THE NEW
80/20 RULE

The Dermer Rule is Michael Dermer’s framework for the AI era: 80% of work should be AI. 20% must be human. The 80% is the scalable, repeatable work AI can perform or accelerate; the 20% is the judgment, relationships, taste, ethics and accountability where being human changes the outcome.

The Dermer Rule says: 80% of work should be AI. 20% must be human. It is a work-design principle, not a literal headcount formula. AI should do as much of the repeatable research, analysis, creation, prediction, personalization, optimization, coordination and execution as it can do reliably. Humans should own judgment, context, trust, empathy, taste, ethics, courage, vision, exceptions and accountability.

No. The Dermer Rule refers to 80% of work, not 80% of people. Its purpose is to move routine and scalable work to AI so humans can spend more time on judgment, trust, creativity, relationships and accountability.

The 80/20 split is directional, not a mathematical requirement for every task or company. Some work may be 95% AI-led; other work may remain mostly human. The value of the framework is forcing an explicit decision about what belongs on each side.

Pareto’s 80/20 Rule explains where results often come from; the Dermer Rule uses 80/20 to redesign who—or what—should do the work. The Dermer Rule asks a modern operating question: what should AI do, and what must humans continue to own?

AI-led work is work where AI performs the first, largest or continuous share of the task rather than merely assisting at the end. Humans can still set goals, constraints and approvals, but the workflow is designed around AI doing the scalable work by default.

Human-owned work is work where a person remains responsible for the judgment, relationship, exception or consequence. AI can inform or prepare that work, but the human is not merely approving a machine output; the human owns the decision and outcome.

The Dermer Rule is designed to answer one question: what should AI do, and what must remain human? It gives leaders, employees and entrepreneurs a practical way to redesign work instead of treating AI as a collection of disconnected tools.

Start with the work, not the tools. Map one important workflow, identify the repetitive research, analysis, drafting, monitoring and execution AI can take over, then mark the points where human judgment or accountability must remain. That creates a real AI operating model instead of scattered tool usage.

Automate high-volume, repetitive work with clear outputs and low downside if something goes wrong. Good first targets include research, summarization, drafting, classification, follow-up, reporting and monitoring. Keep high-stakes exceptions and decisions human until the system proves reliable.

Do not start with the highest-stakes, most ambiguous or relationship-sensitive decisions. Begin with repeatable work where errors are easy to detect and reverse. Keep consequential judgment, trust and accountability human while the organization learns.

Start by auditing the work, not by buying more AI tools. List the recurring tasks in the workflow, move research, analysis, drafting, monitoring and routine execution toward AI, and explicitly mark the judgment, trust and accountability that must remain human.

Start with workflows, not tools. Map the work, identify repeatable steps, decide where human judgment is essential, and then choose AI tools that fit the redesigned process. Tool-first adoption usually creates isolated productivity tricks instead of operating leverage.

Start where there is a large amount of repeatable work and a clear business outcome, then expand. A focused first area makes it easier to measure value, establish guardrails and learn what belongs in the AI 80% versus the Human 20%.

Prioritize AI use cases with high work volume, clear outputs, measurable value and manageable risk. The best early opportunities usually combine meaningful economic impact with repeatability and easy human oversight.

A workflow is ready for AI when the goal, inputs, acceptable output and escalation rules can be defined clearly. If nobody can explain what good looks like or when a human must intervene, automate later.

AI should touch most knowledge-work workflows, but it should not own every decision. The Dermer Rule sets the direction: 80% of work should be AI; 20% must be human. Push repeatable research, analysis, creation, monitoring and execution toward AI, while preserving human judgment, trust and accountability where the consequences matter.

Most knowledge workers should learn to use AI, but not every employee needs the same tools, autonomy or use cases. The right design depends on the work. AI use should expand where it improves speed or quality without weakening safety, judgment or accountability.

AI should be considered in nearly every function, but it should not be forced into work where it adds no value or creates unacceptable risk. The Dermer Rule is a default question, not a requirement to automate everything.

