Domain Playbooks

This is Unit Four of Leading AI at Work, on the book's domain playbooks. Unit Three closed on a rule: AI risk lives in the things a feature touches. This unit asks which things those are when your work is not a hospital. Carry one question through this lecture: what is the first thing to look for in a domain you have never modeled? The answer arrives with a licensing office at the end. By the end you will be able to find any domain's version of the Incident Report and state the rule every playbook repeats. To name the Gold things, cards and AI limits across six domains. And to build a micro-playbook of your own.

Four entries on the objection that opens the playbooks, and the one rule all six share.

An objection, and a quick answer.

LAI 076 "But That's Healthcare"

The M-CaReM-AI method, by J. R. Carver, opens its playbooks with a common objection: the hospital case does not apply to us. In the book's scene, a claims supervisor is asked which decision people argue about, appeal or get audited on, and answers at once: the claim decision. That answer starts her own playbook.

The lesson from that exchange.

LAI 077 Every Field Has Its Incident Report

The book's lesson is that every field has its own version of the hospital's Incident Report. Usually it is the decision that matters most to the people affected: the claim, eligibility or credit decision, or the performance review. Finding that decision is the first move in any domain.

How the playbooks are built.

LAI 078 How the Playbooks Work

The book offers short playbooks for six domains: public services and benefits, healthcare, financial services, insurance, software platforms, and people systems. Each covers typical core things, Gold candidates, example cards, safe AI use and first steps. Start with the one closest to your context, and borrow freely.

The rule in every playbook.

LAI 079 Draft Around, Never Decide

One rule repeats across all six: AI may draft around a consequential decision, but never makes it. AI may draft a denial notice from a decision's fields; the denial itself comes from policy and a person. The draft is where AI saves time; the decision is where accountability lives.

Fifteen entries walk the six playbooks in turn, from benefits to people systems.

Benefits first.

LAI 080 Benefits: Core and Gold Things

The benefits playbook names the Case, the Eligibility Determination, the Benefit or Payment, and the Notice as Gold candidates. For each, a team should be able to point from every key property or status back to a specific rule, policy section or audit requirement. A status nobody can trace is a gap.

Three cards for a determination.

LAI 081 Benefits: Cards for the Determination

The playbook gives three cards for the Eligibility Determination. One creates it from an Application and a Case snapshot, recording household, income, assets, rule-set version, outcome and effective date; it must name a rule-set version and needs a valid Case and Application. One generates the Notice, stating the decision's basis in required language with appeal rights and deadlines. One supersedes a Determination after appeal, keeping the history. That third card is most often missing, and without it an agency cannot say which decision applied on a given day.

Where AI stops in benefits.

LAI 082 Benefits: Where AI Must Not Decide

The playbook says AI must not decide final eligibility, interpret statute, or set benefit levels without validated, human-approved rule engines. Its safe zones are drafting notices from approved templates, flagging missing information, and summarizing case histories for reviewers. Final values of Gold things stay governed by policy.

Healthcare, beyond incidents.

LAI 083 Healthcare: Orders and Consents

Beyond incident reporting, the healthcare playbook names Orders, Consents, Diagnoses and Billing Encounters as Gold candidates. Medication and procedure orders are safety-critical, consents are legally significant, and diagnoses and billing are central to care and revenue. Each carries a different risk, so each needs its own Pack and matrix.

Consent and medication cards.

LAI 084 Healthcare: Consent and Medication Cards

The healthcare playbook's cards capture a procedure consent, with the patient, procedure, risks explained, time, who consented and the witness, logging the language and method, such as in person or by video. Another revokes a consent and alerts downstream systems. A third places a medication order and checks allergies, interactions and dose limits. The book adds a requirement for healthcare AI pilots: each feature states the things it touches, the fields it reads and writes, and how clinicians review its output.

Friendly, and wrong.

LAI 085 Wrong in a Friendly Way

The healthcare playbook warns that a patient-friendly summary can be wrong in a friendly way. An AI summary of a care plan may read warmly and plainly while misstating a dose or a follow-up date. Patient-facing text needs the review a clinical note gets, because patients act on it.

Lending's Gold thing.

LAI 086 Lending: The Credit Decision

The lending playbook makes the Credit Decision Gold: approve or decline, amount, rate and reasons. Its Pack lists the fields used, rationale fields and policy references; its matrix ties it to models, data sources, notices and reports. That structure lets a lender answer a regulator asking how a decision is made.

Keeping a model tied down.

