A tool you can use
The AI Use-Case Triage Canvas
Eight questions to decide whether an AI use case should be piloted, scaled, paused, or rejected. Each one is a decision, not a discussion topic.
Most AI ideas probably should not reach a pilot. A demo can show possibility. It cannot tell you whether the outcome matters, what a mistake costs, or what evidence would justify the next step.
The canvas
Run one use case through eight questions.
- Use caseWhat specific task is the AI performing? Not "customer service." The actual workflow step.
- ObjectiveWhat problem is this meant to solve? What outcome should change?
- ValueWhat is the expected gain, and how will you measure it against a baseline?
- Error costWhat happens when the system is wrong? A false positive and a false negative are not the same problem.
- ReversibilityCan you undo the action if the model is wrong, or is the damage done?
- OversightWho reviews the output, how often, and what triggers escalation?
- EvidenceWhat evidence would be enough to keep going? Define this before the pilot starts, not after you have spent the budget.
- RecommendationPilot, scale, pause, or reject. Make the call.
If a team cannot explain the fallback, it is usually not ready to scale.
Worked example
Klarna's customer-service assistant
Early 2024. Klarna is about to deploy a generative AI assistant, powered by OpenAI, to handle customer-service chats. Imagine you are the AI lead asked to fill in the canvas before launch.
- Use case
- Resolve common support requests in chat inside the Klarna app (refund status, disputes, account and payment questions) without handing off to a person.
- Objective
- Cut time to resolution and the cost of support. The outcomes that should change: resolution time, cost per ticket, and the share of tickets closed without escalation.
- Value
Cost per ticket down. Resolution time down. More tickets closed without a human.
The gain is real only if customer satisfaction holds, repeat contacts do not spike, and complaints do not rise. If deflection goes up while satisfaction drops, you are moving the wrong number.
- Error cost
The assistant gives a customer wrong information about money: refund eligibility, a dispute outcome, an account balance. The customer acts on it and loses money or trust.
The customer bears the cost first, then the company. In Moffatt v. Air Canada, a tribunal held the airline to what its chatbot had told a customer.
- Reversibility
- Low. Customers act on the answer immediately, and retracting it does not undo the damage to trust. A person can step in only if escalation is one click away and customers actually use it.
- Oversight
The customer operations quality team audits a daily sample and reviews trends weekly. Every conversation carries a visible "talk to a human" option.
Pause triggers: satisfaction drops more than 3 points week over week, complaints spike, a regulator asks questions, or an error costs a customer real money.
- Evidence
- 90 days in one market and one product line, with satisfaction flat or better and no rise in repeat contacts or complaints.
08 · Recommendation
Pilot. Not scale.
The cost case is real and the technology demos well. But the error cost is high and reversibility is low. The oversight plan is the difference between a pilot and a scandal.
A responsible team should pilot this use case because the cost gain is real and measurable, but only if satisfaction and complaint volume are tracked from day one as co-equal measures, and one-click human escalation stays in place as a non-negotiable fallback.
What happened
In February 2024, Klarna announced that the assistant had handled 2.3 million conversations in its first month, two-thirds of its customer-service chats, the equivalent work of 700 full-time agents. Those are Klarna's numbers, from Klarna's press release.
In 2025, Klarna said it was shifting its AI focus from cost cuts to growth, and its CEO acknowledged the company had over-automated customer service (Reuters).
I don't know enough about Klarna's internal process to say it made the wrong call. The canvas makes a narrower point: the tension between cost and service quality shows up in question 3, before launch, not after the press release.
Use it
Bring me one use case.
Run one of your own use cases through the eight questions with the people who own it. The question you cannot answer is usually where the work is.
If you want a second view, bring the use case and the decision you are trying to make. We'll use a 30-minute call to see whether I can help.
The canvas is also on LinkedIn as a swipeable document.