AI autonomy should rise as work becomes lower-risk, more measurable and more reversible. Routine work can become fully autonomous; high-stakes, ambiguous or irreversible work needs stronger human checkpoints. The goal is maximum useful autonomy, not maximum autonomy.

AI can work without case-by-case approval when the task is low-risk, measurable, reversible and well monitored. High-stakes decisions, unusual exceptions and actions with legal, financial, safety or reputational consequences should retain a meaningful human checkpoint.

Give AI as much autonomy as the risk and reversibility of the work justify. Low-risk repeatable work can move deep into the AI 80%; high-stakes, ambiguous or relationship-sensitive work needs stronger human ownership.

AI can execute large portions of a process, but organizations still need humans to set goals, handle exceptions, make consequential tradeoffs and own outcomes. Full autonomy is most appropriate in bounded, measurable processes—not open-ended leadership.

AI can run large parts of a company’s operating work, but it should not own the company’s purpose, major tradeoffs or accountability. A company can become highly AI-operated while remaining human-led.

Yes. High-stakes AI output should normally have human review and clear accountability. The higher the consequence, ambiguity or irreversibility, the more the Human 20% matters.

Human-on-the-loop means AI operates autonomously within defined boundaries while a human monitors performance and can intervene. It is useful when case-by-case approval would destroy the efficiency benefit but the process still needs accountable oversight.

Humans should always own consequential judgment and accountability. AI can do the research, analysis and recommendation, but people must own decisions involving values, trust, major exceptions, legal or financial consequence, safety and the outcomes the organization must stand behind.

AI should never be the only accountable owner of a consequential decision. The more a decision affects people, money, safety, rights, reputation or irreversible outcomes, the more important a meaningful human owner becomes.

Humans should own consequential judgment, values, accountability, trust, major exceptions and decisions where context can change the right answer. AI should still do the research and analysis around those decisions, but the person must remain responsible for the outcome.

The hardest judgment for AI is the kind where context changes the right answer — high-stakes tradeoffs, situations with no precedent, and calls that weigh competing values. AI can gather the facts and lay out the options, but a person has to decide what actually matters and own the outcome. That judgment sits squarely in the Human 20%.

Context matters because the same facts can call for different decisions depending on the situation, the relationship, and what's at stake. AI works from patterns in past data; a human reads the specific moment — what's changed, who's affected, what the numbers don't show. When context can flip the right answer, the decision belongs to a person.

AI can support trust through consistency and responsiveness, but durable trust still depends heavily on human credibility and accountability. The more consequential the relationship, the more important it is that a person can explain, decide and stand behind the outcome.

AI can simulate empathetic language, but human empathy matters most when the relationship, emotion and consequence are real. In sensitive moments, people often need to know that another person understands and is accountable—not only that the words sound empathetic.

AI can generate options and imitate patterns; taste is the human judgment about which option deserves to exist. Taste matters in brand, product, design, storytelling and any context where 'technically correct' is not the same as 'good.'

AI cannot bear legal, ethical or organizational accountability in the way a person or institution can. Someone still has to own the decision, explain it and live with its consequences.

Ethics belong to humans because someone has to be accountable for the values behind a decision, not just its efficiency. AI can flag issues and apply rules, but it can't carry moral responsibility or answer for the consequences. When a choice affects people, fairness, or trust, a human must own it.

Courage is human because the hard calls — delivering a difficult truth, making an unpopular decision, taking a real risk — require someone willing to be accountable for them. AI can model the options and the odds, but it has nothing at stake. Standing behind a tough decision is the founder's job, not the machine's.

Vision belongs to humans because deciding where a business should go, and why, is a matter of purpose and conviction, not pattern-matching. AI can surface trends and possibilities, but it can't choose what the company should stand for or commit to a direction. Setting the destination is human work; AI helps you get there.

Exceptions are human work because they're the cases the rules didn't anticipate — where following the standard process would get it wrong. AI is excellent at the repeatable 80%; when something falls outside the pattern, a person has to weigh the specifics and decide. Handling the exception well is a core part of the Human 20%.

Yes. As AI makes routine work cheaper and faster, the scarce human capabilities become more valuable. Judgment, trust, taste, leadership and accountability increasingly become the differentiators rather than the volume of work a person can produce.

