Original LinkedIn edition published
These are original companion notes. Read the full newsletter on LinkedIn
The short answer
When AI increases the volume of work, map the decisions that need human review, define the evidence required at each checkpoint and check whether reviewers have enough time, knowledge and authority. Throughput alone cannot establish whether the workflow is delivering useful, reliable results.
Plan review capacity before increasing output
Anthropic's October study examines changes inside its own company using several sources, including employment records, tool telemetry and hiring scorecards. It describes growing agent activity and more work devoted to specifying, reviewing and monitoring agents. This is evidence from a frontier AI firm with unusual tools and incentives. Adoption was not random, and agent runtime and estimated human working hours are different measures. The findings support questions about work design rather than a general forecast of job losses or productivity gains.
For an employer, a useful unit of analysis is the decision between producing an output and acting on it. Identify who checks source quality, what would trigger rejection and which consequences require escalation. Then measure the review queue, corrections and downstream rework. If generation accelerates while review capacity stays fixed, a growing backlog or hurried approvals may hide the cost of the change. A checkpoint is useful only if its owner can make an informed decision and stop or change the next step.
The Sova-delivered Core-6 is a generic assessment of foundational AI knowledge. It can inform a common baseline and development needs, but it does not measure a person's performance in a particular workflow. Employers can separately observe a relevant task or discuss a case using consistent criteria. Pair that evidence with coaching in how experienced colleagues evaluate uncertainty, so junior staff can develop judgement. The edition's Judgement Bottleneck Review, like A/A/H and PRIME, is a separate companion resource rather than a delivered or scored Core-6 module.
Illustrative example: an L&D content review queue
An L&D team uses an agent to draft twenty learning guides in the time previously needed for two. Its subject expert can still review only four carefully each week. The team separates factual accuracy checks from editorial checks, names the source material and reserves specialist approval for claims with material consequences. It limits the next batch to the team's review capacity, records recurring errors and lets junior colleagues practise checking drafts with feedback before taking responsibility for approval.
Questions to take into your next meeting
- Map where generated work waits for a person to accept, correct, reject or escalate it.
- Define the evidence, knowledge and authority needed at each consequential decision.
- Compare review capacity with output volume, including time spent correcting downstream errors.
- Give less experienced colleagues supervised opportunities to explain and improve their judgement.
Put the idea to work
Follow one batch through the review process
Choose a recent batch of AI-generated work. Record how much was produced, how long it waited for review, what reviewers checked and what needed correction after release. Agree one change to the workflow, such as a clearer acceptance rule, a smaller batch or protected review time, then compare the next batch before expanding production.
If your team doubled AI output tomorrow, which decisions would receive less attention and how would you know?
Sources and further context
- Anthropic: Employment at Anthropic Primary company study discussed in this edition; its observational findings concern Anthropic's own workforce and do not establish economy-wide effects
- European Commission Joint Research Centre: AI and Skills The synthesis cited in the edition provides broader context on AI exposure, labour-market evidence and applied learning needs
Core-6 measures common AI knowledge through Sova. The activities discussed here are separate ways to explore work and behaviour.
What evidence belongs in an AI literacy programme? Try the separate A/A/H and PRIME worksheet See what Core-6 measures