What You Can Do By The End

• Choose monitoring measures
• Design real human oversight
• Write an incident playbook
• Assess lock in and plan an exit
• Manage a model change

What Changes At Scale

• More users, more data, more edge cases
• Less attention per case
• Controls must run as routines
• Scale in steps with gates

Monitor Four Groups Of Measures

• Quality: test set, sampled review, ratings
• Safety: flagged outputs, leaks, misuse
• Cost: spend and unit cost
• Use: users, volume, speed, errors

Logs And Privacy

• Log inputs, outputs, model version, reviewer actions
• Control access and set retention
• Mask personal data where possible
• Logs enable audits and investigations

Drift

• New products, policies and customer language
• Seasonal patterns change inputs
• Re run the test set on a schedule
• A drop triggers investigation

Human Oversight That Works

• A person can understand, question and override
• Reviewers tire and start to click through
• Automation bias: trusting machine output unchecked
• Required for high risk uses under the EU AI Act

Keep Oversight Real

• Show the source next to the output
• Seed known wrong examples as an attention test
• Limit daily volume per reviewer
• Let reviewers reject without penalty
• Track how often reviewers change outputs

What Is An Incident

• Wrong answer sent to many customers
• Personal data exposed
• Biased or unfair result
• Security breach or cost spike

The Incident Playbook

• Report: one clear channel for all staff
• Severity scale and who is called
• Pause switch with a named user
• Notify legal and data protection early
• Blameless review and recorded fix

1 Detect and report
2 Pause if needed (named person)
3 Assess severity (low, medium, high)
4 Contain harm and notify
5 Communicate with affected people
6 Review without blame, fix the cause

Case: Forty Five Minutes Without A Switch

• A deployment error left old code active
• Unintended orders flooded the market
• It took about forty five minutes to stop
• Loss of around 440 million US dollars
• Lesson: practise the pause

Rehearse The Playbook

• Run a tabletop exercise once a year
• Use a realistic scenario
• Time the pause switch
• Fix the gaps you find

Vendor Lock In

• Proprietary formats and tuned prompts
• Data held in the vendor's system
• Unique features and long contracts
• Some lock in is fine if chosen on purpose

Model Change Management

• New versions can be worse on your task
• Pin to an exact model version
• Run the full test set and compare
• Review the cases that matter by hand

Staged Release And Rollback

• Release to a small share of users first
• Watch monitoring measures, then widen
• Keep the old version ready for rollback
• Record change, results and approver
• Track deprecation dates

new model -> test set -> hand review -> 5% of users
   -> watch 1 week -> 25% -> 100%
problem at any step -> roll back to previous version

Recap

• Scale in steps with gates
• Monitor quality, safety, cost and use
• Keep human oversight real
• Playbook, pause switch and rehearsal
• Test a second model, release changes in stages

Next Week: The Capstone

• Build your company AI roadmap
• 90 day plan and steering committee deck
• Rubric and review checklist

Questions And Exercise

• Write a monitoring plan for one use case
• Draft the incident playbook
• Plan a test of a second model