AI customer service that knows when to hand over
Good AI customer service answers your customers from your own help content, says when it does not know, and passes the conversation to a person with the details attached. I’m Mo Tawfik Khalifa, chief product officer and forward deployed engineer at Instruxi. I set up an AI help desk for a fintech platform: I bought the agent instead of building one, wrote the help centre it answers from, and held one rule, that it can read and explain and never change an account. Asked to send a customer’s cash-out to a third party, it refused and passed the conversation to a person.
Your customers are wary for good reason. In a Gartner survey of 5,728 customers in December 2023, 64% said they would prefer companies did not use AI for customer service, and 53% would consider switching to a competitor if they found out a company was going to (Gartner, July 2024). The top concern was that it would get harder to reach a person.
Worried the bot makes things up, or traps your customers?
These are the things that go wrong with support bots, in the words of the people involved, and what I do about each one.
“Hey! We have no such policy” “Unfortunately, this is an incorrect response from a front-line AI support bot.”Michael Truell, Cursor co-founder, after its support bot invented a rule, The Register, April 2025
The bot may only answer from your help articles, and every answer is tied to the line it came from, so you can check it. If no article covers the question, it says it does not know and hands over. It never fills the gap with a guess.
“It should be obvious to Air Canada that it is responsible for all the information on its website.”Tribunal member Christopher Rivers, ordering Air Canada to pay a passenger its chatbot misled, BBC, February 2024
You own what the bot says, so it only says what your published articles say. Anything that promises money, a refund, a discount or an exception goes to a person, who can decide and stand behind it.
Customers’ top concern about AI in customer service is that it will get more difficult to reach a person.Gartner survey of 5,728 customers, July 2024
Asking for a person gets a person, every time, from the first message. The handover carries a short note: what the customer asked, what the bot already answered and why it stopped, so nobody has to explain twice.
By 2027, 50% of organisations that expected to significantly reduce their customer service workforce will abandon these plans.Gartner prediction, June 2025
I plan the AI to take the repeat questions off your team, not to replace the team. We judge it on customer satisfaction and handoffs, not on headcount, and your people keep the conversations that need them.
What you get
- An agent on your help desk that answers from your articles only, with the topics that always go to a person written down.
- The setup in your own help desk accounts, which you own from day one.
- A price fixed in writing after the first call, before work starts.
- Progress on a live link you can click every week, and a written handover so your team runs it without me.
- The person you talk to is the person who sets it up.
How do I set it up, step by step?
- Read your tickets. I pull a few hundred recent conversations and sort them: which questions repeat, which need a person, and which have no help article yet.
- Write the missing articles first. The bot can only be as right as your help content. Most of the work is here.
- Buy the agent, usually. Most help desks now sell one. I pick the one that fits your desk and your volume, and say why. Building your own only pays off in a few cases, which I cover in build or buy.
- Set the rules. Answer only from articles. Topics that always go to a person: safety, complaints, money, anything legal. Read-only to start: it can look up an order and explain, and it cannot change one.
- Replay past tickets. Before any customer sees it, I run real past questions through it and check every answer against its source.
- Launch with a person watching, then read the numbers with you every week and fix the articles behind the misses.
Ask the bike shop.A made-up shop with an AI on its website chat. You are the customer: pick what to ask. Every answer shows the help article line it came from. When no line covers it, the AI says so and a person takes over.
Returns and refundshelp/returns
- You can return most items within 30 days of delivery for a full refund.
- Helmets can only be returned unworn and in the original box, because we cannot check a worn helmet for damage.
- Refunds go back to the card you paid with, within 5 working days of the item reaching us.
Deliveryhelp/delivery
- Orders placed before 2pm on a working day leave the same day.
- UK mainland delivery takes 2 to 4 working days.
- Bikes arrive almost fully built, with the tools for the last steps in the box.
E-bike batterieshelp/batteries
- Charge your battery indoors on a hard surface, and never leave it charging overnight.
- If a battery is hot, swollen or smells of burning, stop charging, unplug it if it is safe to, move away from it, and call 999 if it smokes or catches fire.
- Battery faults always go to our workshop team, never to the chat assistant alone.
A made-up shop, help centre and team. The rules are the ones I set on a real help desk.
The count of answers with no source stays at zero on purpose. That is the rule doing its job: when the AI has nothing to quote, it does not speak, it hands over.
Which numbers should you watch?
I set these up before launch and read them with you every week. None of them is a target I promise before seeing your tickets; they are the dials that tell us what to fix.
- Customer satisfaction (CSAT) on AI conversations, next to the same score for conversations your team handles. If the AI scores lower, we find out why before widening it.
- Handoff rate. The share of conversations passed to a person. Too high means missing articles; too low can mean the AI is answering things it should not.
- Resolved without coming back. Conversations where the customer did not return about the same thing within a week.
- Answers with no source. Should be zero. Any answer not tied to an article is a bug.
- Time to a person after a handoff, in and out of opening hours.
What have I built and written on AI support?
What an AI customer service agent can safely do, what it costs, and the read-only rule I held: it explains, and never changes an account.
Build or buyAn AI chatbot for your website, three waysA help desk widget, a no-code builder or a custom bot on a model API, with prices checked at the source.
GuardrailsAI agent governance, in practiceWhen the agent is allowed to act, not just answer: an owner, least privilege, a check before each action and a record.
If your agent will take actions such as refunds or address changes, the controls are the same ones I use for AI agent security.
Questions people ask
How do you stop an AI support bot making things up?
It may only answer from your help articles, and every answer is tied to the article it came from. If no article covers the question, it says it does not know and passes the conversation to a person. Before launch I replay real past tickets and check every answer against its source.
What happens when the AI does not know the answer?
It says so, and hands the conversation to a person with a short note: what the customer asked, what it already answered, and why it stopped. Your person starts with the full picture and the customer does not repeat themselves.
Can customers always reach a person?
Yes. Asking for a person gets a person, every time. Some topics never stay with the AI at all, such as safety, complaints, money and anything legal. Outside opening hours the customer is told when a person will reply.
Can the AI give refunds or change orders?
Not at first. I start it read-only: it can look up an order and explain your policy, and anything that changes money or an account goes to a person. Actions are added one at a time later, with a limit and a person approving anything over it.
Will it replace my support team?
I plan it to take the repeat questions off your team, not to replace the team. The hard conversations still need people, and the AI hands them over with the details attached so those people can spend their time on them.
How do you measure whether AI customer service works?
Customer satisfaction on AI conversations next to human ones, the share handed to a person, the share resolved without the customer coming back, and answers with no source, which should be zero. I set them up before launch and read them with you every week.
- 30-minute call. Your help desk, your volume, what customers ask most.
- A written plan in 48 hours. What to set up first, which articles are missing, the risks.
- Start, or not. Use the plan with your team or with me. I reply to every email within one working day.
If you want AI on your support without losing your customers’ trust, write to m@tawfik.tv with your help desk tool and roughly how many conversations you get a month.