Quick Navigation
- The right mental model for AI at work
- The five modes — how AI actually helps
- What tasks fit and what don't
- What this looks like for your role
- Who does what — the handoff
- The mistakes that cost people most
- Building a personal system, not just a habit
- The workplace AI workflow in one diagram
- Cheat sheet
The right mental model for AI at work
There's a particular kind of disappointment I see often in organisations that have rolled out AI tools without a framework.
Someone uses AI for the first time, gets a surprisingly good draft of something, and stops checking. Then a month later, a fabricated statistic ends up in a client presentation, or an AI-drafted email goes out with the wrong tone to the wrong person, and suddenly the entire team becomes sceptical.
The problem wasn't the tool. It was the mental model.
Most people approach AI at work in one of two ways that don't work. They either trust it completely — treating outputs as finished work — or they distrust it completely and never reach for it at all. Both positions cost them.
The mental model that actually works is simpler and more familiar than either extreme: think of AI as a new colleague.
Not a genius colleague who is always right. A fast, capable, well-read colleague who has some specific quirks you need to know about: they've read almost everything ever written and can draw on it fluently. They have no memory of what you talked about yesterday — every conversation starts fresh. They sound confident even when they're wrong, and they won't always flag when they're uncertain. They need a clear brief to do good work — vague requests produce vague results. And their work should never go out without your review.
flowchart TD
A["You give a clear brief"]
A --> B["AI produces a first draft"]
B --> C{"Review:\nAccurate? Complete? On-brand?"}
C -->|"Yes"| D["Use it"]
C -->|"Needs work"| E["Edit and refine"]
C -->|"Wrong"| F["Re-prompt or redo"]
The colleague model sets the right defaults. Use AI actively. Give it real tasks. But never abdicate judgment. Every output that leaves your desk has your name on it, not AI's.
That's the whole model. Everything else in this guide is an application of it.
Try it yourself — The colleague test
Think of the last AI output you used without reviewing it properly. What was it? What could have gone wrong if it had an error? Would you have accepted the same output from a new colleague without checking? If not, why did you accept it from AI?
| Output I used without review | What could have gone wrong | What I'd do differently |
|---|---|---|
| e.g. Sent an AI-drafted client email without checking the tone | Wrong tone damaged the relationship | Always read for tone before sending external comms |
If your table is empty, you're either very careful or very lucky. Think harder.
The five modes — how AI actually helps
"Use AI at work" is not actionable advice. You need to know what kind of work AI helps with, and how.
Almost every AI workplace use case falls into one of five modes. Knowing these changes how you prompt — because instead of "help me with this report," you know exactly what you're asking for.
flowchart LR
A["AI Work Modes"] --> B["Draft\nFirst version of anything written"]
A --> C["Summarise\nCompress long content to key points"]
A --> D["Research\nSurface options, ideas, background"]
A --> E["Explain\nTranslate complexity into clarity"]
A --> F["Transform\nConvert content from one form to another"]
| Mode | What AI does | What you do |
|---|---|---|
| Draft | Writes the first version | Edit, refine, verify, approve |
| Summarise | Compresses content to essentials | Check nothing important was lost |
| Research | Surfaces relevant ideas, options, example | Apply judgment, verify facts |
| Explain | Translates complex content into plain language | Validate accuracy, adjust for audience |
| Transform | Converts content from one format to another | Review the output format |
Draft is the most used and most misunderstood. AI does not write finished work — it writes first drafts. The value is that a first draft at AI speed removes the blank-page problem, which is often where the most time gets lost. You still do the real work: the thinking, the editing, the judgment.
Summarise is where AI saves the most time most consistently. A 40-page report summarised into key points, meeting notes converted to decisions and actions, a research thread compressed into a briefing — these are high-value, low-risk uses of AI, as long as you check that nothing important was dropped.
Research is where hallucination risk is highest. AI is genuinely useful for surfacing frameworks, options, and considerations — the shape of the landscape. It is less reliable for specific facts, recent events, or precise data. Use it to generate the list of things to look into, not as the source of record for what you found.
