Resources
How to do this well.
Most of what goes wrong with customer-facing AI is decided in the first week, not in the technology. This is what we have learned about the decisions that matter, written plainly and without a form in front of it.
What it knows, and where it stops
- Your website34 pages
- Policies and terms12 documents
- Written answers8 articles
- Still unanswered3 questions this week
What to put in front of customers first.
Start where you already have good answers and high volume. Ambition is cheaper after something is working.
Begin with the questions you already answer twenty times a week
Opening hours, coverage, price ranges, how your process works, what happens after someone gets in touch. You already have these answers written down somewhere, they are low risk, and they are most of the volume.
Use your real material, not a summary of it
Point it at the pages and documents your business actually publishes. Answers drawn from a rewritten summary drift away from what your customers can see on your own website.
Decide what it must never handle alone, on day one
Anything clinical, legal, financial or contractual. Write that list before you go live, not after the first uncomfortable conversation.
Put a person one sentence away
The single fastest way to lose trust is to make somebody argue with software to reach a human. Make it easy and few people will actually take it.
The goal is not to deflect customers. It is to be worth talking to.
A system that wears people down until they give up will show a wonderful deflection rate and lose you the customer.
How to tell whether it is actually working.
Deflection rate is the wrong measure. It rewards a system that frustrates people until they give up.
Resolved without a person
Conversations that ended with the customer getting what they came for, not conversations that simply ended.
Answered from your material
The proportion of answers grounded in something you published. If this is low, the knowledge is the problem, not the model.
What it could not answer
The most valuable list your business will get. Every entry is a question customers ask that your website does not answer.
Time to a person
When somebody needs a human, how long they wait. This is the number your customers actually feel.
Bookings and outcomes
Appointments made, cases resolved, enquiries qualified. The things your business would have counted anyway.
Repeat questions
The same customer asking the same thing twice is a memory failure, and it is worth watching closely.
What we would tell you to avoid.
All of these are recoverable. They are easier not to do.
Launching it everywhere at once
One site, one team, one clear set of questions. Prove it in front of real customers, then widen it. A bad first week is expensive to undo.
Making it sound like a robot to be safe
Hedged, cautious language reads as evasive. Customers trust a clear answer with an honest limit far more than a careful non-answer.
Hiding that it is an agent
Say what it is. People are perfectly happy talking to one when it is useful and honest about what it cannot do.
Leaving the gaps unread
The list of questions it could not answer is the whole improvement plan. A business that reads it weekly gets better fast.
Experience it
See how it handles yours.
Talk to the agent on this site, or open a workspace and point it at your own material.
Monday
10:04
What other teams put in front of their customers first.
- Start with one page
- Not a six-month programme
- Measured from week one
- Answered, booked, handed over
- Improved from real gaps
- What it could not answer, weekly