Chatbots for Business

A bot should do a few jobs well. Start by counting what customers actually ask.

• Collect 50 real questions from email, chat and phone notes.
• Group them. In many small businesses a few common questions make up a large share of all contacts.
• Choose the jobs: opening hours, prices, booking status, delivery, returns.
• Write what the bot will not do.

Activity. Collect and group 50 questions for your business and rank the top ten.

Standard to meet. Collect fifty real questions; count the groups; rank the top ten; write a "will not do" list.

A good chat is short, clear and kind. Every path ends with an answer or a person.

• Greeting says it is a bot and what it can do in one line.
• Offer 3 to 5 buttons for common questions; allow free text as well.
• Fallback: if the bot does not understand twice, offer a human option and take the name and contact.
• Keep each reply under 40 words.

Activity. Write the greeting, five answers, the fallback and the hand-off for your bot.

Standard to meet. Write a greeting that says it is a bot; end every path in an answer or a person; keep replies under 40 words.

A bot is only as good as its answers. Write each answer once, in a plain spreadsheet, with an owner and a date.

• Columns: question, answer, source, owner, last checked.
• Use the same wording as the official price list or policy; do not invent.
• Review every month. An old answer is a wrong answer.
• Mark questions that must go to a human.

Activity. Build a knowledge base of at least 30 rows with owners and dates.

Standard to meet. Fill all columns, source every answer, and mark human-only rows.

Rule bots follow set paths and are predictable. AI-assisted bots handle free text but can be wrong. Many businesses use both.

• Rule-based: buttons and keywords, easy to test, limited.
• AI-assisted: understands free text, needs checks, may give wrong answers confidently.
• Safe design: AI may find the answer in the knowledge base but must not invent beyond it.
• Always keep a human route.

Activity. Choose a design for your case and write three reasons.

Standard to meet. Argue the choice with risk and cost, and include a human route.

People must know they are talking to a bot and how their details are used.

• Say it is a bot at the start.
• Collect only what is needed: first name and a contact detail, with consent.
• Never ask for passwords, payment cards or health details in chat.
• Say how long chats are kept and how to ask for deletion, following local law.

Activity. Write the privacy notice and a data table for your bot.

Standard to meet. Show the bot disclosure, limit the data, and state retention.

Test before launch. Write 30 questions in real customers' words, including awkward ones, and mark each answer.

• Mix: 20 normal, 5 misspelt or short, 5 outside the bot's scope.
• Mark each as right, partly right or wrong. Pass is 24 of 30 right.
• Fix the wrong ones and test again.
• Keep every version of the test set.

Activity. Run your test set twice and record the scores.

Standard to meet. Test thirty varied questions, complete two score sheets, and list the fixes.

Containment rate is the share of chats the bot resolves without a person. It turns into money saved.

• Containment = resolved by bot / all chats. Example: 620 of 1,000 chats = 62 percent.
• Cost per human chat is USD 3.50 in this example (an estimate; check your market), so 620 chats saved = USD 2,170.
• Also track hand-off quality: did the person receive a useful summary?
• Track customer rating after each chat.

Activity. Calculate containment and saving from sample figures and write the monthly report.

Standard to meet. Correct arithmetic with units; report names the top missed questions.

A bot is never finished. Review missed questions weekly and add or fix answers.

• Read ten failed chats each week.
• Add answers for repeated misses; retire wrong ones.
• Announce changes to staff so they know what the bot says.
• Keep an improvement log with date, change and result.

Activity. Write a launch checklist and a four-week improvement plan.

Standard to meet. Complete the checklist; name owners and dates in the plan.

On the slide: L1 Basic · What a chatbot is

Notes: Explain what a chatbot does and that it must say it is a bot. Notes: try two chatbots and list what each does well and badly.

On the slide: L2 Awareness · Top ten questions

Notes: List the ten most common customer questions and a clear answer for each. Notes: use real enquiries from the business; check that no answer asks for passwords or card details.

On the slide: L3 Foundation · Build a simple bot

Notes: Build a rule-based bot with a hand-over to a person. Notes: hand over when the bot fails twice, or for complaints or sensitive matters.

On the slide: L4 Practitioner · Test with 30 questions

Notes: Test the bot with 30 questions and report the pass rate. Notes: the course standard is at least 24 of 30; learners log each failure and fix.

On the slide: L5 Specialist · Measure value

Notes: Measure the share of chats resolved without a person and the money saved. Notes: 620 of 1,000 chats is 62 per cent, and 620 chats at USD 3.50 each saves USD 2,170; learners redo it with their own figures.

On the slide: L6 Advanced · Run a weekly review

Notes: Run a weekly review of failed chats and update the knowledge base. Notes: sample ten failed chats, add or fix answers and record the change.

On the slide: L7 Super expert · Plan a support service

Notes: Design a bot and human support service with privacy rules, targets and audit. Notes: review disclosure, data handling and hand-over; present a 12-month plan with targets.