Prioritize life, work, and cognitive load

One responsible-looking AI failure starts with a rollout.

New tools arrive. Teams try things. Managers ask for AI-assisted drafts, summaries, analyses, tickets, reports, and recommendations. Updates get longer. Demos multiply. Output rises.

Then the hidden bill arrives.

Engineers face more agent-written pull requests than they can inspect. Product managers read summaries and still return to source material. Support leaders check AI-drafted replies after hours because edge cases are risky. Sales managers receive more account research than they can coach against. Executives get more dashboards and less clarity.

Nobody meant to exhaust the organization. Everyone wanted leverage. But the company automated production without designing absorption. It created more things to read, review, interpret, correct, route, and decide.

That is not transformation. That is cognitive load with a software budget.

Life is part of the system

In the original book, I wrote about prioritizing life and showing care for people. I meant it then. I mean it more now.

Human capacity is not a sentimental side topic. It is one of the conditions that determines whether real work succeeds. A product team can only absorb so many priorities before roadmap quality falls. Engineering can only take so much context switching before defects rise. Customer success can only carry so many escalations before the experience gets uneven.

Good operators watch load — not just hours or task count, but cognitive load: the effort required to understand the work, hold context, make decisions, coordinate with others, and recover from interruption.

A person can have a manageable task list and still be overloaded if every task requires context reconstruction, cross-functional negotiation, unclear decision rights, and constant notifications.

Care is not avoiding standards. Care is designing work so people can meet high standards without burning through judgment as if it were disposable.

AI can move the load, not remove it

AI can reduce busywork when used well. It can summarize, draft, find patterns, prepare options, automate repetitive steps, and give people more room for judgment.

Used poorly, it creates another stream: another inbox, dashboard, agent to supervise, prompt to write, output to verify, meeting to discuss, exception path to manage, alert to triage.

Many AI programs measure generation and forget review. They measure time to draft and forget time to validate. They measure tickets created and forget tickets closed. They measure adoption and forget whether the work feels clearer to the people responsible for the outcome.

The hidden math is plain:

AI output volume times verification time times reviewer scarcity equals operating burden.

Take an illustrative case. If an agent produces fifty recommendations a week and each takes ten minutes to verify, the company created more than eight hours of review work. If the reviewer is the scarce expert, AI may have increased pressure on the person it was supposed to help.

AI does not automatically reduce cognitive load. It relocates it.

Sometimes it moves load from preparation to review, from execution to exception handling, from one department to another, or from today to technical debt and cleanup later. A serious leader asks where the load went.

Multi-agent setups make this easier to miss. When one agent hands work to another, each step can look efficient while the person at the end inherits the combined output, plus the job of working out which agent got what wrong.

Reduce total load, not local effort

An AI-enabled workflow is not improved unless it reduces total cognitive load while protecting outcome quality and human accountability.

A tool can make one person faster while making three people slower. A chatbot can reduce ticket volume while creating customer confusion. An agent can produce analysis quickly while forcing a senior leader to check every assumption.

Local efficiency is not enough. The question is whether the whole system became easier to operate. Did the right people get more focus? Did decisions become clearer? Did quality hold or improve? Did review become manageable? Did customers get a better outcome?

The leader’s job is not to protect people from hard work. Hard work is part of building. The job is to prevent unnecessary load: avoidable confusion, rework, interruption, review burden, tool switching, ambiguity, and heroics caused by poor design.

Cognitive-load audit for AI workflows

Before scaling an AI tool or agent, run a one-page audit.

1. Outcome and workflow

Name the outcome. Map the trigger, inputs, people, systems, decisions, outputs, exceptions, and owner. If the team cannot describe the workflow, do not scale the agent.

2. Current cognitive load

Find the real friction: context switching, unclear priorities, manual reconstruction, duplicate review, tool hopping, waiting, ambiguous quality standards, or escalation confusion.

3. Proposed AI role

Define the role precisely: draft, summarize, classify, recommend, monitor, route, search, update, or execute. Also define what AI will not do.

4. Review and verification burden

Who checks output? How often? Against what standard? How long does it take? Which checks can run as automated evals, and which need a person? What check decides whether output is ready to influence a customer, product decision, code path, or commitment?

5. Interruptions and alerts

List every notification, ticket, dashboard, report, and meeting the workflow will create or change. A useful alert changes behavior. A useless alert trains people to ignore the system.

6. Decision rights and accountability

Clarify who recommends, approves, overrides, and owns the outcome when something goes wrong.

7. Net-load score

• Reduce: Total load goes down while quality holds or improves.
• Shift: Load moves and must be planned for.
• Increase: Review, coordination, interruption, or ambiguity rises.
• Unknown: Run a bounded pilot before scaling.

A short example

A CEO asks for an AI briefing agent that compiles weekly notes from sales, product, support, and engineering. The first version produces a long Monday report with pipeline changes, churn risks, escalations, roadmap issues, incidents, and competitive notes.

It creates information faster. It does not create a better leadership system. The report is too long, source links are uneven, facts mix with interpretation, and the meeting becomes an accuracy debate.

A cognitive-load audit narrows the job. The outcome becomes: help the leadership team identify cross-functional risks and make weekly decisions. The agent produces a decision brief with five sections: material changes, risks requiring executive attention, decisions needed, source-linked evidence, and unresolved uncertainties. Each function owns source quality. The meeting starts with decisions, not narration.

The company stopped asking AI to produce more information and started asking it to reduce the load required to make better decisions.

Chapter 5 operating check

Before expanding an AI workflow, ask the people who will live inside it:

1. What outcome should improve?
2. What work creates the most cognitive load today?
3. What specific role will AI play?
4. What will AI explicitly not do?
5. Who reviews output, and how much review capacity exists?
6. What quality standard will reviewers use?
7. What new alerts, reports, messages, meetings, or dashboards appear?
8. Which signals will change a decision or action?
9. Where might load shift to another team?
10. Which expert could become the bottleneck?
11. Who owns exceptions and corrections?
12. How will we know after thirty or sixty days whether total load went down?
13. What should be paused, simplified, or retired if this scales?

AI should give people more room for judgment, not consume the judgment it was meant to support.