AI in the Workplace: A Practical Guide for Knowledge Workers

This article breaks down how knowledge workers can use AI on the recurring tasks that eat time every week, not just for one-off questions and writing tasks. It includes ready-to-use prompts, guidance for using AI safely, and steps for building AI into your workflows quickly and simply.

Most knowledge work comes with operational overhead that isn't in the job description: meetings, admin, staying informed, managing relationships.

These tasks are often the best place to start using AI, because the output is easy to check and the stakes of getting it wrong are low.

Below are 10 of those tasks, and how to prompt AI to help with each.

Note: Some AI tools need source material attached or pasted directly. Others can pull it themselves from a connected inbox, drive, or open browser tabs. The prompts below work either way.

New articles and updated notes pile up on the same topic, and it can be a real challenge to keep up with what’s happening.

AI tools can help you:

Reconcile scattered sources into one answer
Catch contradictions or version drift across documents
Summarise what has changed since the last update
Flag where sources disagree

Try this query

Information summary query

Use when:You have multiple sources and need agreement, conflicts, and what’s changed.

Role: Research analyst.
Task: Synthesise the five articles below and our internal notes on {topic}. Identify where they agree, list conflicting numbers/claims, and note what’s changed since {last_date_or_version}.
Inputs:
- Articles: {paste titles + URLs or attach files}
- Internal notes: {paste bullets or attach doc}
- Last date/version to compare against: {YYYY-MM-DD or version}
Constraints:
- Time-bound evidence: prefer sources from {start_date} to {end_date}; flag anything older.
- Evidence rules: for each major claim, cite at least 2 independent sources when possible; label news/blogs as secondary.
Output:
1) 3–5 bullet summary of consensus.
2) Table: Claim | Conflicting values/wording | Sources (with dates) | Confidence (high/med/low).
3) “What’s changed since {last_date_or_version}” as 3–7 bullets.
4) “What I’m missing / what to verify next” with 3–5 specific checks.
Verification:
- Flag any sentence that could be read more than one way.
- Ask up to 3 clarifying questions if any input is underspecified.

A status update, a recap, a check-in: most work communication repeats the same few shapes, which makes it easy for AI to write a decent first draft.

AI tools can help you:

Draft emails, updates, or documents from bullet points
Adjust tone or length for a different audience
Rewrite the same update for different readers
Catch unclear phrasing before you send

AI can help turn a conversation into a tangible log of next actions and decisions.

AI tools can help you:

Pull action items and owners out of notes or a transcript
Draft the recap right after a call
Catch commitments nobody was actually assigned
Prep an agenda from the last meeting's open items

Try this query

Meeting recap query

Use when:You have a transcript/notes and need a decision/commitment log plus a ready‑to‑send recap.

Role: Project coordinator.

Task: From the call transcript/notes below, list every commitment made, who owns it, and any date mentioned. Flag anything that sounded like a decision but was never clearly assigned. Then draft a five-sentence recap email I can send as-is or edit.

Inputs:

- Transcript/notes: {paste text or attach file}

- Meeting context: {meeting name, date, attendees if known}

Constraints:

- Be literal: only record commitments/decisions that are explicit or strongly implied; mark implied items as [Implied].

- If owner/date is missing, write “Owner: TBD” or “Date: TBD”.

Output:

1) Commitments table: Commitment | Owner | Date | Status (Confirmed/Implied) | Source line (quote short snippet).

2) “Unassigned decisions” list: Decision | Why it’s unclear | Suggested owner + next step.

3) Draft email (5 sentences): subject line + body, neutral tone, no fluff.

Verification:

- Flag any sentence in the draft email that could be read more than one way.

- Ask up to 3 clarifying questions if critical details (owner/date) are missing for high-impact items.

AI can make it easier to organise a complicated weekly schedule, juggle multiple commitments, and estimate time more accurately.