LAI 087 Lending: Decision and Notice Cards

The lending cards keep an AI underwriting model tied to a defined decision. The creating card records every input, from application and bureau data to risk scores, assigns adverse-action reasons when declining, and ties the decision to a policy version and a model version. Another generates the adverse action notice as the law requires. A new model's features map to the decision's fields, and the matrix records how each model change affects it. The model version is the field teams forget, and without it nobody can say which model shaped a decline.

Bias work is not the finish.

LAI 088 Mitigated Is Not the Same as Fair

The lending playbook advises testing outcomes across groups even after bias work is done. It rests on a point in NIST's AI risk framework: systems in which harmful biases are mitigated are not necessarily fair. A bias fix is a step, not a certificate; outcomes still need checking.

Insurance's Gold things.

LAI 089 Insurance: Claims and Decisions

The insurance playbook names the Policy, the Claim, the Claim Decision and the Payment as Gold candidates. Its cards register a claim, checking the incident date falls within the policy period, and create a decision with reason codes linked to policy clauses. These things move money and shape what customers experience.

Where a triage score belongs.

LAI 090 Triage Fills the Assessment

In insurance, AI usually enters through triage, and the book says where its output goes. Triage models fill Claim Assessment fields, which feed the Claim Decision's logic but do not replace it. Keeping the score on the assessment is what keeps the decision governed.

Platforms' Gold things.

LAI 091 Platforms: Tenants and Subscriptions

For software products, the playbook names the Tenant, the Subscription and the Invoice as Gold candidates, sometimes the Workspace too. The Tenant carries billing and data separation; a plan-change card must handle proration, feature availability and data retention. On a platform, Gold things are often money and the walls between customers.

AI features as features.

LAI 092 Features, Not Magic

The platforms playbook treats AI features as features, not magic, organized through cards and Packs like any other. The method states which things each AI feature reads and writes, such as Tasks, Comments or a Workspace summary. Without that, AI content smears across things with no clear meaning.

People systems.

LAI 093 People Systems: Reviews, Applications, Pay

The people-systems playbook names the Performance Review, the Compensation Record, the Application and the Employee as Gold candidates. They carry career consequences, money, fairness in hiring, and personal data; a review card restricts who may see it. These decisions affect livelihoods.

The error nobody sees.

LAI 094 Screening Errors Are Silent

A screening tool may suggest a shortlist, the playbook says, but status changes stay governed by cards with fairness rules. Screening errors are silent: the qualified applicant filtered out never appears in anyone's review. That is why outcomes must be monitored across groups.

Three entries on the moves every domain shares, a micro-playbook, and a new domain. The opening question is answered here.

The same moves everywhere.

LAI 095 Six Domains, Same Moves

The book sums up its playbooks: different domains share the same moves. Name the things, choose Gold anchors, describe behavior as cards, make change safe with matrices, and let AI assist around the structure, never define it. Everything else, it says, is your industry's accent on the same underlying language.

A micro-playbook in four steps.

LAI 096 The Micro-Playbook for Your Domain

The book closes with a micro-playbook. Pick one or two decision things whose values really matter to the people affected. Make them Gold. Build their Packs and matrices, tracing each key property to its rule. Insist that every serious change and every AI feature starts from them. This course adds a fifth step as practice: write one sentence on what AI may draft around those decisions, and one on what it may never decide. Decision things are the ones people argue about, not the most used, and choosing by daily use is the common way to pick the wrong Gold thing.

A playbook for licensing.

LAI 097 A Playbook for a Licensing Office

A worked example, constructed for this course, builds a playbook for a professional licensing office, a domain the book names but does not detail. Nora asks the book's opening question, which decision do people argue about, appeal or get audited on, and finds the License Decision: issue, deny, suspend or renew. The Disciplinary Action is her second Gold thing. Tracing statuses to regulations shows suspended has two meanings, missed fees and discipline, so it becomes two statuses. She writes three cards, including a supersession that keeps history, then her two sentences: AI may draft letters, summarize complaint files and flag missing records; it may never decide a license or whether a complaint is substantiated. That answers this lecture's question: in any new domain, look first for the decision people argue about.

Look at what you can now do. You can find any domain's version of the Incident Report, the decision people argue about, and state the rule every playbook repeats: AI drafts around it and never makes it. You can name the Gold candidates, key cards and AI limits across benefits, healthcare, lending, insurance, platforms and people systems. And you can build a micro-playbook for a domain of your own, choosing by consequence. Unit Five turns to the operating model.