AI

Where AI should never be allowed to decide

Widget studio: conversation analytics and the chat widget's channels
Widget studio: the chat widget and its conversations.
Shared inbox: conversations from web chat and other channels, with status and age
Shared inbox: every conversation in one queue, with its status and age.
Knowledge base: coverage checklist of the answers the assistant needs
Knowledge base: the answers the assistant can give, and the gaps.
Workflows: automated steps that run when a condition is met
Workflows: automated steps that run when a condition is met.
26 September 2026 · 5 min read

The line

An automated system may act where being wrong is cheap and reversible. It may only draft where being wrong is expensive or public.

That is the whole principle. Everything else is applying it.

Act freely

  • Answering a question you have already written the answer to.
  • Offering a time that is genuinely free in the calendar.
  • Sending a confirmation or a reminder.
  • Filing, tagging, routing and enriching records.

All reversible, all cheap when wrong, all constant.

Draft only

  • Prices beyond what you have published. A quote is an offer.
  • Anything clinical, legal or financial. Not a boundary to negotiate.
  • Public replies in your name. A response to a bad review is read by
  • hundreds of people and cannot be unsaid.
  • Commitments about timescales. "We can have someone there Tuesday" is a
  • promise somebody has to keep.
  • Spending money. Proposing a budget shift is useful. Moving it is not.

Refuse entirely

Some things should not be drafted at all, because having a draft makes sending it too easy:

  • Anything that would misrepresent who the customer is talking to.
  • Anything that fabricates social proof — reviews, testimonials, case studies.
  • Anything designed to make it harder to leave, unsubscribe or complain.

Who is accountable

The answer has to be a person, and it has to be decided before anything goes wrong rather than afterwards.

If an automated reply quotes the wrong price, somebody in your business owns the consequence. That is true whatever the vendor's terms say, because the customer dealt with you. Knowing in advance who reviews what, and who answers for it, is the difference between a mistake and a crisis.

In practice this means three things written down: what the system may do alone, who reviews what it drafts, and how quickly somebody is told when it refuses.

Review what it refused, not just what it sent

Most businesses check the messages that went out. The more useful review is the opposite one.

Every time the system stopped, it believed the question was beyond its boundary. Reading those weekly tells you two things: whether the boundary is in the right place, and what your customers are actually asking that you have never written an answer to.

That second list is usually the most valuable document in the business, and it only exists if the system is honest about what it did not know.

Make the boundary yours

The important part is that these are your settings, not a vendor's defaults. You should be able to state the ceiling above which a quote needs you, and change it without asking anyone.

And when the system stops at a boundary, it should show you what it *would* have said. That is how you learn whether the boundary is in the right place — and most businesses find they move it over time in both directions.

Find out which channel fills your diary

See which channel books your diary, and what each booking costs you.

Where to Draw the Line With Business AI