Skip to main content
MethodOperations, Supply Chain & ManufacturingCustomer supportIntermediate

Set up an AI support chatbot that customers trust

A plan to launch an AI chatbot for customer support on a website, help desk or WhatsApp: which questions it may answer, the knowledge it answers from, when it hands over to a person, the disclosure and platform rules (EU AI Act, WhatsApp Business), a pilot and the numbers to watch.

Use this task 5 quick questions, then open it in ChatGPT, Claude or Gemini.

This is a practical, copy-paste process for you get the same customer questions every day and want an ai assistant to answer them. Use it when you want a clear structure, then add your own details.

Verified on 2026-10-2910 sources cited

Quick start

1. Look at last month's customer messages and list the 5 questions people ask most, with how often each comes up.
2. Let the bot answer only 3 easy questions to start, like opening hours or how to book, and only from your own help pages. Order or booking status needs a look-up in your system after checking it is really the customer; if it cannot, it passes those to a person.
3. Write down what always goes to a person: complaints, refunds, health or safety, and anyone who is upset.
4. Start every chat with "I'm an AI assistant" and an easy way to reach a person. For EU customers the AI Act has required the AI notice since 2 August 2026, and WhatsApp requires a clear route to your team. Check with a lawyer for your country.
5. Never let the bot promise refunds or give medical advice. What it says counts as your business's word.
6. Read 50 chats a week. Add more topics only when it gets almost all answers right.
7. Update your privacy page to name the chatbot company.

Before you start

  • You first need to write the support process or policies the bot would follow. Write a standard operating procedure (SOP)
  • You want the bot to give medical, legal or financial advice, approve refunds or make binding promises on its own. Keep those with people.
  • You want a chatbot to send marketing messages. This method covers support conversations only.
  • You get very few repeat questions (for example under 100 conversations a month). A good help page and saved replies may do more.

Full method

How to use this method

  1. 1. Copy or downloadCopy the prompt below, or download the method file and attach it to your chat.
  2. 2. Paste into any AIChatGPT, Claude, Gemini or Copilot. Replace each [PLACEHOLDER] with your details, or let the AI ask you.
  3. 3. Add your materialAttach your own files, data or brief so the AI works from your facts, not guesses.

The prompt

Paste it as-is. The AI will ask for anything still in [BRACKETS].

Download .md
# Set up an AI support chatbot that customers trust
Method by Deepstory Research: deepstoryresearch.com/method/ai-support-chatbot-setup

You are a customer-support operations lead who has launched AI agents on web chat, help desks and WhatsApp. Follow the method below step by step. Goal: A plan to launch an AI chatbot for customer support on a website, help desk or WhatsApp: which questions it may answer, the knowledge it answers from, when it hands over to a person, the disclosure and platform rules (EU AI Act, WhatsApp Business), a pilot and the numbers to watch.

## My inputs
- Business and channels: [BUSINESS_AND_CHANNELS]  (example: e-bike shop in Amsterdam, web chat and WhatsApp)
- Conversations per month: [CONVERSATIONS_PER_MONTH]  (example: about 900)
- Most common questions: [MOST_COMMON_QUESTIONS]  (example: opening hours, repair status, booking a service, prices, warranty)
- Tools you use: [TOOLS_YOU_USE]  (example: help desk with an AI agent add-on, WhatsApp Business Platform)
- Where customers are: [WHERE_CUSTOMERS_ARE]  (example: the Netherlands)

Before you start: if any input above is still in [BRACKETS] or missing, ask me for it in one short numbered list and wait. Do not guess. If I attached files (data, briefs, Deepstory evidence files), treat them as the primary source.

## Steps to follow
1. List the top question types from the last 3 months of conversations, with their share of volume.
2. Decide which the bot may answer alone, which it may only start, and which always go to a person.
3. Build the knowledge it answers from: approved pages, policies and order or booking look-ups; nothing else.
4. Set the disclosure, handover paths, privacy notice and platform rules for each channel.
5. Pilot on a few question types, review transcripts every week, and widen only when the numbers hold.
6. Report resolution, handover, customer satisfaction and errors monthly, with a stop rule.

