AI & Digital Work
Making technology useful, with human judgment at the centre.
Read practical guidance on using AI at work with clear responsibilities and meaningful human checks. This section focuses on the task, the information being used and the evidence needed to trust a result. It is written for professionals improving research, reporting and knowledge workflows, including teams that need a workable fallback when an AI-assisted step fails.
Start with the work you need to do
Reader worksheets →Page 1 of 1 · Articles 1–24 of 24. Browse the numbered pages to explore the complete collection.
How do you check an AI summary for missing information?
Check what an AI summary says and what it leaves out, using a completed fictional briefing, an omission map and a corrected staff update.
Can 95% AI accuracy still miss important requests?
A worked 200-request example shows why high AI accuracy can hide missed exceptions, and how precision, recall and review capacity change the decision.
An AI incident record should separate evidence from explanation
A completed AI incident response record that separates observed evidence from hypotheses, calculates two exposure denominators and defines containment and recovery checks.
The AI demo ends before the contract risk does
Move beyond the AI demo with buyer requirements, data-use boundaries and acceptance evidence. Australian guidance and model-clause scope explained.
A human shortlist can still hide automated rejection
UK hiring teams can review a shortlist while leaving rejected applicants without meaningful consideration. Examine each outcome and its evidence before describing the tool as decision support.
An AI pilot is already a privacy decision
Assess a New Zealand AI pilot before personal information enters it. Map data flows, current IPP3A duties, evidence gaps and conditions for approval.
An AI governance decision log in practice
A fictional US customer-support pilot shows how one AI decision record can connect scope, evidence, conditions and an accountable approver.
Agentic AI needs a stopping rule before it needs more autonomy
An AI agent can turn a bad assumption into a sequence of actions. Singapore’s framework points organisations toward bounded powers, accountable humans, technical controls and end-user responsibility.
An invoice total can look right and still be the wrong number
Keep original invoice evidence, validate extracted amounts and currencies, and route uncertain fields to review before downstream AI or payment use.
Synthetic data still needs a privacy argument
Synthetic does not automatically mean anonymous. Define the use, source permissions, disclosure testing and release limits before sharing generated datasets.
“We use AI” leaves the important privacy questions unanswered
Explain what an AI service does with personal information. Match purposes, sharing, retention and human review statements to the actual Canadian service workflow.
AI-written alt text can sound fluent and still miss the point
The same image can need different alternative text depending on what it does on the page. Review the purpose before accepting generated wording.
Build an AI system inventory that supports EU AI Act decisions
Record AI use, roles, purpose, affected people and current status before deciding how the EU AI Act applies.
Check that withdrawn documents leave the AI retrieval index
Trace document withdrawals through the retrieval index and verify that an assistant stops using obsolete material for current answers.
Recheck the workflow when an AI model changes
Compare the changed system against a known task baseline, separate critical failures from cosmetic differences and document a release or hold decision.
Build an AI literacy exercise around the job
Use a realistic work task to identify what staff need to recognise, check and escalate, then retain proportionate evidence of the learning.
Keep document-reading AI inside its authorised task
Treat retrieved documents as untrusted input. Bound permissions, review consequential actions and test whether an assistant stays within its authorised task.
Map where workplace AI data remains after a chat is deleted
Trace prompts, uploads, logs and exports across an AI workflow before deciding retention periods or claiming that deleting a chat removes every copy.
Make human handoff work when an AI workflow stops
Define who receives an AI escalation, what information they need and how work pauses when the human review queue has no capacity.
Test an AI generated spreadsheet formula before using the result
Check an AI generated formula against known answers, boundary cases and workbook structure before using its result in an operational decision.
Verify decisions before circulating AI meeting notes
Separate discussion, proposals and agreed actions in AI meeting notes, then confirm owners and deadlines before circulating the approved record.
Reduce sensitive data before an AI drafting task
Create a minimum input copy for an AI drafting task, check hidden identifiers and record the permitted destination before uploading workplace information.
Test AI citations before releasing a workplace knowledge assistant
A claim-level test method for a workplace knowledge assistant, with separate checks for evidence support, document currency and access permissions.
Build an AI assisted workflow with meaningful human checks
Pilot AI assistance on a bounded operational task with approved inputs, source checks, measured review effort and a clear route to stop or correct errors.
Review the workflow as well as the answer
Begin with a defined use case and approved inputs. Decide which information may enter the workflow, who reviews the output and which decisions require a person with the relevant knowledge. The presence of a reviewer is not enough if that person cannot inspect the sources, recognise an error or stop the process.
For research and knowledge assistants, a citation should lead to evidence that supports the claim being made. Our coverage examines source checking, evaluation examples and the difference between a plausible answer and a supported one. Tests should reflect the questions users actually ask, including cases where the system should say it does not know.
Use the guides to plan a bounded pilot and record what you learn. Keep track of errors, review effort and changes to the information or task. Articles connect these practical controls with management-system responsibilities where relevant. They do not imply that using a particular tool makes a workflow accurate, confidential or suitable for every decision.