Practical AI at work, answered.

Run both on identical representative cases with the same brief, then score output quality, attribution accuracy, review minutes, and export fit. Keep the winner's prompt and checklist.

Check where content lives, retention windows, who can access it, training-data use, deletion and export paths, and compliance attestations. Prefer explicit written terms over marketing claims.

They can be for non-sensitive drafts when policy allows, but free tiers may differ on retention, limits, and controls. Keep client and confidential work on approved accounts with clear terms.

Assign it one owning workflow, route look-alikes into it or out, and review seats quarterly against verified use. Every renewal needs a reviewer and a baseline.

On each vendor's official pricing page on the day you decide, since plans change often. Examples: meeting assistants at https://otter.ai/pricing and https://fireflies.ai/pricing; suite AI at https://www.microsoft.com/en-us/microsoft-365-copilot.

Pick the suite you already live in when it covers the task with review in place, and add a specialist only when it removes a concrete bottleneck the suite cannot. One owner per workflow beats five overlapping subscriptions. Sources: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview, https://workspace.google.com/blog/product-announcements/reimagining-content-creation

Give every tool one owning task, review usage monthly against completed work, and cancel anything without a verified reviewer and checklist. Compare total cost including seats, add-ons, and review time, and always confirm current prices on the vendor's official pricing page.

Check capture method, speaker labels, language coverage, where recordings live, retention terms, and how actions export to the task tracker. Trial it on real meetings and judge review burden, not demo fluency. Official pricing examples: https://fireflies.ai/pricing, https://otter.ai/pricing

Allow only explicitly approved tools and data, start with sanitized examples, keep a named reviewer, and store outputs where policy says. Review each vendor's data and retention terms before connecting real client material.

Feed only approved scope, pricing, and timelines, require every figure to trace to the source, and mark anything unconfirmed for human sign-off before the proposal leaves.

Ask where data lives, how long it is kept, who can see it, how to export and delete it, what review controls exist, and the total 12-month cost including seats and add-ons.

Confirm the task was permitted, the output matches the brief, every fact traces to the source, unresolved points stay visible, and a named human approves the final version. Include stop rules for incomplete sources and high-consequence decisions without domain review.

Treat the output as incomplete by default. Fix speaker labels first, verify attribution with someone who attended, and skip AI-led decisions where the record cannot support them.

Get permission first and follow organizational rules. Announce AI note-taking at the start and respect objections, especially with clients and in regulated contexts.

Abandon it when sources are incomplete, review takes longer than writing, policy forbids the data, or the decision is high-consequence without domain review. Record the SKIP reason so the team does not retry blindly.

When the task is stable, short, and fully structured, a template with blanks beats generation. Use AI where drafting, extraction, or reformatting varies each time.

Count it as unfinished, but flag its missing estimate rather than assigning zero hours. Report known estimated hours separately from tasks without estimates, and mark the complete remaining-hour total as unknown; see https://www.cihanuygur.com/blog/ai-agents-work-queue/.

Report the duplicate and ask how it should be handled before producing a final workload total. Do not silently count it twice or delete a row; see https://www.cihanuygur.com/blog/ai-agents-work-queue/.

Not from a queue that contains only task status and estimated hours: those figures describe unfinished work, not available capacity. Remove overload or staffing claims unless separate approved inputs establish capacity, priorities, and deadlines; see https://www.cihanuygur.com/blog/ai-agents-work-queue/.

Calculate the known first-12-month base cost as twelve times the quoted monthly subscription plus the one-time setup fee, showing the source labels. List unpriced extras separately and do not treat a below-budget base price as proof of all-in affordability; see https://www.cihanuygur.com/blog/ai-prepare-decision/.

Require the source label for the quoted term or calculation in every material cell, and use not specified when the source has no stated term. If a cell is unsupported, request a corrected labeled table rather than accepting an apology without a replacement artifact; see https://www.cihanuygur.com/blog/ai-prepare-decision/.