AI tools can help you:

Turn competing deadlines into an order of operations
Flag scheduling conflicts before they become a problem
Separate what's genuinely urgent from what feels urgent
Sanity-check whether a plan is realistic

Try this query

Weekly planning query

Use when:You need ranking, conflict detection, and a reality check for a 5‑day week.

Role: Operations planner.

Task: Rank my weekly work by urgency and impact, flag likely conflicts, and tell me if this is realistic for a five-day week. If not, say what to cut or push.

Inputs:

- Work items: {list each item with deadline, rough time estimate, and any dependencies}

- Constraints: {working hours/day, meetings, hard deadlines, time zones}

- Success criteria: {what “done” looks like this week}

Constraints:

- Use a simple scoring rubric: Urgency (1–5), Impact (1–5), Effort (hours). Show the score for each item.

- Assume a 5-day week with {X} productive hours/day after meetings.

Output:

1) Ranked list: Item | Urgency | Impact | Effort (h) | Score | Deadline | Dependencies.

2) “Likely conflicts” list: conflict | items involved | why it conflicts (time/dependency).

3) Reality check: “Realistic? Yes/No.” If No: “Cut/push” recommendations with rationale (which items, how many hours freed).

4) A proposed day-by-day plan (Mon–Fri) with time blocks.

Verification:

- Flag any assumptions you made (e.g., meeting load, focus time).

- Ask up to 3 clarifying questions if deadlines/dependencies are unclear.

AI isn’t a replacement for human judgment, but it’s good at assembling, sorting, and comparing complex information to make the decision process easier.

AI tools can help you:

Bring together the relevant information to make a decision
Surface where Reasoning agrees or disagrees
Stress-test an assumption before you build on it
Organise the factors at play

Try this query

Decision analysis query

Use when:Making a complex choice between two possible strategic options.

Role: Strategy analyst.

Task: I’m deciding between {Option A} and {Option B} for {decision/context}. Give the strongest case for each on its own terms, then tell me where those cases actually disagree: a difference in facts, assumptions, or priorities. Do not average them into a compromise.

Inputs:

- Decision context: {goal, constraints, stakeholders, timeline}

- Option A description: {key features, costs, risks, evidence}

- Option B description: {key features, costs, risks, evidence}

Constraints:

- Treat each option sympathetically; no straw-manning.

- Separate facts (verifiable), assumptions (beliefs), and priorities (values/trade-offs).

Output:

1) “Case for A” (5–7 bullets): benefits, risks, key assumptions, evidence cited.

2) “Case for B” (5–7 bullets): benefits, risks, key assumptions, evidence cited.

3) “True disagreements” table: Disagreement | Type (Fact/Assumption/Priority) | Why it matters for the decision | How to resolve/test.

4) “Decision criteria” checklist: the 3–5 criteria that should drive the choice given my context.

Verification:

- Flag any claim that lacks evidence or depends on a critical assumption.

- Ask up to 3 clarifying questions if the decision context is underspecified.

Today’s more sophisticated AI tools can go beyond simply answering questions to create slides, reports, documents or spreadsheets.

AI tools can help you:

Turn research or notes into a first-draft deck, spreadsheet, or report
Go straight from raw material to a formatted document
Structure a deliverable before you spend time on polish
Pull sources together as you build it

Try this query

Presentation outline query

Use when:You need a presentation outline with source-backed claims.

Role: Presentation strategist.

Task: Using the research below, build a five-slide outline for a stakeholder presentation: title slide, three content slides (clear headline + three supporting points each), and a recommendations slide. Note which source backs each claim.

Inputs:

- Research: {paste excerpts/links or attach files}

- Audience: {executives / technical / mixed}

- Goal: {what decision or action you want from stakeholders}

Constraints:

- Each supporting point must cite a source (title + date or URL).

- Keep headlines action-oriented (verb-first).

Output:

- Slide 1 (Title): Title | Subtitle | Presenter | Date.

- Slide 2 (Content): Headline | Point 1 (source) | Point 2 (source) | Point 3 (source).

- Slide 3 (Content): Headline | Point 1 (source) | Point 2 (source) | Point 3 (source).

