Chatbots and human support: how UK stores should split the work

The framing of chatbot against human is wrong in the same way that comparing a search box to a shop assistant is wrong. They are not competing for the same job. The useful question is which enquiries have exactly one correct answer, because those are the ones software can handle and the rest are not close.

Questions with one answer

On most stores these are the large majority of contacts by volume and almost none of them by difficulty:

• Where is my order, and when does it arrive?
• What is the returns window and how do I start one?
• Do you deliver to this postcode, and what does it cost?
• Is this in stock in this size?
• What are your opening hours?

Every one of those has a factual answer sitting in a system you already run. Answering them at two in the morning without waking anybody is the real case for automation, and it is a good one. It is not a case for AI: a decision tree wired to your order data answers all five, deterministically, and cannot invent anything.

Questions with no single answer

The remainder share one property: the right response depends on judgement about this specific customer.

• A complaint, where the outcome is a decision about a refund or a replacement.
• A pre-sale question where the customer is choosing between two products and wants advice.
• Anything involving a distressed or angry customer, where the objective is to keep them rather than to close the ticket.
• Anything unusual enough that the answer is not written down anywhere.

A bot handling these produces the failure mode everyone has experienced: a loop that will not connect you to a person. That costs more than the salary it saved, and it does the damage to the customers who were most engaged.

The escape hatch is the whole design

One rule covers most bad chatbot experiences. There must be a visible route to a human on every screen, and asking for one must work immediately rather than after three more clarifying questions.

Two more that follow from it: hand over the transcript so the customer does not repeat themselves, and if nobody is available, say so and take a message rather than pretending to be a queue.

What a UK store has to get right legally

Chat is a data collection surface and gets treated as one. Points worth raising with whoever handles your compliance:

• Transcripts are personal data. They routinely contain names, addresses and order details, so they need a lawful basis, a retention period and a place in your privacy notice.
• Third-party widgets process data on your behalf. If the vendor is outside the UK, that is an international transfer and needs the right contractual terms.
• Do not use a chat widget to collect marketing consent by implication. Answering a support question is not agreement to receive email.
• Tell people when they are talking to software. In the UK this is an expectation rather than a specific statute. If you sell into the EU, transparency rules for AI systems that interact with people apply there, so take advice on where you stand rather than assuming the UK position covers you.

Choosing a tool

Tidio, Intercom, Drift and ManyChat all appear on UK shortlists and differ mostly in what they are really for: Intercom is a support desk, Drift leans to sales conversations, ManyChat lives on messaging platforms, Tidio targets smaller stores. Judge them on four things.

1. Does it read your real order data, or does it only answer from a script?
2. How fast is the handover to a person, and does the transcript go with it?
3. Where is the data stored, and what are the retention controls?
4. What does the widget cost you in page weight? Chat widgets are among the heaviest third-party scripts on a typical store, and Interaction to Next Paint is measured on the page they load into.

How to decide whether it worked

Not by conversion rate, which moves for a dozen reasons at once and will happily give you a number that flatters the decision. Measure the things the bot is actually responsible for: how many contacts it resolved without a handover, how long people waited for a person when they asked for one, and whether the same question keeps arriving. The third is the most useful, because a question asked repeatedly is usually a missing page rather than a support problem.