No: preserve both labeled statements and keep access status unresolved until the responsible person confirms it. A newer note is a clue to investigate, not proof that the earlier contradiction has been resolved; see https://www.cihanuygur.com/blog/ai-review-contract/.

Check that each source label maps to the correct input file and an actual supporting passage for the claim or gap. A label without matching evidence is not enough, so inspect source-index.csv rather than relying only on the chat summary; see https://www.cihanuygur.com/blog/ai-review-contract/.

Keep each quoted currency and amount exactly as stated, and convert only when an approved exchange-rate source and date are provided. Otherwise compare like-for-like amounts and mark cross-currency totals unresolved; see https://www.cihanuygur.com/blog/ai-prepare-decision/.

Keep both versions labeled with their source and date, identify the conflicting fields, and ask which version is authoritative. Do not overwrite the earlier quote or choose the newer one without confirmation; see https://www.cihanuygur.com/blog/ai-prepare-decision/.

Label the claim unsupported or UNKNOWN and place it in the open-items section for review. Do not strengthen it through polished wording or a source label that does not point to evidence; see https://www.cihanuygur.com/blog/ai-review-contract/.

Keep it as a draft when decisions, owners, or dates still need confirmation, or when the transcript is partial or ambiguous. A concise brief is not ready for action merely because it sounds confident; see https://www.cihanuygur.com/blog/ai-meeting-decision-brief/.

Write the trigger, outcome, permitted context, tools, permissions, evidence required, stop condition, and human owner. If those boundaries are unclear, automation only moves the ambiguity into the model's output; see https://www.cihanuygur.com/blog/ai-workflow-operating-system/.

Ask for the loaded column names, row count, delimiter, empty rows, and any parsing warnings before requesting totals or conclusions. This makes a shifted delimiter or malformed header visible; see https://www.cihanuygur.com/blog/chatgpt-csv-input-troubleshooting/.

Compare the reported column names and sample rows with the original file, especially when the header uses semicolons, tabs, or another separator. If several fields appear as one column, fix or re-export the file before asking for analysis; see https://www.cihanuygur.com/blog/chatgpt-csv-input-troubleshooting/.

Put the approved brief and the draft in separate labelled sections, then ask for a requirement-by-requirement comparison before requesting a rewrite. Use fictional or approved material and review every proposed change; see https://www.cihanuygur.com/blog/chatgpt-review-draft-brief/.

Quote the conflicting draft text, point to the relevant brief requirement, and flag the conflict for a human decision. Do not silently rewrite a substantive commitment or invent a compromise; see https://www.cihanuygur.com/blog/chatgpt-review-draft-brief/.

Duplicate status depends on which fields define the same business record, and the model should not guess that identity. State the fields explicitly, such as email, company, and amount, and review the assumption; see https://www.cihanuygur.com/blog/claude-code-csv-duplicates/.

Write results to a new path and refuse to overwrite the input by default. Keeping the source unchanged preserves a recovery point while you review the output and tests; see https://www.cihanuygur.com/blog/claude-code-csv-duplicates/.

It should name the missing column and stop before producing a misleading result. A clear input error is safer than silently treating an incomplete file as valid; see https://www.cihanuygur.com/blog/claude-code-csv-duplicates/.

The trigger is the event or input that starts the work; the outcome is the observable artifact the workflow must produce. Naming both prevents a vague goal such as ‘use AI for meetings’ from becoming an untestable process; see https://www.cihanuygur.com/blog/ai-workflow-operating-system/.

Give it only the context and permissions needed for one bounded task, and keep the action or sending step behind human review. Expand access only when the workflow and exception path are understood; see https://www.cihanuygur.com/blog/ai-workflow-operating-system/.

Choose a recurring job that is valuable, repeatable, and safe enough to review, then define its outcome and baseline before selecting a tool. Avoid broad transformation projects with no named job or accountable owner; see https://www.cihanuygur.com/blog/ai-workflow-operating-system/.