- Slide 4 (Content): Headline | Point 1 (source) | Point 2 (source) | Point 3 (source).

- Slide 5 (Recommendations): Recommendation 1 (source) | Recommendation 2 (source) | Recommendation 3 (source) | Next step/ask.

Verification:

- Flag any claim that lacks a source or relies on outdated evidence (> {max_age}).

- Ask up to 3 clarifying questions if the goal/audience is unclear.

AI can reduce tedious manual tasks, such as entering or categorising data, labelling files, or copy-pasting information from one tool to another automatically.

AI tools can help you:

Categorise expenses and flag duplicates or missing entries
Draft routine emails
Turn a rough note into something submittable
Fill out repetitive forms from information you already have

Try this query

Categorisation query

Use when:Cleaning and summarising expenses for reporting.

Role: Finance operations assistant.

Task: Categorize each line of my expense export by {your categories}, flag duplicates or missing categories, and give me a summary total by category I can paste into my report.

Inputs:

- Expense export: {paste CSV-like rows or attach file}

- Categories: {e.g., Travel, Software, Meals, Office, Other}

- Reporting period: {month/year}

Constraints:

- If a line is ambiguous, assign the most likely category and mark it [Assumed: {reason}].

- Flag potential duplicates (same vendor/amount/date within ±1 day).

Output:

1) Categorized list: Date | Vendor | Amount | Category | Notes (e.g., [Assumed], [Duplicate?]).

2) Summary table: Category | Total amount | Count of lines.

3) “Issues to fix” list: missing categories, duplicates, outliers.

Verification:

- Ask up to 3 clarifying questions if the category list is incomplete or if currency/tax treatment is unclear.

AI is useful for catching inconsistencies within a single document, and reconciling claims across multiple sources, like flagging where a draft drifts from the style guide.

AI tools can help you:

Catch inconsistent numbers or claims across a document
Flag unclear phrasing before something ships
Check a draft against a checklist or style guide
Spot gaps between an early section and a later one

Try this query

Draft QC query

Use when:You need a targeted QC pass without a full rewrite.

Role: Senior editor.

Task: Check my draft against the style guide/checklist below. Flag numbers or claims that contradict each other elsewhere in the document, any sentence you had to re-read to understand, and anything the checklist requires that’s missing. List issues by location. Do not rewrite the whole thing.

Inputs:

- Draft: {paste text or attach file}

- Style guide/checklist: {paste bullets or attach doc}

Constraints:

- Reference locations by section/paragraph/line where possible.

- For contradictions, quote both conflicting passages.

Output:

1) “Contradictions” table: Location A | Quote A | Location B | Quote B | Why it conflicts | Suggested fix.

2) “Hard-to-read sentences” list: Location | Quote | Why it’s unclear | Suggested rewrite.

3) “Missing checklist items” list: Checklist item | Where it’s missing | Suggested addition.

Verification:

- Ask up to 3 clarifying questions if the checklist is ambiguous or the document structure is unclear.

Getting up to speed on an unfamiliar topic used to mean a slow read or a favour from a busy colleague.

AI tools can help you:

Get a working understanding of an unfamiliar tool, process, or topic fast
Prep questions you should be ready to answer
Explain a concept at the level you need
Fill in background before talking to someone more expert than you

Try this query

Information briefing query

Use when:You need a fast, plain-language briefing before a meeting.

Role: Subject-matter explainer.

Task: I have a meeting tomorrow about {topic} and I’m not deeply familiar with it. Explain the essentials in plain language: what it is, why it matters now, and the terms people will assume I know. Give me three questions I should be ready to answer or ask.

Inputs:

- Topic: {specific topic}

- Audience/industry: {e.g., fintech, healthcare, SaaS}

- Time sensitivity: {what’s new in the last 6–12 months}

Constraints:

- Plain language, no jargon without definitions.

- Time-bound: prefer developments from {start_date} to {end_date}; flag anything older.

Output:

1) “What it is” (3–5 bullets).

