Contents

Appendix E

The same method on four other uses

The service-credit assistant stays the spine of the book. These four short cases apply the same four questions - what may the output do, who owns it, what do we know, when does permission end - to other common uses. What changes is the data, the way each use fails and what a failure costs.

The company, Maya, Daniel, Leila, Priya, Nadia, the CFO and Tom are the book’s people. The scenarios are invented for illustration, as the case itself is. No figures are claimed.

1. Forecasting

The CFO asks whether a model could forecast service-credit claim volumes for next quarter so staffing can be planned.

What is different here

The data is numbers over time, not documents. Nothing is retrieved. The tool returns a prediction, and people will tend to treat a number with decimals as more certain than it is.

How it fails

A forecast fails quietly. It can fit the past well and still be wrong, especially if the past was unusual. In the case, last year's claim counts were inflated while procedures conflicted, so the model learns a pattern that the repaired process has already removed. Nobody sees the error until staffing is wrong.

QuestionIn this case
What may the output do?Recommend. It feeds a staffing decision that a person makes. It must not set the budget by itself.
Who owns it?The CFO owns the planning decision. Daniel owns the data pipeline. Leila confirms the service process has not changed. One named person can switch the forecast off.
What do we know?Known: the history. Assumption: next quarter resembles it. Unknown: whether the repaired procedure changed claim behaviour. State all three beside the number, with a range, not a single figure.
How do we test it?Compare against a simple baseline, such as the same quarter last year adjusted by hand. If the model does not beat the baseline by a margin the CFO cares about, it has not earned a place.
When does permission end?Review each quarter against what actually happened. Review early if the procedure, the customer mix or the pricing changes.

2. Image analysis

Customer operations proposes that a system look at customer-submitted photos of damaged goods and flag credit claims that look fraudulent or not damaged.

What is different here

The data is images from customers' own phones. They contain personal information the customer never meant to share: a face in a mirror, a house number, a name on a parcel. Quality varies with lighting, angle and camera.

How it fails

The errors are not evenly spread. A system can be accurate on clear photos and poor on dark or blurry ones, which tend to come from customers with older phones. A wrong 'not damaged' flag can lead to a refused credit. The cost of a false flag falls on the customer, who may not know a machine contributed.

QuestionIn this case
What may the output do?Assist at most: highlight a photo for a person to look at. If a flag is allowed to refuse a claim, that is Determine, a different decision with a much higher bar.
Who owns it?The claims lead owns the outcome. Priya's team checks flagged photos. Daniel owns configuration and access to the images. Nadia owns the question of what personal information the photos may be used for.
What do we know?Known: accuracy on the supplier's examples. Unknown: accuracy on your customers' real photos, by lighting and device. Test on a sample that includes poor photos and say so.
Where is the consequence?Check the cost of a miss in each direction. Count how many flagged claims a person overturns. Provide a way for the customer to ask for a human look.
When does permission end?Review when a new product line, new packaging or a new photo upload route changes what the photos look like, and on a fixed date in any case.

Which of the four uses is closest to something in your work, and what may its output do?

Guidance

Choose the use you can describe in one sentence. Say whether the output informs a person, drafts something for review, or acts by itself. Then name the one person who could switch it off.

Your answers are saved only in this browser on this device. Nothing is sent anywhere.

3. Content generation

The marketing team asks to use a writing assistant to draft customer emails, web copy and social posts. In the book's register this is the separate 'marketing assistant' entry, not a benefit of the service-credit use.

What is different here

The output is text that other people will read as the company's own voice. The checking question changes from 'is this retrieval right' to 'is this true, owned and fit to publish'.

How it fails

The model produces fluent, confident text that may state things the company cannot support: a price, a promise, a claim about a product, or wording close to someone else's. Fluency hides weakness. Edits can drift the brand voice. Public channels are harder to take back than internal ones.

QuestionIn this case
What may the output do?Assist. It produces a draft. A named person reads every public item before release. For internal first drafts, the check can be lighter. The audience sets the check, not the tool.
Who owns it?The marketing lead owns what goes out. Nadia owns the list of claims the company may not make. Daniel owns access and which text may be pasted in (no customer data, no unreleased pricing).
What do we know?Known: the team likes the drafts. Unknown: whether the drafts are better, faster including checking time, or just more. Measure the edit rate and the time to approved, not time to first draft.
What would change the answer?A new channel (public post instead of internal note), a new audience (customers instead of staff), or any claim about performance, price or law.
When does permission end?Short review, because the output is public. Review at once on a published error, a complaint, or a change of supplier or terms.

4. Autonomous agents

A supplier offers an 'agent' that can read the service inbox, look up accounts, issue small credits, email customers and update the case record without a person in each step.

What is different here

Everything in the earlier cases was a tool that produced an output for a person. An agent takes actions. The question changes from 'is the answer right' to 'what can it reach, and can it be stopped or undone'.

How it fails

Errors chain. One wrong reading of an email leads to a wrong lookup, a wrong credit and a wrong message, each step reasonable given the last. By the time someone looks, the effects are in several systems and customers' inboxes. Credentials make the reach large. Small permissions become large ones when combined.

QuestionIn this case
What may the output do?Act is the highest authority mode, so start lower. Run it in Inform or Assist inside a test copy first. Give permission for each kind of action, not for 'the agent'. Money, messages to customers and record changes each need their own decision.
Who owns it?Maya owns the decision to grant any action. Daniel owns its credentials and can cut access at once. A named person on each shift knows how to stop it. If the owner leaves, the permission lapses until someone is named.
What do we know?Known: what the supplier demonstrated. Unknown: what it does with an ambiguous email, a hostile email or a customer who is upset. Test those cases on purpose and keep the logs.
What limits apply?Spend and volume caps, an approved list of actions, no new access without a decision, and a record of every action that a person can read afterwards.
When does permission end?Short, in weeks, with a hard end date in the access setting. Review early on any unexpected action, any change to what it can reach, or any provider update.

What stays the same across all four

ForecastingImage analysisContent generationAutonomous agent
Highest sensible mode nowRecommendAssistAssistInform or Assist, in a test copy
What the data isNumbers over timeCustomers' photosText going publicLive systems and inboxes
Main way it failsQuietly wrong, looks certainErrors uneven by groupFluent and unsupportedErrors chain across steps
Main safeguardBaseline and rangeTest on poor photosNamed reader before releasePer-action permission and a stop

What would you compare that use against, and what would tell you its permission should end?

Guidance

Name a simple baseline you could check by hand. Then name one change, such as new data, a new audience or a new tool connection, that would send the use back for review.