The set of tasks in the Human 20% may change, but the need for human ownership does not disappear. As AI improves, humans should keep moving away from routine production and toward higher-consequence judgment, relationships and accountability.

AI should make many routine operational decisions; humans should own consequential decisions. Let AI analyze options, apply rules and make bounded low-risk decisions. Keep strategic tradeoffs, major people decisions, risk acceptance and accountability in the Human 20%.

AI can do much of the analytical work behind a decision; humans should own consequential choices. The Dermer Rule puts analysis, prediction and option generation in the AI 80%, while context, tradeoffs, values and accountability stay in the Human 20%.

AI can inform this decision, but it should not be the sole accountable decision-maker. Use AI for research, consistency checks and scenario analysis; preserve meaningful human judgment because the consequences are high and context matters.

AI can prepare agendas, summarize information, surface issues and track actions, but humans should own debate, conflict, commitment and decisions. The value of an executive meeting is not merely information transfer—it is alignment and accountable choice.

AI can improve board preparation and analysis, but governance judgment must remain human. Directors are responsible for challenge, independence, fiduciary judgment and accountability.

Leaders should use AI recommendations as decision input, not automatic authority. AI is excellent at comparing information and surfacing options; leaders must still apply context, values, risk tolerance and accountability.

Leaders should treat disagreement with AI as a prompt to examine assumptions, not as evidence that either side is automatically right. Ask what data and logic produced the recommendation, test alternative scenarios, and then make the decision with explicit human accountability.

Choose the model that performs your actual workflow best—not the model with the loudest brand or highest benchmark. Test a small set of realistic tasks for accuracy, speed, cost, privacy, tool access and reliability. Most companies should expect to use more than one model over time.

Most companies should be willing to use multiple AI models when different models are materially better for different work. Standardize enough to control security and cost, but do not lock every workflow to one vendor if another model produces meaningfully better outcomes.

No. There is no single best AI model for every business task. The right choice depends on the job: reasoning depth, speed, cost, context size, multimodal capability, tool use, privacy and reliability all matter.

Use a reasoning model when the task requires multi-step analysis, tradeoffs, planning or difficult problem solving. Do not pay reasoning-model cost and latency for routine extraction, formatting, classification or simple drafting.

Use faster, cheaper models for high-volume, low-risk and easily checked work. Reserve more capable models for work where better reasoning materially changes the outcome.

Different departments may need different models, but the company should manage them through a common security and governance standard. Finance, engineering, marketing and customer service can have different requirements without becoming four uncontrolled AI ecosystems.

Reevaluate important AI workflows regularly because model capability, price and reliability change quickly. Re-test when a major model changes, when workflow economics shift, or when performance no longer meets the business standard.

Avoid unnecessary vendor sprawl, but do not create strategic dependence on one model if portability is practical. A small approved portfolio can preserve negotiating leverage, resilience and access to best-in-class capabilities.

Choose AI tools by workflow outcome, not by brand. Test quality, speed, cost, privacy, integrations and reliability against the actual work.

Use public benchmarks only as a screening signal; choose models with your own workflow tests. A model can rank highly overall and still perform poorly on your documents, customers, tone, data or decision criteria.

Use an off-the-shelf AI tool when it already solves the workflow well and the integration, security and economics are acceptable. Build custom only when your process, proprietary data, control needs or strategic advantage justify the extra complexity.

Build custom AI when the workflow is important enough that proprietary data, integration, control or differentiation creates real economic value. Do not custom-build commodity capabilities that a mature product already provides well.

Use the fewest AI tools needed to cover materially different workflows well. Tool proliferation increases cost, security risk, training burden and fragmentation. Consolidate overlapping tools unless the performance difference matters.

Yes—standardize the approved AI stack, data rules and access model, while allowing controlled exceptions for better-fit workflows. Standardization should reduce risk and friction without preventing teams from using materially better capabilities.

Evaluate enterprise AI providers on workflow performance, data handling, security, admin controls, reliability, integrations, portability and total cost. Model quality matters, but so do the controls required to use it safely at scale.

