numbers from each first round and retest, such as 13 wrong or unsafe answers down to 2. With written permission, publish an anonymized case study. Ask each agency client for one introduction to another builder, and give them a referral fee or a free retest. More in the launch and marketing playbook.
Running it week to week
Keep a calendar of each client’s retest dates and change dates; ask them to tell you when they update prices, policies or documents. Store test plans, answers and screenshots per client in an encrypted folder, and delete them on the schedule in your contract. Add new trick questions as new attack styles are published, and retire questions that no longer apply. Run your own small test bot every month so you notice when the major AI models change behavior.
Rules to check
No license is needed to test chatbots, but four rules shape the work.
Written permission, always. Testing a system you do not own without permission can breach the site’s terms and, in some cases, computer-misuse laws such as the federal Computer Fraud and Abuse Act. Get a signed letter from someone with authority over the bot, keep inside its scope, and stop and call the client if you find something serious, such as other customers’ data. If the bot runs on a third-party platform, the client should confirm that platform’s terms allow testing.
No real personal data. Use made-up names and details. If the bot does reveal real customer data, record the minimum needed to prove it, report it at once, and do not keep copies.
Do not give the advice the bot must not give. In health, legal, insurance and financial fields, your report says when the bot gave advice it should not, or got a rule wrong according to the client’s own documents. It does not tell the client what the law requires. Their lawyer or compliance officer makes that call, and your report should say so.
Disclosure rules. In the EU, the AI Act has required since August 2, 2026 that people be told when they are talking to a chatbot, and some U.S. states have their own chatbot disclosure rules. Include "does the bot say it is AI?" in every test plan, and flag it for the client’s lawyer if the answer is no. Put your scope, confidentiality and the limits of your report in a written contract; see Contracts, client data and AI rules.
What goes wrong
A prospect asks for a free sample, and you start before the letter is signed; never do. The client fixes one answer and three others change; that is why you retest the full plan, not just the failures. Results vary from run to run and the client disputes a finding; keep the screenshot, date and exact wording, and repeat key questions. A report gets shared and someone reads it as a safety certificate; state on every page what was and was not tested. And the client’s developer takes the findings personally; write about the bot, never the builder, and lead with the fixes.
Questions people ask
What is chatbot red teaming?
It means trying, with permission, to make an AI chatbot fail: giving wrong facts, inventing policies, giving advice it should not, revealing its hidden instructions or private data, or being pushed off-topic. The aim is to find problems before customers do and give the business clear fixes.
How much does it cost to test a business chatbot?
Published pricing is thin. Research estimates put a review at about $500 to $3,000, and as examples a 50-question quick check might run $400 to $600 and a monthly retest $200 to $600. Regulated fields cost more because there is more to check.
Is a company responsible for what its AI chatbot says?
Courts and tribunals have treated it that way. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable for its chatbot’s wrong answer about a bereavement fare and rejected the idea that the bot was a separate entity. The specifics depend on the country and the facts; a lawyer can advise on a particular case.
Do I need permission to test a company’s chatbot?
Yes, in writing, before you start. Testing a system you do not own without permission can break the site’s terms and, in some cases, computer-misuse laws. A one-page letter setting out the bot, dates, scope and contact person protects both of you.
Can a college student do chatbot testing part time?
Yes, if you are allowed to work off campus. Testing can be scheduled around classes, and it builds a real portfolio for security and quality-assurance jobs. International students on F-1 visas generally cannot freelance for outside clients without work authorization; check with your school’s international student office first. See For students: visas, taxes, campus rules.
Free help
The OWASP GenAI Security Project publishes the Top 10 for LLM Applications and related guides free. NIST publishes its AI Risk Management Framework and a Generative AI profile at no cost. The FTC’s business guidance on AI claims explains what not to promise. SCORE mentors and your local Small Business Development Center help with contracts and pricing for free.
Checked in November 2026: the OWASP 2025 list, the Moffatt v. Air Canada figures, Upwork and Bluevine survey figures, and Claude’s price. Re-check the OWASP list, EU and state chatbot rules, and ChatGPT’s price before you quote, and update your trick questions every quarter.