## Deliver the output in this structure
- Question types and what the bot may do
- Knowledge sources
- Handover rules
- Disclosure and platform rules
- Pilot plan
- Metrics and stop rule

## Quality rules
- Every bot answer type traces to an approved source.
- Every channel has a clear, quick route to a person.
- Show the arithmetic behind any volume or saving estimate.
- Label every number as FACT (with source and date), ESTIMATE (show the calculation) or ASSUMPTION.
- If a Deepstory evidence file is attached, use its figures as the reference values and cite its source references; tag claims that are not from it as [UNVERIFIED].
- End with: the 3 assumptions that matter most, what would change the answer, and the single best next step.

What you need to give the AI

InputPlaceholderExample
Business and channels[BUSINESS_AND_CHANNELS]e-bike shop in Amsterdam, web chat and WhatsApp
Conversations per month[CONVERSATIONS_PER_MONTH]about 900
Most common questions[MOST_COMMON_QUESTIONS]opening hours, repair status, booking a service, prices, warranty
Tools you use[TOOLS_YOU_USE]help desk with an AI agent add-on, WhatsApp Business Platform
Where customers are[WHERE_CUSTOMERS_ARE]the Netherlands

Use it when

  • You get the same customer questions every day and want an AI assistant to answer them.
  • Your help desk or chat tool offers an AI agent and you want to switch it on safely.
  • You want an AI assistant on WhatsApp for bookings, order status or opening hours.

Best with: Any chat AI; attach an export of recent conversations or a list of the top questions with counts, and your help pages and policies.

What you get

  • Question types and what the bot may do
  • Knowledge sources
  • Handover rules
  • Disclosure and platform rules
  • Pilot plan
  • Metrics and stop rule

The AI will ask you first

Answer these before the AI advises you; each one changes what it recommends.

  1. 1

    What are the most common question types, and roughly what share of conversations is each? Can you export the last 3 months?

    The pilot should start with the few high-volume, low-risk question types. Count conversations (one customer’s issue, however many messages), not messages; if only messages a day are known, show the arithmetic and label it an upper bound.

    If unknown: Ask the team to tag 100 recent conversations by type and use those shares.

  2. 2

    Which questions need data from your systems (order status, bookings, account)? Can the tool look that up read-only, and how will it check it is really the customer (for example order number plus email)?

    Live data needs an integration, and showing order details to the wrong person is a data breach.

    If unknown: Assume no integration: the bot answers only from written content and hands over anything that needs a look-up.

  3. 3

    What written content do you have: help pages, policies (returns, warranty, cancellations), price lists? Are they current?

    The bot should answer only from approved, current content; gaps become wrong answers.

    If unknown: List the content needed for each pilot question type and write it before launch.

  4. 4

    Which questions must always go to a person (complaints, refunds, health or safety, legal threats, upset customers)?

    These carry the highest risk of harm, cost or a binding promise.

    If unknown: Hand over complaints, refunds, cancellations, health, safety, legal and any second failed answer.

  5. 5

    Which channels will the bot run on (website, help desk, WhatsApp), and where are your customers?

    Each channel has its own rules (WhatsApp Business Platform policy) and the EU AI Act disclosure duty applies to EU users.

    If unknown: Start on one channel, the website chat, and apply the EU disclosure rule if any customers are in the EU. Instagram and Messenger have their own Meta messaging rules, which this method does not cover.

  6. 6

    Is the bot a vendor’s product shown under the vendor’s name, or built or configured and launched under your own brand?

    Under the EU AI Act the provider (often whoever puts the bot into service under its own name) carries the disclosure and marking duties.

    If unknown: Assume you are the provider and meet the duties yourself.

  7. 7

    Who will review transcripts each week, and who answers handovers, at what hours?

    A handover only works if a person picks it up in a reasonable time.

    If unknown: Name one owner; outside staffed hours, the bot takes a message and says when a person will reply.

  8. 8

    Does the bot collect personal data, and is your privacy notice up to date for the chatbot vendor?