2) “Why it matters now” (3–5 bullets with recent developments).

3) “Terms you’ll be assumed to know” (term | one-line definition).

4) “Three questions to be ready for” (question | why it matters | talking points).

Verification:

- Flag any claim that lacks evidence or depends on a contested assumption.

- Ask up to 3 clarifying questions if the topic is too broad or the industry context is missing.

Context scattered across old emails and notes can be hard to reconstruct on short notice, right before a call.

AI tools can help you:

Pull together context before a conversation: what was discussed, promised, or left open
Prep a briefing before a client or stakeholder call
Track open items across an ongoing relationship
Draft a check-in that references real prior context

Try this query

Pre-call relationship briefing query

Use when:Preparing for a call with a person or account using past interactions.

Role: Account strategist.

Task: Before my call with {person_or_account} today, pull together what we discussed last time, anything promised on either side, and anything unresolved from our past emails/notes. Flag anything you’re not fully confident about so I can double-check.

Inputs:

- Person/account: {name, company, role}

- Past interactions: {paste last meeting notes + key emails or attach files}

- Today’s goal: {what you want from the call}

Constraints:

- Be literal: only record promises/unresolved items that are explicit or strongly implied; mark implied items as [Implied].

- If confidence is low, mark [Low confidence: {reason}].

Output:

1) “Last time” summary (3–5 bullets).

2) “Promises” table: Who | Promise | Date | Status (Done/Open) | Source snippet.

3) “Unresolved items” list: Item | Why it’s unresolved | Suggested next step.

4) “Call plan” (3 bullets): opening, key ask, fallback.

5) “Confidence flags” list: item | why confidence is low | how to verify.

Verification:

- Ask up to 3 clarifying questions if the history is incomplete or ambiguous.

Every prompt above assumes the tool has access to real, sometimes sensitive, company information, so a few habits up front keep that access from becoming a problem later.

Check for an approved tools list before you start. Most companies post one somewhere: an IT wiki, a security portal, an onboarding doc. If you can't find one, ask directly rather than assume there isn't one.
Log in through your company's SSO, not a personal account. A personal login doesn't inherit your company's data controls, audit trail, or offboarding process, so the work stays outside IT's visibility even when the work itself is company business.
Know your company's data tiers, not just "sensitive versus not." Most companies split data into more than two categories, for example public, internal, confidential, and restricted, and what's fine to paste into one tool may not be fine for another. Check with IT or security if the tiers aren't documented somewhere obvious.
Strip identifiers before you paste, when you can. Swap real names, account numbers, and client details for placeholders like {client_name} before pasting into a tool that hasn't been formally approved.
Turn off model-training toggles. Most AI tools have a setting, often labeled something like "improve the model" or "use my data for training," that opts your inputs into training data by default. Check the privacy or settings panel and turn it off unless the tool has been specifically approved without that concern.
Review what a tool can access before connecting it. If it books things, sends things, or touches your inbox, check what permissions it's requesting, and revoke access for anything you're not actively using.
Ask if you don't know whether something's approved. A quick message to IT or security is usually preferable to finding out after the fact that a tool wasn't cleared.

Do-and-don’t checklist for secure workplace AI use, emphasising company SSO, approved tools, data redaction, disabled training, and limited permissions.

Most knowledge workers already use AI casually: a one-off question, a quick email draft, copy the result, done. Treating a task as a two-week pilot, not a one-time question, is what turns that casual use into meaningful time savings.

Here’s how to start:

For five working days, log every repetitive task, how long it takes, and how tedious it feels. There are multiple time-tracking tools available, but a running note works fine: "Monday, 20 min, sorted receipts, mildly annoying." By Friday, a pattern usually stands out on its own.

Review the notes and choose which time-consuming, tedious task to automate first.

There are a few factors to consider:

Does the task happen multiple times a week?
Is the process relatively consistent every time?
Is it quick to check the output?
Does it depend on judgment calls that vary by situation?
Would an error be easy to catch and fix?
Is the source data complete and reliable?