Price matters only after the model meets the required quality and risk standard. A cheap model that creates rework or bad decisions is expensive; an expensive model used for trivial tasks is wasteful.

Latency matters when AI sits inside a live customer or employee workflow; it matters far less for asynchronous analysis. Match speed requirements to the workflow rather than treating fastest as automatically best.

Employees should choose only from approved AI tools unless there is a controlled exception process. Freedom at the workflow level is useful; unmanaged vendor, data and security choices are not.

Use AI to synthesize and navigate research; use primary sources and search to verify important facts. AI is often faster for forming an answer, but high-consequence research still requires checking the underlying evidence.

Use AI search when you need synthesis across sources; use traditional search when you need direct source discovery, exhaustive browsing or precise navigation. In important work, the best process often uses both.

Use deep research when the question needs many sources, comparison, synthesis or a documented evidence trail. It is unnecessary for simple facts or tasks where you already have the authoritative source.

Trust AI research only to the extent that its important claims can be traced to credible evidence. Treat the synthesis as useful work product, not as a substitute for source quality and verification.

Verify the claims that drive the decision, especially numbers, dates, quotes, legal or medical facts and surprising conclusions. Open the cited sources and check whether they actually support the claim.

For consequential or externally used research, yes—AI should provide traceable sources wherever possible. Citations make human review faster and reduce the risk of accepting fluent but unsupported claims.

Prefer primary, authoritative, current sources and ask whether the source actually supports the specific claim. Reputation alone is not enough; relevance, recency and direct evidence matter.

AI should do most of the searching, summarizing and comparison; humans should own source judgment, interpretation and consequential conclusions. That is the Dermer Rule applied to research.

Yes. Continuous monitoring is an excellent AI use because the work is repetitive, high-volume and time-sensitive. Humans should define what matters, validate important signals and decide how the company responds.

AI can replace a large share of routine research work, but strong analysts still add value through question selection, source judgment, interpretation and implications. The job should shift from collecting information toward deciding what the information means.

Use AI every day for preparation, research, drafting, summarizing, analysis, follow-up and routine coordination. The objective is to remove work around your real work, not create a new hobby called 'using AI.'

Start with repetitive tasks you do frequently: email drafts, research, meeting preparation, notes, summaries, first drafts and data analysis. High-frequency work compounds the productivity gain fastest.

AI can organize and recommend priorities, but you should own what deserves your attention. Give it goals, deadlines and commitments; keep the final tradeoff in the Human 20%.

AI can schedule, reschedule, prepare and protect focus time within explicit preferences and constraints. Sensitive relationship choices and unusual conflicts may still need human judgment.

Yes. Email drafting is a strong AI 80% task because the human can provide intent and quickly review the output. Important, sensitive or relationship-defining emails should preserve human judgment and voice.

AI can answer routine email within defined rules; consequential, sensitive or ambiguous messages should escalate. The right design is not 'AI answers everything' but 'AI handles the normal path.'

Yes. AI should gather background, summarize prior interactions, identify open issues and suggest questions before important meetings. Humans should own the conversation and judgment in the room.

Yes. Transcription, summaries, decisions and action-item extraction are ideal AI tasks. Review high-stakes commitments and make sure the system distinguishes discussion from an actual decision.

AI should create the first draft of most routine presentations and reports. Humans should own the message, prioritization, interpretation and what the audience needs to understand.

AI can analyze spreadsheets, explain trends, find anomalies and draft models, but material financial or operational conclusions need verification. Use AI to accelerate analysis, not to eliminate accountability for the numbers.

Yes. AI can classify, summarize, tag and retrieve documents when permissions and data controls are appropriate. Keep authoritative versions and retention rules explicit.

If AI creates more drafts, tools, notifications and review than value, the workflow is badly designed. Measure time and outcomes, eliminate duplicate steps and automate end-to-end rather than adding AI on top of the old process.

AI should be considered in every significant process, but not forced into work where it adds no value or creates unacceptable risk. The default question should be 'what can AI own here?' not 'where can we sprinkle an AI feature?'

An AI-first company designs work assuming AI handles scalable cognitive labor by default and humans own the work where human judgment changes the outcome. It is an operating model, not a software purchasing strategy.