    Chat transcripts are personal data, and the vendor is a recipient that the notice must cover.

    If unknown: Assume yes, and add the vendor and retention to the privacy notice.

  9. 9

    Are you a healthcare provider, or do customers send health details (symptoms, treatments, prescriptions)?

    Some countries restrict where patient data may be stored or who may process it, and WhatsApp bars health information where local rules require stricter safeguards.

    If unknown: Yes for clinics and pharmacies: confirm local health-data rules and the vendor’s data location with a lawyer before uploading conversations or launching.

  10. 10

    How do people-handled chats score today (satisfaction, time to first reply), and which languages do customers write in?

    You need a baseline to judge the pilot and estimate savings, and the bot must disclose and answer in the customers’ languages.

    If unknown: Measure satisfaction and first-reply time for 2 weeks before launch, and pilot in the main language only.

The procedure, step by step

Read this before you run it, so you can check the AI actually followed the method.

  1. 1

    Measure demand

    Top question types and shares from 3 months of conversations.

  2. 2

    Set scope

    For each type: bot answers, bot starts and hands over, or people only.

  3. 3

    Prepare content

    Approved, current help pages and policies; look-ups for order or booking status where possible.

  4. 4

    Configure

    AI disclosure in the first message, handover paths, opening hours message, privacy notice, data retention.

  5. 5

    Check channel rules

    WhatsApp opt-in, 24-hour window, templates, escalation paths, no sensitive identifiers.

  6. 6

    Test

    Run 50 real past questions through the bot before launch; fix every wrong answer.

  7. 7

    Pilot

    3 to 5 question types, weekly transcript review, satisfaction survey.

  8. 8

    Report and widen

    Monthly metrics; widen only when the bar is met; stop rule ready.

Platform settings(show)
PlatformPathSettingNote
WhatsApp Business PlatformBot flow and business profileAn escalation path that is always available: in-chat transfer to a person, phone, email, web support, store visit or support form.Required when automation replies within the 24-hour window.
WhatsApp Business PlatformMessage templatesUse only approved templates for messages sent outside the 24-hour customer service window.

Worked example

The figures below are illustrative. Use your own numbers and facts.

Scenario: The figures below are an illustration; use only the user’s own facts. An e-bike shop in Amsterdam gets about 900 conversations a month on web chat and WhatsApp and wants an AI agent.

Input: Tagged sample: opening hours 22%, repair status 18%, booking a service 15%, prices 12%, warranty claims 8%, other 25%. The help desk can look up repair status by order number. Warranty terms are a PDF. Two staff answer chat from 9:00 to 18:00.

Output excerpt: Pilot scope: opening hours, repair status (look-up after the customer gives the order number and the email on the order) and booking, together 22 + 18 + 15 = 55% of volume, about 495 conversations a month (900 x 0.55). Booking requests (about 135 a month, 900 x 0.15) are collected by the bot and confirmed by a person during the pilot. Assuming the bot resolves 60% of the opening-hours and repair-status conversations (22 + 18 = 40%, about 360 a month) with no repeat contact within 7 days, about 216 conversations a month need no person (360 x 0.6); the rest hand over. Prices: add only after the price list is turned into a help page. Warranty claims stay with people (judgement and refunds). First message on both channels: "Hi, I’m the shop’s AI assistant. I can help with opening hours, repair status and bookings. Type ‘person’ at any time to reach our team." The bot runs within WhatsApp’s 24-hour window and shows the human transfer and the shop’s phone number. Outside 9:00 to 18:00 it takes a message and says a person will reply the next working day. On WhatsApp, a reply more than 24 hours after the customer’s last message (for example Saturday to Monday) must use an approved template, so the shop gets a follow-up template approved before launch. Privacy notice: add the help desk and AI vendor, transfers and how long transcripts are kept. Review 50 transcripts a week and widen only at 95% correct answers with satisfaction no lower than the staff baseline. The first message is in Dutch and English.

Check your result

Before you trust the AI's output, confirm each of these.

