World Model Readiness
Engraved functions instrument

For your company

AI in the Functions

Module · Where the bot ends, a human begins

The AI-in-Customer-Service Check

A support bot handles the easy questions cheaply, right up until it confidently gives a wrong answer to an angry customer with nowhere to go. The economics are real and so is the reputational downside. This module checks the five controls that decide whether automated support helps or corrodes trust: a defined scope, a working escalation path, honest disclosure, quality monitoring of what the bot actually says, and a loop that learns from the conversations it got wrong.

Question 1 of 5 · Bot scope is defined

Is there a clear line for what your support bot may and may not handle?

A bot without a defined scope answers everything, including the refund policy it guessed and the medical question it should never touch. Scope written as explicit boundaries, enforced by the system, is what keeps a helpful assistant from becoming a liability with a chat window.

Question 2 of 5 · Escalation actually works

When the bot cannot help, can a customer reach a human without a fight?

The fastest way to turn a minor issue into a lost customer is a bot that loops, deflects and hides the exit. A named, easy path to a human, triggered by the customer or by the bot recognising its own limits, is the safety valve the whole system depends on.

Question 3 of 5 · Customers know it is AI

Do customers know when they are talking to a bot rather than a person?

Pretending a bot is human buys a few smoother minutes and a lasting trust cost when the customer works it out, and they do. Clear disclosure is increasingly a legal duty as well as a decency one, and it sets expectations that make the whole interaction go better.

Question 4 of 5 · Answers are monitored

Do you know how often your bot gives customers wrong or unhelpful answers?

A bot fails silently: no complaint, no ticket, just a customer who quietly gave up or acted on bad information. Without sampling and monitoring the actual answers, you are trusting a system you have never audited to speak for you thousands of times a day.

Question 5 of 5 · Failures feed learning

When the bot fails a conversation, does anything change as a result?

Every bot has conversations it handles badly. The question is whether those failures end as a shrug, or as a fix to the knowledge base, the scope or the escalation rule. A support bot that does not learn from its worst days repeats them indefinitely.

For the statistics · one click each

Three questions for the public picture

These do not affect your score. They feed the anonymised, aggregated statistics; groups under 8 respondents are never shown.

What share of customer contacts does AI handle without a human?

None
Under a quarter
A quarter to half
More than half
We do not track it

Do you tell customers when they are talking to a bot?

Yes, up front
Only if they ask
No
We are not sure
No support bot yet

How easily can a customer reach a human when the bot cannot help?

One step, quickly
Possible but slow
Hard to find
No human option
No support bot yet

Your context

Used to calibrate the report. Company size and sector remain in the anonymized dataset; your email does not.

What the five levels look like

Every dimension in this assessment is scored 1 to 5. This is what the levels mean, dimension by dimension. The graded report diagnoses where your own answers land and what to do about it.

Bot scope is defined

  1. 1Answers anything
  2. 2Vague intent
  3. 3Topics listed
  4. 4Scoped and enforced
  5. 5Scoped, tested at edges

At the low end: A bot that will attempt any question will eventually answer one it had no business touching. Define the topics it owns and the topics it must hand off, and enforce that line in the system. What good looks like: A scope tested at its edges is what lets you trust the bot on the front line. Revisit it as you add capabilities; every new skill widens the surface where it can go wrong.

Escalation actually works

  1. 1No human exit
  2. 2Buried and slow
  3. 3Human on request
  4. 4Bot escalates itself
  5. 5Seamless handoff with context

At the low end: A bot with no exit to a human traps your most frustrated customers with your least capable agent. Add a clear, fast route to a person before you widen what the bot handles. What good looks like: A seamless handoff that carries the conversation context is what makes automation feel like service rather than a wall. Watch the escalation rate; a sudden climb tells you the bot's scope has drifted past its competence.

Customers know it is AI

  1. 1Poses as human
  2. 2Ambiguous
  3. 3Disclosed if asked
  4. 4Clearly disclosed
  5. 5Disclosed with easy opt-out

At the low end: A bot that lets customers believe it is human is a trust debt that comes due the moment they notice. Disclose that it is AI up front; the honesty costs nothing and the deception costs a relationship. What good looks like: Clear disclosure with an easy route to a human is the standard regulators and customers now expect. Keep the opt-out genuinely easy; disclosure without a real alternative is only half the promise.

Answers are monitored

  1. 1Never checked
  2. 2Complaints only
  3. 3Occasional sampling
  4. 4Regular quality review
  5. 5Scored against a standard

At the low end: A bot nobody audits is a system speaking in your name with no quality control at all. Start sampling its conversations weekly; the first read will tell you whether you have a helper or a hazard. What good looks like: Answers scored against a written standard turn quality from a hope into a number you can manage. Feed the low scores into the learning loop; monitoring only pays off if the failures change something.

Failures feed learning

  1. 1Failures ignored
  2. 2Noticed, not acted on
  3. 3Fixed ad hoc
  4. 4Reviewed and improved
  5. 5Closed-loop improvement

At the low end: Ignored failures are failures you have chosen to repeat. Start the simplest log of conversations the bot got wrong, and review it before you next touch the knowledge base. What good looks like: A closed loop from failed conversation to concrete fix is what makes the bot better next quarter than this one. Track how fast the loop closes; that latency is the real measure of your support quality.