Authentic Intelligence · Companion sources
The book teaching guide
Start with the decision. These short teaching summaries connect ideas from the book to questions leaders can use in practice.
What this guide provides
These are prepared paraphrases of ideas in Authentic Intelligence: A Leader’s Guide to Ethical AI, with chapter references and application questions. They are not excerpts or page citations, and they do not reproduce the manuscript. Alongside the existing AIM framework and practical tools, they supply the source material for Ask Authentic Intelligence.
The framework supports judgment. It does not certify a system as ethical, compliant, safe, or technically valid. Applying a principle to a particular organization takes evidence and domain expertise.
For broader questions about AI, accountability, consent, human judgment, and the author’s perspective, explore the author Q&A and ethical thought experiments.
Chapter 1
When AI Moves the Decision
The decision can move before the signature
Start by asking who gets seen. A ranking, a hidden queue, a default button, or an incomplete summary can shape an outcome before anyone signs off. The decision system includes its interface, workload, incentives, data, and vendor arrangements. Review what the professional can see, what happens when they do nothing, and how hard it is to disagree. Calling a tool an assistant does not settle the authority it has in practice.
A question to work with: Trace one case from intake to outcome. Where could someone disappear from meaningful consideration?
Chapter 2
Regulated Trust
Compliance, risk, and legitimacy
Legal review matters, but it answers a different question from whether people can understand, challenge, and obtain correction of a decision. Risk management examines exposure and controls. Ethical legitimacy asks whether the institution can justify the decision process to those affected. A system may have completed procurement and still lack usable explanations or recourse. Bring those gaps into the decision before treating approval as a finished argument.
A question to work with: Could you explain this process to the person bearing its cost, using reasons you are prepared to own?
Chapter 3
Practical Ethics for Leaders
Rights first, reasons next, results over time
AIM uses more than one ethical lens. Rights establish protections that a business benefit cannot simply cancel. Reasons require a defensible justification and meaningful challenge. Results require attention to benefits and harms over time. Virtue and care bring leadership character, dependence, vulnerability, and context into the review. An accuracy gain or a faster process is useful evidence; it does not resolve every obligation the organization owes.
A question to work with: Which protection would be weakened by the proposed tradeoff, and what would change for the affected person?
Chapter 4
The Leadership Stack: Responsible and Authentic
Responsible leadership and authentic leadership
Responsible leadership widens the field of people whose interests count. Authentic leadership helps leaders retain moral agency under pressure: understand their limits, describe the AI role honestly, consider contrary evidence, and hold to a defensible moral standard. Authenticity here is more demanding than simply being yourself. A leader can support useful automation and still pause a proposal whose consequences the organization cannot own.
A question to work with: Whose perspective is missing from this review, and what evidence would change your recommendation?
When human review becomes a rubber stamp
A person can be formally responsible and practically unable to disagree. Look at workload, training, source access, default actions, performance targets, and the cost of an override. An unusually low override rate could reflect accurate recommendations or discouraged judgment; the number alone cannot tell you which. Examine actual decisions and whether a reviewer could recognize an error, challenge it, and correct the outcome.
A question to work with: Ask a reviewer to show what happens when they disagree with the system.
Chapter 5
Gate 0: The Decision Before the Decision
Type A: administrative support
Type A helps organize work without materially shaping a consequential outcome. Scheduling, file organization, and policy search can fit this role. The actual use matters: routing that causes a consequential delay may have crossed the boundary. Administrative does not mean risk-free. Privacy, security, accuracy, accessibility, records, ownership, and vendor controls still matter.
A question to work with: What consequence could an incorrect output have for the person relying on this workflow?
Type B: decision support
Type B informs human judgment while leaving the outcome open to meaningful review. Reviewers need access to evidence, enough time, competence, independence, authority to disagree, and a duty to use that authority. A generated summary should help someone examine the case; it should not become the only version of the case they are allowed to see. If defaults or workload make the recommendation decisive, review the use as possible Type C.
A question to work with: Can the reviewer reach a different decision in practice, and can you show a real example?
Type C: automation, including automation in practice
Type C determines, triggers, or effectively determines a consequential outcome. It includes explicit automation and human workflows where thresholds, defaults, volume, or incentives make the AI output decisive. Type C is not automatically prohibited. It carries the highest burden of evidence, rights protection, intelligible reasons, recourse, monitoring, executive ownership, vendor control, and stop authority. Reversibility and the burden imposed on people matter.
A question to work with: What happens before an authorized person reviews the case, and can the affected person obtain a timely correction?
How the 0–2 scoring works
A score of 0 means the condition is absent, symbolic, unsupported, or untested. A 1 means it is partly present but incomplete or not reliably operational. A 2 means it is documented, operational, owned, supported by evidence, and connected to review. Record the evidence and gap behind each score. Do not average away weak Rights or Reasons with strong efficiency results. A weak score should change the decision, its scope, or the conditions of use.
