The previous version told fictional stories and claimed that only one job in its list had been eliminated by automation. That was not a defensible basis for career advice.
A better approach is to look for task patterns. The same pattern can appear in an administrator's, analyst's, manager's or developer's job.
Twelve patterns worth examining
- Drafting from supplied material: producing a first version of an email, report or proposal.
- Summarising: compressing meetings, documents or cases into key points.
- Classification: assigning categories, priorities or routing labels.
- Extraction: pulling named fields from documents or messages.
- Comparison: identifying differences across products, policies or versions.
- Translation and rewriting: changing language, reading level, tone or format.
- Search and synthesis: gathering evidence across multiple sources.
- Pattern detection: highlighting unusual records, trends or omissions.
- Code assistance: explaining, generating or testing small pieces of code.
- Workflow orchestration: choosing or triggering tools across several steps.
- Customer interaction: drafting responses or handling bounded routine requests.
- Decision preparation: assembling options and evidence for a responsible person.
These patterns describe capability, not a guaranteed outcome. A technically possible task may remain human-led because of poor data, regulation, cost, trust or accountability.
Exposure is not replacement
The ILO's 2025 global index combines task-level data, worker input, expert discussion and model-based scoring. It found that one in four workers were in occupations with some generative-AI exposure, but only a smaller share sat in its highest exposure category. Transformation was judged more likely overall than replacement (ILO).
The OECD has also studied changing skill demand in occupations highly exposed to AI. Its evidence reinforces an important point: exposure reaches educated professional work, including managers, accountants, developers and HR professionals. It is not confined to supposedly routine low-skill jobs (OECD).
Score a task in context
For each task, ask:
- Are the inputs digital and consistently available?
- Can the expected output be described clearly?
- Can a reviewer detect a wrong result?
- Is the cost of error acceptable?
- Does the tool have approved access to the data?
- Is the responsible owner still clear?
- Would automating it remove useful human contact or learning?
A task with a clean input and obvious acceptance test is different from a negotiation where the unspoken context is the work.
Track how the job changes
Repeat the review every quarter or when a relevant tool, policy or process changes. Record:
- tasks newly assisted or automated
- new checking and exception work
- skills becoming more important
- responsibilities that moved between people
- evidence of benefit or harm
Automation can remove an activity while creating monitoring, escalation and governance work elsewhere. Count the whole process.
The proof boundary
This list helps structure a local job review. It does not predict headcount, salary, redundancy or the timing of adoption in a particular employer. Those depend on decisions and constraints no general article can observe.
For a practical place to compare these patterns with real Microsoft workplace use, join Microsoft Copilot Adopters.