Keep missing, blank, and explicit zero as separate states during parsing and reporting. A missing estimate is unknown; it must not be silently converted to zero and included in a calculated total; see https://www.cihanuygur.com/blog/chatgpt-csv-input-troubleshooting/.

Recalculate a sample or the full total outside the model using the confirmed columns and compare it with the reported result. Review the formulas, exclusions, assumptions, and row count before treating the calculation as an approved report; see https://www.cihanuygur.com/blog/chatgpt-csv-input-troubleshooting/

Ask for themes with verbatim excerpts, interview IDs, and a separate interpretation field; prohibit reconstructed or polished quotations. Review the source recording or transcript for every quote used in a decision, message, or product claim. Sources: https://www.cihanuygur.com/blog/chatgpt-work-report-brief/

A web URL import reads that page's text; it does not crawl linked pages or import embedded videos and images. Add each relevant page as its own source, then open the imported sources to check what arrived. Source: https://support.google.com/gemininotebook/answer/16215270?hl=en&co=GENIE.Platform=Desktop

Official source

A YouTube source imports the video's caption transcript, not the on-screen demonstration, and videos without speech are not supported. For a visual-only tutorial, write the relevant steps or supply an approved document that explains them rather than expecting the video URL to capture each click. Source: https://support.google.com/gemininotebook/answer/16215270?hl=en&co=GENIE.Platform=Desktop

Official source

Google's source-import documentation says footnotes and comments from Google files are not imported. If those passages matter, copy the approved wording into a separately named text source with its document reference; do not silently fold a reviewer's suggestion into the approved document text. Source: https://support.google.com/gemininotebook/answer/16215270?hl=en&co=GENIE.Platform=Desktop

Official source

Microsoft's Outlook FAQ documents Copilot scenarios for the user's primary mailbox, not shared, delegated, group, or archive mailboxes. Check the mailbox type with your administrator before troubleshooting prompts, and do not forward shared-mailbox content to a personal account as a workaround. Source: https://support.microsoft.com/en-us/outlook/frequently-asked-questions-about-copilot-in-outlook

Official source

In Home's Enterprise Search, open All sources, turn Web search off, and select your workspace as the source. You can narrow it further to the relevant page or teamspace instead of relying only on an instruction in the question. Source: https://www.notion.com/help/enterprise-search

Official source

Not necessarily: Notion says initial connector ingestion can take up to 36 hours, depending on the amount of content. Check the connection with the workspace owner and allow for that initial ingestion before concluding that a failed first search means the app is empty. Source: https://www.notion.com/help/notion-ai-connectors

Official source

In a private ChatGPT Project, open Add sources, choose Google Drive or Slack, authorize access if prompted, and paste the file, folder, or channel link. Check the project sources list afterward so the project uses the intended reference. Source: https://help.openai.com/en/articles/10169521-deep-research-faq

When the app, account, plan, and workspace allow it, deep research can use supported connected apps' read actions as sources and does not use app write actions as part of research. Confirm provider permissions separately because connecting an app does not override them. Source: https://help.openai.com/en/articles/10500283-deep-research-faq

A prompt can produce a useful draft. A workflow makes the complete job repeatable, reviewable, and accountable: it has a trigger, an outcome, controlled context, a quality standard, and an owner.

Say what the output is for, what information to use, what shape the result should take, and what is forbidden. Name the audience, limit sources to approved notes, require time boxes and decision owners, and forbid inventing deadlines, people, or commitments.

Read the output beside its source and verify every decision, owner, date, and rationale. Watch for proposal drift, consensus drift, ownership drift, and certainty drift. Keep unresolved points visible and require the responsible person to approve the final version.

Use only data your organization explicitly allows in the selected tool. Start with fictional or sanitized examples and keep confidential, personal, client, financial, HR, and legal material out unless policy permits it. Keep a named human reviewer on every workflow.

TRY means a bounded experiment beside the existing process is worth another look. SKIP means the task, tool, or review burden does not fit. USE means your own checks support keeping it for that specific task with review still in place. Apply it per task, not per tool.