Using AI tools improves individual tasks; redesigning work changes the workflow, roles, approvals and economics around AI. The second produces much larger and more durable gains.

The business owner of the workflow should decide the AI 80% and Human 20% with input from technology, risk and the people doing the work. IT should enable the system, not unilaterally define the business judgment.

Revisit important workflows whenever model capability or economics materially change, and at least as part of regular operating reviews. A workflow designed around last year's AI can become obsolete quickly.

Every consequential AI workflow should have a named human owner, even if most execution is autonomous. The AI system needs operational ownership; the outcome needs human accountability.

Maintain a simple inventory of AI workflows, purpose, data, model/tool, permissions, risk level, owner, human checkpoint and business metric. This makes scaling and governance far easier than discovering AI usage after something goes wrong.

Use a risk-tiered approval process: fast paths for low-risk workflows and deeper review for sensitive data, external actions or consequential decisions. Governance should accelerate safe use rather than make every experiment wait for a committee.

Tie every AI pilot to a named workflow, owner, baseline metric and decision about what happens if it works. A demo without an operating path is not transformation.

Move from experimentation to adoption by redesigning priority workflows, standardizing infrastructure, training people and measuring outcomes. Scale working patterns across the company rather than multiplying isolated pilots.

An AI agent is software that can pursue a goal by reasoning through steps, using tools and taking actions rather than only returning an answer. The important business question is not whether something is called an agent, but how much authority it has and where human escalation is required.

Traditional automation follows predefined rules; an AI agent can interpret context, choose among actions and adapt its next step. That flexibility creates more leverage—and more need for boundaries, monitoring and escalation.

Use agents for repeatable multi-step work where goals, tools, permissions and escalation rules can be defined clearly. Keep open-ended judgment, sensitive relationships and consequential exceptions in the Human 20%.

Give an AI agent only the authority needed to complete a clearly defined task, with tighter limits as consequences increase. Permissions, spending limits, approved actions and escalation triggers should be explicit before autonomy expands.

Yes, in bounded situations with approved rules, identity, tone, limits and escalation—but not as unrestricted autonomy. External communications can create reputational and legal consequences, so the system should distinguish routine interactions from situations requiring human ownership.

AI agents can take transactional actions within explicit limits, but high-value, unusual or irreversible actions should trigger human approval. The control design should reflect dollar value, reversibility, fraud risk and customer consequence.

Monitor AI agents by logging actions, measuring outcomes, setting hard permissions and defining exception triggers. The agent should know when confidence is low, rules conflict or consequences exceed its authority, and then escalate to a named human owner.

You cannot reliably know from confidence or tone alone; verification must be designed into the workflow. Use trusted sources, deterministic checks, comparisons, sampling and human review where errors matter.

Reduce hallucinations by grounding AI in reliable sources, narrowing the task, requiring citations or evidence, and validating outputs. For high-consequence work, never make fluency a substitute for verification.

Good AI governance matches the control to the consequence. Low-risk work should move quickly; high-risk, irreversible or sensitive work needs stronger testing, oversight, documentation and named accountability.

The company and designated human owners remain responsible for decisions and actions taken through AI. Accountability cannot be delegated to a model; every consequential workflow should have a clear human owner.

AI governance should define approved tools, data rules, risk levels, permissions, review requirements, monitoring and accountable owners. Governance should make safe use easier, not bury adoption under unnecessary approvals.

Manage AI bias by testing outcomes across relevant groups, reviewing inputs and criteria, and preserving human review for consequential decisions. A consistent automated process can still be consistently unfair.

AI use becomes high risk when errors can materially affect safety, rights, employment, money, legal exposure, privacy or reputation. Those uses should have stricter testing, approvals, monitoring and human accountability.

Only give an AI system data that the company is authorized to use under that provider’s privacy, security and retention terms. Data classification and approved-tool policies should determine what is allowed, not employee guesswork.

Do not put secrets, regulated data, personal information or confidential customer/company information into unapproved public AI tools. Use approved enterprise systems with appropriate contractual, access and retention controls for sensitive work.

