The Frontier · Part II

AI in high-stakes industries

In finance, healthcare, insurance, and law, one sentence your AI produces can break a statute even if no one acts on it. This part teaches you to ship AI into those rooms with controls that hold up under an examiner.

The outcome

Ship an AI product into a regulated industry with the boundaries, gates, fairness evidence, and audit trail that let your compliance team and a regulator sign off on it.

Who this is for

PMs and builders shipping AI into finance, healthcare, insurance, or law, or any product where a wrong answer is expensive and public.

What you’ll walk out with

You arrive withYou leave with

Advice line left to the modelEnforce the advice line in the product
No boundary the product enforcesSort every request into three zones
Custody questions unansweredPick a model your risk team approves
Answers from training dataGround answers in an approved corpus
Unmeasured disparate impactProve fairness group by group
No record to show an examinerKeep an audit trail that replays

Start with chapter one The High-Stakes Clearance · fillable PDF

A real incident

A chatbot stated a refund policy that did not exist, and a tribunal made the airline pay for it. The chapters next door are how you keep one sentence from doing the same to your product.

01 / 09

A real incident

Air Canada's website chatbot told a grieving customer he could claim a bereavement discount after his flight was over. No such policy existed.

A tribunal held the airline liable for what its chatbot said. The core system was not broken; it produced a fluent, confident answer. The failure was that no boundary stopped it from stating a rule the company did not have, and nobody had decided who owned that sentence.

Moffatt v. Air Canada, BC Civil Resolution Tribunal, 2024

02 / 09

What happened

The bot described a refund path that contradicted the airline's own policy page.

THE CHATBOT, TO A GRIEVING CUSTOMER

If you have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued.

The linked policy said the opposite.

Moffatt booked full-fare flights and then applied for the refund the bot had described. Air Canada refused, because its real bereavement policy does not apply once travel is complete. The bot had stated a rule the company never offered.

03 / 09

What actually failed

The language model worked. Every control around it was missing.

No output boundary. Nothing checked the bot's answer against the one policy page that governed the question.

No ownership. Air Canada argued in front of the tribunal that the chatbot was a separate entity responsible for its own words.

No hand-off. A regulated question about fare rules had no path to the real policy or a human.

A better model would not have prevented this, because the gap was never in the model. It was in the boundaries nobody drew around it.

04 / 09

The pattern

A company owns every sentence its AI produces to a customer, whether or not the customer acts on it.

$650.88

in damages the tribunal ordered for the bot's misrepresentation

$812.02

total once interest and tribunal fees were added

The sum was small; the precedent was not. The tribunal called the airline's separate-entity defense a remarkable submission and rejected it. In a regulated setting, a wrong answer is a liability the company holds.

Moffatt v. Air Canada, BC Civil Resolution Tribunal, 2024

05 / 09

The one idea

In a regulated industry, one sentence your AI produces can create liability even if no one acts on it.

The model's fluency is the hazard, not the safeguard. What protects you is the set of boundaries, gates, and records you build around it before launch.

06 / 09

What this part builds

Five controls that turn a vague worry into a clearance a regulator can read.

Boundaries

Sort every request into may-do, must-refuse, and hand-off, with refusals that still help.

Approved model and corpus

Pick a model your risk team can approve and ground answers in sources you can defend.

Gates and fairness

Enforce policy the model cannot, and give every denial a specific, defensible reason.

Evidence and audit

Validate on your own population and keep a record that replays any interaction.

07 / 09

The objection worth answering

"A disclaimer covers us. We already say the bot might be wrong."

Air Canada had policy pages that contradicted its bot, which is a form of disclaimer, and the tribunal still held the airline liable. A fine-print caveat does not undo a clear answer a customer reasonably relied on.

In finance, healthcare, insurance, and law, the standard is higher than in retail. The control has to live in the product, not the terms of service, as a boundary that stops the wrong sentence before it reaches the user.

08 / 09

What actually fixes it

Each gap in this story maps to a chapter that closes it.

No must-never lineWhy high-stakes

→ Name your regulator and write a must-never list with a real penalty behind each line.

No refusal pathSetting boundaries

→ Sort every request into may-do, must-refuse, and hand-off, and write refusals that help.

No output gateSecurity guardrails

→ Build input and output gates that enforce the policy the model itself cannot.

No recordTesting evidence

→ Validate on your own population and keep a record that replays any interaction.

The gap that cost Air Canada in public is a page you fill in private.

09 / 09

From story to part

Every gap in this story is a chapter ahead.

Eight chapters take you from a stakes memo to a signed High-Stakes Clearance: boundaries the product enforces, a model and corpus you can defend, gates and fairness evidence, and an audit trail a regulator can read.

Start with chapter one →

↑ The Builder's Stack The Frontier map →