Record usual completion time, prompt preparation time, and checking and correction time across representative examples. Do not count fast generation as saving when it creates slow review, and consider whether a plain template would solve the task with less effort.

It predicts likely text from patterns learned during training, and web-enabled assistants can also retrieve current pages before answering. That is why the same question can produce a fluent answer that still needs checking against cited sources.

Hallucination means fluent invented details such as names, dates, owners, or commitments. Reduce it by limiting sources to approved material, demanding uncertainty flags instead of guesses, and verifying every important fact against the source before sharing.

Start from the task and the approved data path, not from model hype. Use the workplace-approved stack, test one recurring task against your current process, and keep the tool whose complete workflow, including review, earns its place.

Controlled context means the model may use only the approved notes you supply, nothing else. It matters because open-ended inputs invite invented facts, while bounded inputs make checking fast and ownership clear.

Pick one task, one baseline, one approved tool, and one reviewer. Run the AI version beside the current process, log corrections, and decide TRY, SKIP, or USE with explicit stop rules.

Allowed tools and data, who may approve what, where outputs are stored, and how review is recorded. Check your own organizational policy rather than copying a generic template.

List the day's tasks with durations, let the planner propose slots around fixed meetings, then approve the plan personally and protect one focus block. Review scheduled-versus-done weekly, not daily.

Audit recurring meetings quarterly, require agendas for kept ones, batch 1:1s, and let auto-schedulers defend two focus blocks weekly. Fewer meetings beat smarter scheduling of too many.

Display dual zones, rotate painful slots across regions, and record decisions in writing for absent members. Fairness is a policy, not a feature.

One verified handoff such as form-to-tracker or meeting-notes-to-tasks with error alerts and a human reviewer. Prove one flow monthly before automating anything customer-facing.

Google says AI Inbox uses messages in the Primary tab and excludes spam, archived, deleted, muted, or snoozed messages; it also does not support attachments, delegated inboxes, or encrypted emails. Check those boundaries before treating the list as complete. Source: https://support.google.com/mail/answer/16845247

Describe the Gem's persona, task, context, and desired format. You do not need all four every time, but these fields make the repeatable job and expected artifact clearer. Source: https://support.google.com/gemini/answer/15235603?hl=en

No. Use the preview to inspect the behavior, then click Save to create the Gem. A preview is a trial of the instructions, not a saved workflow or evidence that the resulting output is correct. Source: https://support.google.com/gemini/answer/15235603?hl=en

Place the cursor in the text or highlight the section you want to change, select Ask in the toolbar, and request a focused rewrite, expansion, simplification, or grammar edit. Review the result in context before keeping it. Source: https://support.microsoft.com/en-us/microsoft-copilot/using-copilot-pages

Open Compose, enter a short prompt in the Help me write bar, and select Create. State the recipient, purpose, required action, tone, and any specific email or file that supplies facts, then verify the draft before sending. Source: https://support.google.com/mail/answer/13955415

Yes. Google's Docs Gemini guidance includes referencing other files to edit or refine content. Select the relevant approved file, state the exact change you want, and compare the result with both the working document and the reference before accepting it. Source: https://support.google.com/docs/answer/14206696

In Copilot Chat, select the plus button beside the prompt, choose More, then Draft page or code. Describe the durable artifact you need, and review the resulting Page directly before sharing or using it as a team document. Source: https://support.microsoft.com/en-us/microsoft-365-copilot/get-started-with-microsoft-365-copilot-pages

In an eligible Sheets account, enter an AI function such as =AI("Classify this customer inquiry as a compliment, exchange request, or return request.", C2) and select Generate and Insert. Inspect representative classifications and the source range before treating the column as a dataset. Source: https://support.google.com/docs/answer/15877199

Run `claude --continue` from the project directory to resume the most recent session there. If you need to choose among sessions, use `claude --resume` or `/resume`, and confirm the repository and pending work before making changes. Source: https://code.claude.com/docs/en/tutorials