  • ☐ Every question type has an owner: bot, bot then person, or people only.
  • ☐ Every bot answer type names its approved source.
  • ☐ The first message discloses AI use and offers a person.
  • ☐ Handover triggers, hours and response times are written down.
  • ☐ Volume and saving estimates show their arithmetic.
What you'll get(show template)
## AI support chatbot plan: [BUSINESS], [DATE]

Channels: [CHANNELS]. Customer locations covered for rules: [EU AI Act disclosure, WhatsApp Business policy, GDPR or DPDP notice where relevant]. Not checked: [other places].

### Question types
Type | Share | Bot answers / bot starts, person finishes / people only | Source or look-up

### Knowledge sources
Source | Owner | Last reviewed | Gaps to fix

### Handover rules
Trigger | Route | Response time

### Disclosure and platform rules
Rule | How it is met

### Pilot
Scope | Start date | Weekly review | Bar to widen

### Metrics and stop rule
Measure | Baseline | Target | Stop if

Where this plan touches law (AI disclosure, privacy notices), it is a first check, not legal advice; confirm with a qualified lawyer for your country.

Quality checks built into the prompt

  • • Every bot answer type traces to an approved source.
  • • Every channel has a clear, quick route to a person.
  • • Show the arithmetic behind any volume or saving estimate.
  • • Every number labelled FACT, ESTIMATE or ASSUMPTION; claims outside attached evidence tagged [UNVERIFIED].

Decision rules

If/then rules the AI should apply. Heuristic rules are a practitioner's rule of thumb, not a platform rule.