Treat AI as another powerful system that needs data classification, identity, access control, vendor review and monitoring. The productivity benefit does not remove normal security and privacy obligations.

For business use involving company information, enterprise-controlled accounts are usually safer than unmanaged personal accounts. Central controls help manage access, data handling, retention, offboarding and auditability.

Shadow AI is employee use of AI tools or workflows that the organization has not approved or cannot govern. Reduce it by giving employees useful approved alternatives, clear rules and fast pathways for requesting new tools.

Give AI the minimum system access required for its task and separate read, write, financial and external-communication permissions. Expand permissions only after the workflow proves reliable and monitoring is in place.

Assume AI makes impersonation cheaper; use independent verification for unusual payments, credential changes and high-risk requests. Human verification becomes more important when voice, video or written communications can be generated convincingly.

Measure AI by business outcomes, not by how often people use it. Establish a pre-AI baseline and track time, cost, quality, speed, revenue or capacity created by the redesigned workflow.

No. Headcount reduction is only one possible outcome and often the wrong primary AI metric. Measure output, cycle time, quality, revenue, cost, customer experience and the amount of human capacity moved to higher-value work.

Output per employee can be useful, but it should be paired with quality and business-outcome measures. More output is not value if it creates errors, rework, noise or low-quality customer experiences.

Measure the workflow before AI: time, cost, volume, quality, error rate and business outcome. Without a baseline, AI ROI becomes a collection of anecdotes rather than a management metric.

Judge AI cost against the economic value of the workflow, not against token or subscription cost alone. Include model usage, software, integration, monitoring, human review and error costs, then compare them with time saved, capacity created and business outcomes.

AI is improving the business when measurable outcomes improve after accounting for new review, error and technology costs. Track value at the workflow level rather than celebrating adoption volume.

AI is more likely to replace or reshape many tasks inside jobs work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

Yes. AI is more naturally analyzed at the task level because most jobs combine work that is easy to automate with work that remains human. The Dermer Rule focuses on redesigning the work inside a role rather than assuming the entire occupation disappears.

AI is more likely to replace or reshape many tasks inside Which jobs are most exposed to AI work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

Roles with high physical complexity, trust, accountability, human relationships and contextual judgment are generally harder to automate end to end. But nearly every role will still have tasks that AI can absorb.

Judgment, communication, trust, leadership, problem framing, taste, domain expertise and accountability become more valuable as routine production gets cheaper. AI literacy matters too: people need to know how to direct, verify and challenge AI.

AI is more likely to replace or reshape many tasks inside entry-level jobs work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside middle management work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside knowledge workers work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside freelancers work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside content creators work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside consultants work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside salespeople work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside marketers work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside accountants work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside lawyers work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside software developers work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside customer service representatives work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

AI is more likely to replace or reshape many tasks inside How should people redesign their jobs around AI work than to make the human role disappear all at once. Routine research, analysis, drafting and administration move toward AI; judgment, trust, context, exceptions and accountability become a larger share of the human role.

Teach employees to map tasks, prompt clearly, verify outputs, protect data, use approved tools and recognize when human judgment is required. The goal is not prompt tricks; it is learning how to redesign work safely.

Treat AI adoption as work redesign and change management, not software deployment. Set clear expectations, provide approved tools and training, redesign goals and roles, and measure business outcomes.

Companies can require appropriate AI use where it is part of the job, but they should define approved use cases, tools and quality standards. Mandating vague 'AI use' encourages performative adoption; redesigning specific workflows creates value.

Disclosure should depend on context: internal routine work may not need it, while customer-facing, regulated, creative-authorship or high-stakes uses may. The policy should be clear enough that employees do not have to guess.

Preserve human skill by requiring people to understand the work, verify AI output and periodically perform or test critical capabilities without AI. Oversight becomes meaningless if employees no longer know enough to recognize an error.

Hire a person when the work requires durable ownership, judgment, relationships, leadership, physical execution or frequent high-consequence exceptions. Use AI when the work is predominantly repeatable, scalable and measurable.

AI transformation should have executive ownership but cannot belong only to IT. Business leaders must own workflow outcomes, technology leaders the platform, risk leaders the controls and people leaders the role redesign.