On Gmail for the web, open Settings, choose See all settings, open General, and turn Grammar suggestions off. This disables grammar suggestions, but it does not verify the facts, names, dates, or commitments in a draft. Source: https://support.google.com/mail/answer/16887555

Open the thread and click Summarize this email, then treat the AI Overview as a catch-up aid: open the messages and verify dates, owners, and commitments before acting. It requires a supported language and both settings turned on in Gmail settings, Smart features in Gmail, Chat, and Meet and Smart features in Google Workspace, and turning those off also disables other AI writing and proofreading tools. Source: https://support.google.com/mail/answer/16561387

On a computer, open a thread whose AI Overview shows a task with a deadline, click Remind me beside the task, then click Edit reminder if the suggested date needs changing. The task then appears in Google Tasks and Google Calendar, where you can mark it complete or incomplete, and the reminder shows at the top of the inbox when the date arrives and stays for 4 days or until dismissed. It requires a task with a due date in the thread plus a supported language and both Smart feature settings. Source: https://support.google.com/mail/answer/16561387

On a computer, open the sheet, click Ask Gemini, describe the formula with explicit cell or range references, and insert it only after checking the result against a known value. It needs an eligible Google Workspace or Google AI plan or Workspace Experiments access, a supported language, and works best in a native Google Sheets file, so save an Excel workbook as Google Sheets first. Insert the output into the sheet, because side panel history is lost when you reload, close the file, or go offline. Source: https://support.google.com/docs/answer/14356410

On the Copilot Page, select Share and choose Copy link to share that page as a standalone artifact, or Copy component to place a live component in another Microsoft 365 app. This lets colleagues work with the page without giving them access to your Copilot chat session. Sharing still depends on your organisation's licensing and permissions for the destination app, so check the recipient can open it. Source: https://support.microsoft.com/en-us/microsoft-365-copilot/share-a-microsoft-365-copilot-page

After Gemini proposes a refinement in Google Docs on the computer version with eligible access, review the suggested changes and choose Accept all or Reject all. Read the edited passage in context before accepting it, especially when the change affects evidence, qualifications, names, or commitments. Source: https://support.google.com/docs/answer/14206696

In Google Drive on the web, click the Catch me up badge next to a Google Doc or Sheet when it appears. Review the listed changes and comments against the document before treating the update as complete, because the feature summarizes activity since your last view rather than replacing source review. Source: https://support.google.com/drive/answer/16686465?hl=en-GB

Open Copilot in Excel and ask a specific question such as, “Are there any outliers or anomalies in this data?” Name the columns or range you want analyzed, then inspect the returned chart, summary, or explanation against the source rows before using it in a decision. Source: https://support.microsoft.com/en-us/excel/copilot/data-insights-with-copilot-in-excel

Open the project’s more-options menu, choose Project settings, and add instructions such as the audience, tone, output format, and review rules. Keep the instructions specific to that project, then test them on a small task before relying on the project for recurring work. Source: https://help.openai.com/en/articles/10169521

In a Google Doc on your computer, place the cursor where the image belongs and choose Insert → Image → Generate an image. Enter a concrete prompt with the subject, setting, distance, materials, or background, select Create, and click an image to insert it; verify that the result fits the document before sharing. Source: https://support.google.com/docs/answer/15559738?hl=en

Select the text in Word, open the Copilot menu beside the selection, and choose Visualize as a Table. Review the proposed table, then choose Keep it, Regenerate, or Discard; if it needs adjustment, describe the change in the Copilot compose box before accepting it. Source: https://support.microsoft.com/en-us/Word/copilot/rewrite-text-with-copilot-in-word

Open the spreadsheet, click Ask Gemini, and prompt it to create a pivot table that names the fields and calculation you need, such as sales by region. Review the proposed rows, values, filters, and source range against the original data before inserting it. Source: https://support.google.com/docs/answer/14356410

Open Google Slides, describe the presentation in the Gemini side panel, and add relevant files from Drive as sources. Review and edit Gemini’s outline, then approve it to generate the deck; keep the source list and slide claims available for fact-checking. Source: https://support.google.com/docs/answer/17111393