A question to work with: What evidence would move this condition from partly satisfied to operational? Name its owner.
Design a bounded pilot
A pilot is a controlled test of whether obligations can be met in the actual workflow. Define the use, AI role, population, permitted and prohibited actions, duration, human review, evidence, recourse, success and harm measures, monitoring, and pause or rollback triggers. Make expansion a separate decision. A useful experiment should not quietly become the operating model before its conditions have been met.
A question to work with: Write down what the pilot cannot do, its end date, and who can stop it.
Make exceptions explicit and time-limited
The book allows documented exceptions in difficult situations, including cases where not acting could also cause serious harm. Name the weak condition, why proceeding is justified, the risk of not acting, people affected, accountable executive, operational owner, time limit, recourse, monitoring, escalation, rollback authority, and review date. An exception is a visible judgment under defined conditions. It is not a waiver that silently becomes permanent.
A question to work with: What evidence or date ends this exception, and who is required to reconsider it?
Chapter 6
The Triangle: Rights, Reasons, and Results
The Triangle applies all three lenses
Rights, Reasons, and Results apply across the affected person, decision professional, and institution or community. Read the public Triangle guide for the teaching summary used by the companion.
Chapter 7
The Orbit: What the Workflow Does Not Show
Workforce effects and loss of judgment
AI can remove drudgery and create room for better service. It can also remove the work through which people learn judgment. Examine training, entry-level pathways, overreliance, surveillance, workload, and who receives opportunity or coaching. A productivity gain does not tell you whether the organization is becoming more capable. Leaders remain responsible for the work system they create.
A question to work with: Which skills will people still need when the tool is wrong or unavailable, and how will they practice them?
Environmental stewardship
Use a proportionate review of the infrastructure your AI use depends on. For significant deployments, ask about energy, water, emissions, hardware, model-call volume, and lower-compute alternatives. Tie the evidence to the organization's actual commitments. A small administrative use and a high-volume enterprise platform need different depths of review. Do not invent precise environmental numbers when the supplier cannot provide them.
A question to work with: What can the vendor substantiate about the footprint of this use, and would a simpler approach meet the need?
The AI portfolio and accumulating dependencies
The Orbit covers environmental stewardship; workforce and opportunity; security and resilience; community legitimacy; ecosystem and platform governance; and AI portfolio accumulation. Several individually reasonable tools can create shared dependencies, repeated burdens, or fragile operations. Examine vendor concentration and the combined effect on people across workflows. Approval for one function should not authorize every feature in a platform.
A question to work with: Which decisions and services would be affected if the shared vendor failed or materially changed its system?
Data stewardship and security
Map the data a system can access, why it needs them, where they are processed, and whether they are retained or reused for training. Include logs, generated outputs, sub-processors, permissions, accidental disclosure, and manipulation of inputs. Security is part of the institution's responsibility to the people whose information it holds. Verify the actual controls and domain obligations rather than relying on a general promise of privacy.
A question to work with: What sensitive information could enter prompts or outputs, and who can access, retain, or reuse it?
Continuity, rollback, and responsible exit
A tool can become operationally critical before anyone makes that decision explicitly. Plan how work continues if the model is unavailable, unreliable, or paused. Preserve staff capability, pending-case handling, records, affected-person protections, and a usable exit from the vendor. A pause that abandons people halfway through a decision is not a complete rollback plan.
A question to work with: If you turned the system off tomorrow, who would handle the people and cases already in it?
Chapter 8
When AI Screens Opportunity: Hiring and Talent
Define hiring criteria before scoring them
Before automating a judgment about candidates, ask whether the organization has defined and justified the criteria. Vague ideas such as fit, polish, or potential do not become more defensible when turned into scores. Review each function separately: scheduling, summarizing, matching, ranking, detecting, and excluding. Preserve applicant standing, relevant reasons, accessibility, evidence, human judgment, and a usable correction route.
A question to work with: What job-relevant evidence supports this criterion, and how would you explain it to a candidate?
Chapter 9
When AI Shapes Access: Lending, Underwriting, Insurance, and Eligibility
Lending and access to credit
Map the decision AI changes: organizing a file, recommending a risk assessment, setting terms, or deciding access. Examine proxies, missing information, historical labels, and who experiences delay or exclusion. Decision-relevant reasons must reflect the actual basis of the decision, not a plausible explanation written afterward. Preserve correction and recourse, and examine cumulative treatment across lending, pricing, collections, and hardship. Domain-specific legal obligations still require qualified review.
A question to work with: If the data are incomplete or wrong, how can the borrower correct them before bearing an avoidable consequence?