A weekly status update or a recurring expense report fits. A once-a-quarter board deck doesn't. The stakes and the review time are both too high for a first pilot.

Write a sentence describing a correct or usable output, decided before seeing what the AI produces.

Pull three real examples of the output done manually (three status updates, three expense reports), and write down what they have in common: length, structure, level of detail.

Then ask what would make it unusable. Sometimes it's easier to define good by naming the failure. Include the opposite of each failure in the definition of a good output.

For example, for a status update, that might be: "Covers what shipped and what's blocked, in two paragraphs, with minimal rewriting needed."

For an expense report, it might be: "Every line has a category, and duplicates are flagged."

Start with a plain description of the task and the "good" definition from step 3, then run it against two or three real instances.

After each run, compare the output against that definition and note what's off: missing a section, too long, wrong tone. Add a line to the prompt addressing that specific gap, then run it again.

Stop once the output needs only minor editing, not rewriting. Save the final version somewhere easy to find again: a pinned doc, a notes app, or a shared prompt library, for instance.

Pick one repeatable thing to verify manually. For an expense report, that might mean comparing the category totals against the raw export before submitting it. Do that same check on every output for two weeks, even once it starts to feel reliable.

For each instance, track how long the task took start to finish, including the check step, and how it compares to the time logged in step 1. At the end of two weeks, add it up to decide whether it’s worth keeping the Automation as is. If not, adjust and retry, or drop it and pick a different task.

Pick the task that eats the most time and shows up most consistently, run it through the pilot steps, and note how much time or effort AI saves.

Perplexity Computer gives a knowledge worker more capacity without adding hours to the day. It connects to the tools already in daily use, including Gmail, Outlook, Google Drive, Notion, and Slack, and can pull the relevant material, draft the output, and flag anything that needs a second look before it goes out.

Try Perplexity Computer for your daily work.

What is AI in the workplace?

It's the use of AI tools, like connected assistants and chatbots, to help with the recurring parts of a job: staying informed, communicating, planning, reviewing, and administrative work, in addition to the one-off questions most people already ask.

What are examples of AI in the workplace?

Pulling action items out of a meeting transcript, turning raw notes into a status update, categorising an expense report, building a source-backed presentation outline, and prepping for a call using past context are five common ones.

What are the benefits of AI in the workplace?

The clearest benefit is time back on tasks that follow a predictable pattern: summarising, drafting, categorising, and reviewing work that takes real time to do by hand but doesn't require deep judgment.

What are the pros and cons of AI in the workplace?

The upside is faster first drafts and less time on repetitive tasks. The downside shows up when the check step gets skipped: an unverified summary treated as fact, a draft sent without review, or sensitive data pasted into a tool that was never approved for it.

How do you use AI in the workplace?

Start with one recurring task, track how it's currently done for a week, define what a good output looks like, test a prompt against a few real instances, and set one fixed check to run on every output during a two-week pilot.

What are the risks of AI in the workplace?

The main risks are treating AI output as verified when it isn't, sending sensitive data through an unapproved tool, and letting an AI summary substitute for a decision, an approval, or a relationship judgment that still needs a person.

Is AI going to replace jobs?

For most knowledge work, AI is replacing specific repetitive tasks inside a role rather than the role itself. The tasks in this guide, communication, admin, first drafts, are the ones most likely to shift, while judgment, accountability, and relationship-building stay with the person doing the job.

How to use AI in the workplace

1) Staying informed

2) Communication

3) Meetings

4) Planning and prioritisation

5) Analysis and decision-making

6) Producing deliverables

7) Administrative overhead

8) Reviewing and quality control

9) Learning

10) Managing relationships

How to get started with AI without irritating your IT teams

How to use AI in the workplace more effectively

1) Track a full week before making any changes

2) Pick one task

3) Define what "good" looks like before testing it

4) Build and test the prompt

5) Decide on one quality control check

6) After two weeks, compare against the baseline.

Putting this into practice

FAQs