In the meeting chat, open the Recap tab, select Copilot, choose View prompts, and select Recap the meeting. Treat the result as a draft: verify mentions, decisions, and assigned action items against the meeting record before sharing them. Source: https://support.microsoft.com/en-us/teams/meetings-events/recap-a-teams-meeting

Open Copilot in Word, switch to Chat only, and ask for a summary such as five bullets focused on decisions or open questions. Review the cited passages in the document before relying on the summary, because Copilot can generate inaccurate or unsupported content. Source: https://support.microsoft.com/en-us/word/welcome-to-copilot-in-word

In the Copilot pane, open Settings, choose Personalization, enter short preferences for formats, dates, tables, formulas, or level of detail, and select Save. Check the next result in a workbook because a specific prompt can override saved preferences, which apply to your Copilot sessions rather than changing the workbook itself. Source: https://support.microsoft.com/en-us/excel/copilot/copilot-in-excel-personalization

No. A citation shows where a sentence came from, so open it and compare the exact wording with the claim before you reuse the answer. Split the reply into one row per sentence and mark each as supported, overstated, or uncited, keeping anything you cannot trace as NOT FOUND. Source: https://www.cihanuygur.com/blog/copilot-chat-citation-grounding-check/

Upload both exports, state which file is old and which is new, match rows on one key column, and name the fields that count as material. Ask it to convert dates to YYYY-MM-DD and to treat letter case and spacing as formatting, then require separate changed, added, removed, and ignored sections so a deleted row cannot vanish from the report. Source: https://www.cihanuygur.com/blog/chatgpt-data-analysis-tracker-version-diff/

Paste the open questions section into the Ask Gemini side panel and request one row per numbered question with the exact wording, the decision needed, the owner, and a due date. Require NOT STATED wherever the draft names nobody, insert the table into the document so the result survives a refresh, and then ask the draft's authors to fill the gaps. Source: https://www.cihanuygur.com/blog/gemini-docs-open-questions-owner-table/

A project keeps the files you added, so an older month's export stays available until you remove it. Use a dated filename, delete the previous month from the project sources before uploading the new file, and make each prompt print the reporting month and source filename first. Source: https://www.cihanuygur.com/blog/chatgpt-projects-monthly-report-brief/

Google states that side panel history is lost when you refresh the browser, close and reopen the document, or go offline. Insert the output into the document while you work, because the inserted text is what survives. Source: https://support.google.com/docs/answer/14206696

Dictate real vocabulary including client names, acronyms, and noisy samples, then score word accuracy, formatting, and correction effort. Trial on the team's worst audio, not the vendor's demo clip.

Bots cover every platform visibly but may be blocked by client IT. Bot-free device capture avoids guests yet covers only meetings attended. Native platform AI is cleanest for compliance but locked to one suite. Pick by room, not by feature list.

Separate periods for audio, transcripts, summaries, and integration copies, each tied to a purpose with deletion tested. Avoid forever defaults; some meetings deserve zero capture.

Read the exact tier's terms for training and retention separately, confirm the live admin setting, and record both with dates. A headline no can hide narrower scope in the details.

Delete a fictional pilot record through the UI, then check search, exports, CRM copies, and backups. Deleting the vendor transcript alone does not clean downstream copies.

Chat is the baseline secure assistant while the full experience grounds in organizational data across apps with deeper integration. Confirm current licensing and grounding behavior on official Learn pages before planning. Sources: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview

When drafts, sheets, slides, and Drive context already live in Google apps and the task needs that context with controls. Verify current features on the official Workspace blog first. Sources: https://workspace.google.com/blog/product-announcements/reimagining-content-creation

Require primary sources with dates for every claim, open each citation, and drop anything unverifiable. A brief with three verified sources beats ten confident links.

Structure notes and criteria first, review for bias and accuracy, and keep decisions with documented human rationale. Never automate rejections without review and policy.