Insurance and claims triage
Ask what triage actually does to a claimant. Helping an adjuster decide where to look is different from automatically delaying payment or sending someone into a burdensome process. Review data relevance, proxy effects, explanation, accessibility, correction, and the authority to change a route. Measure who waits, how long they wait, and what happens after a challenge, alongside processing speed.
A question to work with: Does the flag merely suggest a review, or does it change payment, access, or burden before review?
Chapter 10
When AI Protects and Profiles: Fraud, Compliance, AML, and Security
Fraud protection and false-positive burden
Fraud prevention serves a real protective purpose. A false positive becomes an ethical event when it blocks access to money, triggers repeated scrutiny, or leaves a person without timely correction. Separate alert prioritization from actions such as freezing an account. Where immediate action is necessary, define narrow conditions, urgent authorized review, a safe explanation, correction, monitoring, and an accountable owner.
A question to work with: How quickly can a wrongly restricted person reach someone who can actually restore access?
Compliance copilots and AML review
Generated case summaries can reduce administrative work, but fluent language can make uncertainty look established. Keep generated narratives as drafts until an analyst verifies the underlying evidence. Examine whether alert prioritization causes some cases to receive no meaningful review. Limited disclosure may be justified in protective systems; internal reasons, authorized review, monitoring, correction, and accountability still need to function.
A question to work with: Which claims in the generated case note can the analyst verify against the original evidence?
Chapter 11
Beyond Finance: Care, Scarcity, Safety, and Public Trust
Healthcare, outreach, and missing data
Apply the Triangle to patient, decision professional, and institution/community. In scheduling or outreach, missing data or difficulty reaching someone can signal need rather than low priority. Ask whether a recommendation allocates attention or effectively removes access to care. Include accessibility, appropriate reasons, human review, continuity, and recourse. This is a governance application, not clinical advice or an authorization to use a system.
A question to work with: Who receives less attention under this workflow, and how will you find unmet need the data do not capture?
Utilities and essential-service priority
For outage response or service prioritization, examine the household, the responsible professional, and the utility/community. Capacity and safety constraints are real; they make the reasons for prioritization more consequential. Review vulnerable households, missing information, correction routes, security, resilience, emergency authority, and the ability to continue service when the AI or vendor fails. Local domain expertise must shape the review.
A question to work with: How will the process recognize a household whose urgent needs are not visible in its usual data?
Benefits, education, and public opportunity
AIM extends to systems that allocate eligibility, education, credentials, and public services. Identify the affected person, the decision professional, and the institution or public. Examine access, understandable reasons, meaningful challenge, missing-data effects, and whether people with less power can correct a mistake. Adapt the review to the domain rather than assuming that a hiring control transfers unchanged.
A question to work with: What would it take for a person with limited time, access, or institutional knowledge to challenge an error?
Chapter 12
Building the AIM Operating System
The minimum governance artifact set
For high-stakes Type B and C uses, preserve ten connected records: purpose and boundary; role classification and decision rights; Gate 0; Triangle and Orbit; affected-person protections; data and model evidence; vendor controls; monitoring and escalation; rollback and retirement; and the AIM decision memo. Scale the detail to the burden of the use. The point is that another person can reconstruct what was allowed, why, and who must act.
A question to work with: Which commitment currently exists only in a conversation, with no owner or evidence attached?
Make AIM part of the operating rhythm
Build AIM into seven moments: pre-procurement, pilot, deployment, monitoring, change control, renewal or expansion, and rollback or retirement. Each moment needs a question, an owner, and a record of the judgment. Revisit material changes in data, thresholds, interface, population, role, recourse, or vendor terms. The original approval remains meaningful only while its conditions still hold.
A question to work with: Where in your existing process could a material AI change happen without anyone revisiting the decision?
AIM for small teams and institutions
Keep the paperwork proportionate while preserving the protections. A small team can begin with one defined use, a short intake, role classification, named owner, vendor evidence, human review, and a monitoring and shutoff plan. Put weak-evidence exceptions in front of accountable senior leaders. Narrow the AI role when the institution cannot explain, monitor, challenge, or correct consequential outcomes.
A question to work with: What is the narrowest useful version of this proposal that your team can actually govern?
What the board needs to see
A useful board view shows the decisions AI influences, their role classifications, outstanding conditions, high-stakes uses, vendor concentration, monitoring triggers, complaints and recourse, exceptions, and upcoming changes or renewals. Include uses that were narrowed, paused, rejected, or rolled back. The board needs evidence of how governance changed decisions, alongside evidence of benefits.
A question to work with: What would tell the board that a previously approved use needs to be narrowed or stopped?
Bring a decision to the conversation
For a fuller working record, use the decision-memo template or the free companion workbook.