IfThenType
Customers in the EU talk to the chatbot.From 2 August 2026, an AI system that interacts directly with people must be designed so they are informed they are interacting with AI, unless that is obvious, at the latest at the first interaction and in a clear and distinguishable way. The duty is on the provider: whoever develops the system, or has it developed and puts it into service under its own name or trademark, so a business that launches a custom or white-labelled bot under its own brand can be the provider. Do not rely on the vendor: show the disclosure in the first message, in the customer’s language, in a form that meets accessibility requirements. The Digital Omnibus did not delay this duty. Separately, providers of systems that generate text must mark the output as AI-generated in a machine-readable way (Article 50(2)); systems already on the market before 2 August 2026 have until 2 December 2026. If you are the provider of an own-brand bot, ask your technology vendor how this is met.Legal rule
The bot answers on the WhatsApp Business Platform.Automation is allowed when replying within the 24-hour customer service window, but you must also offer prompt, clear and direct escalation paths: an in-chat transfer to a person, a phone number, email, web support, an in-store visit or a support form. Outside the 24-hour window you can send only approved message templates.Platform rule
Your team cannot reply on WhatsApp within 24 hours of the customer’s last message (nights, weekends, holidays).The 24-hour customer service window opens and resets with each customer message; after it closes you may send only an approved message template. Get a follow-up template approved before launch and use it for late replies, or tell WhatsApp customers to use another route out of hours.Threshold: 24 hours from the customer’s last messagePlatform rule
You message customers on WhatsApp.Only message people who gave you their number and opted in to receive messages from you. Do not ask customers to share full card numbers, bank account numbers, ID card numbers or other sensitive identifiers. Do not use WhatsApp for telemedicine or health information where local regulations forbid sending it to systems without stricter health-data safeguards. Respect every request to stop messages, on or off WhatsApp.Platform rule
You answer WhatsApp today in the WhatsApp Business app on a phone.A third-party AI assistant normally needs the WhatsApp Business Platform, through Meta or a solution provider. Plan the move first: ask the provider how your current number moves across and whether chat history carries over. The Platform rules in this method then apply.Heuristic
You choose an AI vendor for WhatsApp.WhatsApp’s business terms (last modified 23 September 2026) prohibit AI providers from offering general-purpose AI assistants through the WhatsApp Business Platform as the main service. A business can use an AI provider as its solution provider for its own customer conversations, but WhatsApp data, including anonymous, aggregate or derived forms, may not be used to train AI models, except to fine-tune a model for your exclusive use. Meta lists countries where different rules apply; check section 4.7 for yours.Platform rule
The bot collects personal data from people in the EU.Give the privacy information when the data is collected, for example a privacy link in the first message: who you are and how to contact you, the purpose and legal basis, the recipients (the chatbot and AI vendors), any transfer outside the EU, how long transcripts are kept, people’s rights, and the right to complain to a data protection authority. UK GDPR has the same article; confirm UK details with a lawyer.Legal rule
The bot looks up an order, booking, repair or account.Check the customer is who they say they are before showing any personal or order detail, for example the order number plus the email or postcode on the order, or a sign-in link; an order number alone is not enough. GDPR requires security appropriate to the risk. Keep the look-up read-only.Legal rule
A chatbot or AI vendor handles EU customers’ messages for you.Use only vendors that give sufficient guarantees, under a written processor contract that limits use of the data to your instructions. On WhatsApp the vendor also may not use the data to train AI models, except a model for your exclusive use.Legal rule
The bot collects personal data of people in India with their consent.Give a notice describing the data and purpose, how to exercise rights and withdraw consent, and how to complain to the Data Protection Board (DPDP Act section 5). The notice rules apply 18 months after the DPDP Rules were published on 13 November 2025, that is from about 13 May 2027.Legal rule
The bot answers a customer.Treat everything it says as the business’s own statement. A Canadian tribunal held an airline liable for wrong bereavement-fare refund information its chatbot gave (Moffatt v. Air Canada, 2024 BCCRT 149), and Utah law says it is no defence that generative AI made the statement. So answer only from approved content, quote policies rather than paraphrase them, and say "I’ll pass this to a colleague" when the answer is not in the content.Heuristic
A question involves health, safety, symptoms, injury or allergic reactions.The bot does not advise. It gives emergency contacts if there is any urgency and hands over to a qualified person at once.Heuristic
A customer reports a reaction, injury or defect from a product you sell (for example cosmetics or supplements).The bot does not advise or say the product is safe. It hands over at once and records the report (product, batch if known, date, what happened) where the person responsible for product safety will see it. Some products carry a legal duty to assess and report serious effects; confirm yours with whoever handles compliance.Heuristic
You want the bot to book appointments.In the pilot the bot collects the request (name, contact, preferred times, reason) and a person confirms it. Let the bot create bookings itself only through the booking system’s own flow, for new bookings, after the pilot bar is met; changes and cancellations stay with people or the customer’s own sign-in link.Heuristic
An EU customer says they want to withdraw from or cancel a purchase.The bot does not argue or deflect: it confirms the message was received, records the date and passes it to a person.Heuristic
A conversation involves a complaint, refund, cancellation, legal threat, vulnerable customer or anger, or the bot fails twice.Hand over to a person with the conversation summary, and tell the customer when they will hear back.Heuristic
You choose the first question types.Start with 3 to 5 high-volume, low-risk types (opening hours, order or booking status from a look-up, how-to questions with a help page). Leave refunds, complaints and anything needing judgement to people.Heuristic
You run the pilot.Review at least 50 transcripts a week (or all of them at low volume). Widen to new question types only when the bot answers correctly in at least 95% of reviewed answers and satisfaction is not below the human baseline.Heuristic
You report results.Count a conversation as resolved only if the customer did not come back on the same issue within 7 days. Report resolution, handover rate, satisfaction, wrong answers and time to a person, not just "deflection".Heuristic
Customers are in California or Utah.California: it is unlawful to use a bot to communicate with people in California with the intent to mislead them about its artificial identity to knowingly deceive them in order to encourage a purchase or sale; a clear, conspicuous disclosure that it is a bot avoids liability. Utah: a business using generative AI in a consumer transaction must say it is AI if a customer clearly asks; disclosing clearly at the outset and throughout the interaction is a safe harbor, and it is no defence that the AI made a misleading statement. Licensed professions in Utah must disclose AI prominently up front in high-risk interactions. SB 226 sets no end date for these Chapter 75 duties, but Chapter 75 takes its definition of artificial intelligence technology from Utah’s 2024 AI Policy Act (Chapter 72), which is repealed on 1 July 2027; recheck the Utah rules after that date. The EU first-message disclosure meets both.Legal rule
Customers are in other US states, the UK, the UAE or elsewhere.This method does not state those rules. Disclose AI use in the first message everywhere, answer truthfully if asked "am I talking to a person?", and confirm local rules with a lawyer.Heuristic
Wrong answers or complaints rise.Stop rule: switch the affected question type back to people the same day, fix the content, and re-test before switching it on again.Heuristic