Check the vendor's official pricing page on decision day and screenshot it with the date. Plans change; memory does not update. Examples: https://otter.ai/pricing, https://fireflies.ai/pricing

Buy when a vendor covers the task with acceptable data terms. Build only when the workflow is proprietary, repeated, and cheaper to own after review costs. Prototype on real cases first.

Describe the outcome and constraints, choose the allowed sources, and review the proposed plan before research begins. Modify the scope or source list first; the plan is the checkpoint that prevents a polished report from answering the wrong question. Source: https://help.openai.com/en/articles/10500283-deep-research-faq

In the research prompt window, use Sites to manage the allowed domains. Restrict research to those sites when the source boundary matters; use Prioritize these sites when they should lead but broader search is still acceptable. Source: https://help.openai.com/en/articles/10500283-deep-research-faq

Start Claude Code with `claude --permission-mode plan`, or switch to plan mode with Shift+Tab until the status shows plan mode on. Review the proposed steps before approving edits, especially when the task could touch many files. Source: https://code.claude.com/docs/en/tutorials

Ask it to inspect the project’s current test files, identify uncovered behavior, and add cases for edge conditions in the existing style. Run the tests and fix failures, then review whether the assertions prove the behavior you actually need rather than only increasing coverage. Source: https://code.claude.com/docs/en/tutorials

Check the engagement contract for AI language first, classify material as public, internal, confidential, or regulated, and verify the tool tier's training and retention settings before any client file touches it. Silence in a contract is not permission.

Search the signed NDA, MSA, or engagement letter for AI, automated processing, third-party tool, and subprocessor language. If it says nothing about AI, treat that as unresolved and clarify before proceeding.

Stop processing, preserve an internal record, identify exactly what was exposed, check the vendor deletion path, and notify the accountable person. Decide separately whether the client must be told, and write this checklist before an incident.

Announce capture, confirm the engagement permits it, define retained artifacts and access, and offer a no-record path when declined. A bot-free tool still needs consent and retention rules.

As a drafting aid with guardrails: no client-identifying facts in public tools, and every citation and quote verified against the primary source before filing. Courts have sanctioned lawyers for AI-invented citations, so verification is the workflow.

Keep humans on judgments affecting employment, verify outputs for bias and accuracy, limit data to what policy allows, and document reviewer and rationale. High-impact people decisions need human ownership and records.

Treat patient information as regulated: use only compliant accounts with signed agreements where required, minimize data, and keep clinical judgment fully human. Confirm terms with qualified counsel before any deployment.

Source material, brief version, reviewer, approval path, workspace, and next review date per workflow. Evidence turns activity into something defensible.

Exclude it by default until data terms, access controls, and review paths are vetted for that matter. Speed never justifies ungoverned exposure of deal or dispute material.

Drafting hours saved, correction minutes added, reviewer findings per output, and missed follow-ups before and after. Keep the workflow only when the complete job improves.

A chatbot answers when asked. An agent carries multi-step tasks across apps with permissions, such as drafting, filing, and follow-ups. Give agents narrow scopes, approval gates, and audit trails, and keep high-consequence actions human-approved.

Only for narrow, reversible, low-risk steps with explicit approval gates and logs. Keep sends, payments, filings, and external commitments human-approved while agents prove reliability on supervised tasks.

Grounding ties answers to permitted sources such as files, meetings, and mail with citations. It matters because ungrounded answers invent with confidence while grounded ones can be checked.

Well-built workplace AI inherits existing permissions, so people see only what they may access. Verify this behavior on the official docs before connecting sensitive libraries. Sources: https://learn.microsoft.com/en-us/microsoft-365/copilot/microsoft-365-copilot-overview

The task, the brief version, the source, what drifted, who was affected, the correction, and the checklist change. File it where the team will find it next time.

Personal-tier tool use on work material without approval. Replace with one governed alternative, migrate real workflows, publish the data rule, and measure verified adoption.

Keyboard flows, screen-reader behavior, caption quality, voice alternatives, and admin controls for accommodations. Trial with the people who need them, not for them.