Some companies need a dedicated AI leader; others are better served by clear executive ownership embedded across business, technology, risk and people. The title matters less than having authority, accountability and a company-wide operating model.

Prompting matters, but context, examples, workflow design and verification matter more. The goal is repeatable quality, not impressive one-off prompting.

A good business prompt states the objective, relevant context, source material, constraints, desired output and what 'good' looks like. When possible, give examples and tell the model what to do when information is missing.

Give AI the context that can change the answer—not every document you can find. Relevant goals, facts, examples, constraints and definitions improve output; irrelevant context can dilute attention and increase cost.

Yes. Too much irrelevant or conflicting context can make AI output worse. Curate the information the task actually requires and keep authoritative sources distinguishable from background material.

Yes. Examples are one of the best ways to communicate quality, format, tone and judgment to AI. Use representative examples, but still verify that the model understands the underlying rule rather than merely copying surface patterns.

Give AI explicit company standards, examples, terminology, audience context and authoritative source material. Do not expect a general model to infer your strategy or brand accurately from a vague prompt.

Yes. AI becomes much more useful when it can retrieve approved company knowledge rather than rely only on general model memory. Access should be permission-aware, current and limited to information the user or workflow is allowed to use.

Organize knowledge around authoritative sources, clear ownership, current versions, useful metadata and permissions. AI cannot reliably fix a knowledge base that is contradictory, stale or impossible for humans to navigate.

Retrieval-augmented generation lets AI pull relevant information from your own approved knowledge at answer time. Use it when answers need current, proprietary or source-grounded business information rather than generic knowledge.

AI memory is useful when continuity improves the work, but business memory should be governed like other company data. Define what may persist, who can access it, how it is corrected and when it should be deleted.

Test the same task with better instructions and authoritative context before blaming the model. If multiple strong models fail with good context, the workflow may be underspecified; if one model consistently fails, model choice may be the issue.

AI can capture nearly every meeting where consent and policy allow it, but not every meeting needs recording or automation. Sensitive personnel, legal, board or confidential discussions may require different rules.

Use AI note taking where decisions, commitments or knowledge are worth preserving and participants understand the recording policy. Avoid automatic recording when privacy, sensitivity or trust outweighs the documentation benefit.

Routine notes can be lightly reviewed; material decisions, commitments and sensitive summaries should be checked. AI can confuse discussion, suggestion and final agreement.

AI can run the information layer of routine meetings—agenda, metrics, updates and follow-up—but humans should own debate, motivation and decisions. If a meeting exists only to exchange status, AI may eliminate the meeting entirely.

Yes. AI can summarize channels, identify decisions, surface unanswered questions and flag commitments. Access must respect channel permissions and sensitive conversations.

AI can draft internal announcements, but leaders should own messages involving change, culture, people or trust. The more emotionally consequential the message, the more important authentic human ownership becomes.

AI can automatically send routine customer communications inside approved rules and data. Complaints, negotiation, unusual requests and relationship-sensitive moments should escalate.

Require review when the email is high-stakes, novel, sensitive, legally meaningful or difficult to reverse. Routine low-risk communications can become autonomous after monitoring proves reliability.

Yes. AI can draft and send routine follow-ups, distribute notes and create tasks automatically. Confirm material commitments before the system represents them as agreed.

AI can create and route tasks from agreed decisions, but managers should own priorities and accountability. Automation should not quietly turn every suggestion into work.

Yes. AI can monitor commitments, deadlines and blockers and escalate exceptions. Humans should own performance conversations and decisions.

Yes. Recurring management synthesis is an ideal AI 80% workflow. AI should gather and summarize; leaders should decide what matters and what action follows.

Yes. AI can triage inbox volume, summarize threads and surface messages requiring your judgment. It should use clear importance rules and preserve privacy and sender context.

Automate routine information movement; keep human collaboration focused on alignment, trust, conflict and decisions. AI should reduce meetings and message volume, not merely summarize more of them.

No. AI should do most content production, but humans should own the point of view, lived experience, taste and reputation behind it. A company that automates the human reason to listen will produce more content and less meaning.