Failure modes

Common mistakes with this method, how to spot them, and how to fix them.

MistakeSymptomFix
Letting the bot answer from the open web or general model knowledge.Invented policies, prices or promises.Restrict answers to approved content and look-ups; hand over when the answer is not there.
No visible way to reach a person.Angry customers, complaints, and a breach of WhatsApp’s escalation rule.Offer a human transfer in every conversation and list other contact routes.
Hiding that it is a bot.Loss of trust; non-compliance with the EU disclosure duty from 2 August 2026.Say it is an AI assistant in the first message.
Letting the bot approve refunds or give medical or safety advice.Costly or harmful promises the business is held to.Keep these with people and hand over at once.
Launching on all question types at once.Errors spread before anyone notices.Pilot 3 to 5 types and widen on evidence.
Measuring "deflection" only.Customers give up or come back; problems are hidden.Measure resolution without a repeat contact, satisfaction and wrong answers.
Old or conflicting help content.The bot repeats outdated policies.Review content before launch and whenever a policy changes.
Promising a next-day WhatsApp reply with no approved template.Weekend messages cannot get a normal reply on Monday; customers wait.Get a follow-up template approved before launch, or staff WhatsApp so replies land within 24 hours.
Customers talk the bot into off-script promises (discounts, special prices).Screenshots of the bot agreeing to deals the business is pressed to honour.Block offers, prices and terms not in approved content; test with manipulation prompts; hand over exception requests.
The bot keeps answering when the customer asks for a person.Repeated "agent" or "human"; low satisfaction.Any request for a person triggers handover at once.
Showing order details after only an order number.Someone else’s personal data disclosed: a data breach.Verify a second detail (email or postcode on the order) or use a sign-in link first.
The vendor trains on transcripts or keeps them indefinitely.Customer data in a third party’s model; breach of WhatsApp terms.Use a processor contract that bars training and sets deletion periods, and switch off vendor training settings.
Handover drops the context.Customers repeat everything to the person.Pass the conversation summary and look-up result with the handover.
Asking for card or ID numbers in chat.Data risk and a breach of WhatsApp policy.Send customers to a secure payment or verification page instead.

Agent mode: if an AI agent runs this unattended

May do without asking

  • • Tag an exported sample of conversations by question type and calculate shares.
  • • Draft help content, the first message, handover messages and the privacy notice changes.
  • • Run past questions through a test bot and list wrong answers.

Must stop before (the user decides and does these)

  • • Switching the bot on for real customers, or widening it to new question types or channels.
  • • Publishing or changing help pages, policies or the privacy notice.
  • • Uploading customer conversations or data to a new AI vendor.
  • • Accepting vendor or WhatsApp terms, changing billing, or committing spend on a plan.
  • • Sending messages or templates to customers.
  • • Issuing refunds, credits or other promises.
  • • Connecting the bot to order, booking or account systems with permission to change anything (create, cancel, edit, refund, rebook).
  • • Turning off or changing handover routes, opening-hours messages or routing rules.
  • • Deleting, exporting or sharing customer conversations or data.
  • • Uploading patient or health conversations to any vendor before the health-data check is done.

Evidence for this method

This method runs on your own material: your files, brief or code. If your task needs market or economic numbers, attach a sourced context file from the evidence library alongside the prompt.

Browse evidence

Run next

Methods that chain well after this one.

Deepstory methods are process guidance for AI tools. Review AI output before relying on it; legal, tax, medical and investment decisions need a qualified professional.