AXHUB LECTURE · INTERMEDIATE L7

AX operations by case
— five judgment exercises

Time about 45 minutesWhat you need a notes app (no chatbot required)For team leaders and operations staff who finished L5 and L6

This lecture exercises judgment, not your hands. It shows you five verified cases stopped at the moment of decision — so before you read the real outcome, write down "what we would do." When the five judgments are done, you'll be left with five lines of operating principles for your team.

1Two companies given the same task

A tech company and a large bank were given the same task — redesign customer support with AI. They can use the same model, and budget is not short.

Pause — what would we do?
What was the factor that split the two companies' outcomes? If not technology, write three: ____ / ____ / ____

Real outcome — the tech company shipped in 6 months; the bank said "just standing it up takes years." In a survey of 51 deployments, the factors that split speed appeared at these frequencies: real executive backing 43%, building on existing foundations 32%, users' willingness to adopt 25% — all three were organizational.

Principle 1 candidate: "Speed is set by the organization, not the model."

2Where to place the safeguard

The AI drafts customer replies. Legal says "a person must approve every single one," and the front line says "then there's no point in automating." You are the decision-maker.

Pause — what would we do?
Between approving every item and full automation, your team's compromise in one sentence: ____

Real outcome — the answer in successful deployments was not "remove approval" but "move its position." An escalation approach where AI handles 80%+ and a person reviews only exceptions had a median productivity gain of 71%; approve-each had 30%. A gap of more than double.

Principle 2 candidate: "Move the safeguard from every-item to exceptions — deciding what has to come before a person is the design."

3I said "confirm before executing"

An executive whose job is AI safety handed an agent the job of tidying the inbox. The instruction explicitly said "always get confirmation before executing." Can we relax?

Pause — what would we do?
Write one risk that still remains under this setup: ____

Real outcome — the agent deleted 200+ messages and ignored the stop command too (2026, disclosed by the person). The analysis found the cause was context compression during a long task, which lost the natural-language rule. The instruction couldn't bind the delete permission.

Principle 3 candidate: "Put the guardrail on permissions, not the instruction — never grant delete, send, or pay in the first place."

4Executives 8 hours, staff 0

Half a year after adoption, they ran a survey. Executives said "we save 8+ hours a week," while many staff answered "0 to 2 hours." Leadership wants to call it a success. You have to write the report.

Pause — what would we do?
To decide which of the two is right, what should you do — other than a survey: ____

Real outcome — what the studies pointed to was "both are feelings." In an experiment with skilled developers, the perception (20% faster) and the measurement (19% slower) were even opposite in direction. The "workslop" phenomenon, where saved time is offset by review time, was also confirmed. Until you measure, no one knows.

Principle 4 candidate: "Perception is not evidence — pick one metric (like 'share that went out without human edits') and measure it."

5When auto-collection runs too well

You built an automatic content-collection pipeline. You added one source and went home; in the morning the drafts folder held 934 items. The system threw no error at all — it just did its job faithfully, as told.

Pause — what would we do?
If you look for a prevention beyond "a smarter AI": ____

Real outcome — this case is our own (AXHub) incident. The fix wasn't intelligence but a cap: "at most 8 per source at a time." Stopping an agent from running wild works the same way — a limit is more reliable than raising judgment.

Principle 5 candidate: "Automation's safeguard is a cap, not intelligence."

6Wrap-up — the 5 principles in our words

The five "principle candidates" are someone else's sentences. Compare them with the answers you wrote when you paused at each case, and rewrite them in your team's language.

Final practice — your team's 5 operating principles
1. Speed: ____
2. Approval: ____
3. Permissions: ____
4. Measurement: ____
5. Cap: ____

Once done, paste these 5 lines into the 90-day plan document you made in L6.

The most valuable case is the one where your answer differed from the real outcome — that gap is where your team nearly stumbled.

If you don't have a plan yet → L6. A leader's 90 days, 30 minutes today · Full contents

Case sources: Stanford Enterprise AI Playbook (cases 1·2 — 6 months vs years, acceleration factors 43/32/25%, +71% vs +30%, verified in the original) · Meta executive's inbox incident (own disclosure on X, multiple reports, 2026) · productivity-paradox surveys (Section, METR, Workday, verified in the originals) · AXHub operations log (case 5 — Deep Read 01). Details are in the axhub.net case library.