There is too much AI-generated content when volume rises while distinctiveness, trust or usefulness falls. Measure whether the content says something only your company or people could credibly say.

Give AI proprietary ideas, examples, stories, data, audience context and a clear point of view—not just a topic. Generic inputs produce generic content.

Disclosure should depend on platform rules, legal obligations and whether synthetic authorship or media could materially mislead the audience. Do not imply a human personally created or said something when that distinction would matter.

AI can produce and distribute thought leadership, but it cannot invent credible lived experience or a genuinely owned point of view. Humans must supply the idea; AI can scale its expression.

Yes, if the system is grounded in enough authentic writing, speech, stories, beliefs and examples. The founder should own the ideas and boundaries even when AI performs the production.

Yes, with the executive's consent and clear rules governing what the synthetic voice may say. Voice cloning should scale approved content, not manufacture positions the person never authorized.

Disclose synthetic voice when required by law or platform policy and whenever failing to disclose could materially mislead the audience. Trust is part of the Human 20%.

Yes for repeatable formats once templates, brand rules and quality checks are established. Humans should still own the underlying idea, taste and exceptions.

AI should own production, variation and distribution; humans should own point of view, story, taste and credibility. That is the Dermer Rule applied to content.

Yes, for bounded internal tools and prototypes—but production systems still need appropriate security, architecture and ownership. AI lowers the technical barrier; it does not eliminate technical consequences.

AI can build many simple business applications without a traditional developer, especially internal tools and prototypes. Complex integrations, security, scale and mission-critical systems still require technical judgment.

AI can write production-quality code, but production safety depends on review, tests, architecture, security and deployment controls. Code generation is not the same as system accountability.

Review should be risk-based: more for security-sensitive, financial, safety-critical and architecture-changing code; less for routine well-tested changes. Automated tests and static analysis can move much of the review itself into the AI 80%.

AI can handle a growing share of maintenance, debugging and routine changes if the codebase, tests and observability are strong. Humans should retain architecture, security and consequential change ownership.

Yes. Internal tools are one of the strongest places to use AI coding because workflows are known and iteration can be fast. Apply normal data, access and security standards.

Yes. AI agents and AI-generated integrations can automate work that previously required custom software. Durable workflows still need monitoring, permissions and maintainable architecture.

Use no-code when the workflow fits stable connectors and rules; use AI coding when customization or logic exceeds the platform. Choose the simplest maintainable approach that meets the requirement.

Yes. AI can materially reduce the cost of prototyping and building custom software for small businesses. Maintenance, security and integration costs still exist, so cheaper creation should not mean careless deployment.

Do not delegate architecture, security, permissions or mission-critical technical accountability entirely to AI. AI can produce most of the technical work; humans must own the system consequences.

Use AI voice for scalable, repeatable audio where authentic human recording does not materially change trust or meaning. Keep consent, disclosure, brand and authorization rules explicit.

Often yes for routine narration, localization and high-volume derivative content. Use a real human when emotion, performance, authenticity or a sensitive message is the reason the voice matters.

Use a real person when the message is personal, high-stakes, emotional, reputational or dependent on authentic presence. Use synthetic voice when the value is scalable communication rather than the act of personally speaking.

Yes. AI can create and iterate marketing imagery quickly, subject to brand, rights and truthfulness controls. Humans should own visual taste and whether the image could mislead.

AI can create product imagery, but it must not misrepresent material product characteristics. The more the image influences a purchasing decision, the more accuracy matters.

Yes. Multimodal AI can extract, compare and flag information from images or video, but high-consequence interpretations require validation. Use it aggressively for scale; preserve human judgment where errors affect safety, rights or major decisions.

AI can understand many documents, charts and screenshots well, but complex layouts, tiny labels and ambiguous visuals can still cause errors. Verify important extracted values and conclusions.

Verify factual claims, identity, product accuracy, rights and context before publishing synthetic media. Visual plausibility is not evidence that an image or video is true.

Disclose synthetic media when rules require it or when a reasonable audience could otherwise be materially misled. The standard should protect trust, not merely satisfy a technical labeling rule.