Explain is where AI often surprises people with its usefulness. Complex technical content converted into plain language for a non-technical audience, a legal clause put into plain English, a financial model's assumptions explained simply — these tasks can take a human expert an hour and AI three minutes. The output still needs expert review, but the time saving is real.
Transform is the quiet workhorse. Bullet points into prose. Prose into slides. A technical document into a FAQ. A long email into a brief summary. These are low-risk, high-frequency uses that most people underuse once they learn the mode.
Naming the mode before you prompt is a small habit with a noticeable effect on output quality. It focuses both your brief and the model's attention on exactly what you need.
Try it yourself — The mode map
List five tasks you did this week. For each one, which mode is it? Which mode appears most often? That's where to build your prompt library first.
| Task | Mode (Draft/Summarise/Research/Explain/Transform) | How well did AI help? |
|---|---|---|
| e.g. Drafting weekly status update | Draft — AI wrote 80% of it | Saved 15 minutes |
| e.g. Summarising board report | Summarise — caught all key points | Saved 30 minutes |
What tasks fit and what don't
Knowing the five modes tells you how AI helps. This section tells you where — which specific workplace tasks are good fits and which aren't.
flowchart TD
TASKS["Workplace Tasks"] --> HIGH["High Fit\nUse confidently"]
TASKS --> MED["Medium Fit\nUse carefully — verify thoroughly"]
TASKS --> LOW["Low Fit\nAI assists, you decide"]
HIGH --> H1["Drafting emails and messages"]
HIGH --> H2["Summarising documents and meetings"]
HIGH --> H3["Writing reports and proposals"]
HIGH --> H4["Brainstorming ideas and options"]
HIGH --> H5["Reformatting and restructuring content"]
HIGH --> H6["Explaining concepts in plain language"]
MED --> M1["Researching background context"]
MED --> M2["Analysing qualitative feedback"]
MED --> M3["Drafting contracts or policies"]
MED --> M4["Interpreting data or reports"]
LOW --> L1["Final legal or compliance decisions"]
LOW --> L2["Live data lookups or real-time facts"]
LOW --> L3["Complex numerical calculations"]
LOW --> L4["Judgment calls — hiring, performance, escalation"]
The pattern that explains these groupings: high-fit tasks are ones where fluent language generation is most of what's needed, the output is easy for you to verify, and the cost of an error is low. Low-fit tasks are ones where precise facts, real-time data, or moral accountability matter — and where a confident wrong answer could cause real harm.
What determines task fit, practically:
| Question | High fit | Low fit |
|---|---|---|
| Needs live or recent data? | No | Yes |
| Needs precise facts or figures? | No | Yes |
| High consequence if wrong? | No | Yes |
| Can you verify the output yourself? | Yes | No |
| Follows a recognisable pattern? | Yes | No |
One question that cuts through all of it: can I verify this output? If yes, proceed — carefully or confidently depending on the stakes. If no, treat whatever comes back as a rough hypothesis only, not something to act on or share.
Try it yourself — Your task fit audit
For your top five AI use cases, rate the fit. Any "Low" with "No" on verification is a task to stop using AI for — or to add human review.
| Task | High / Medium / Low | Can I verify the output? | Cost of error? |
|---|---|---|---|
| e.g. Drafting client emails | High | Yes — I know the client | Medium — wrong tone damages relationship |
What this looks like for your role
The five modes apply universally. What changes by role is the context, the audience, the format, and the specific tasks they map onto.
flowchart LR
AI["AI Capabilities"] --> MGR["Manager\nBriefs · Feedback · Decisions"]
AI --> ANA["Analyst\nSummarise · Structure · Explain data"]
AI --> MKT["Marketer\nCopy · Campaigns · Content"]
AI --> ENG["Engineer / Developer\nCode · Debug · Docs"]
AI --> OPS["Operations\nProcesses · SOPs · Templates"]
AI --> HR["HR and L&D\nJob descriptions · Training materials"]
Managers tend to find the most consistent value in three places: drafting performance feedback (which is time-consuming to get right), summarising team updates into concise briefings for leadership, and preparing meeting agendas with pre-structured questions. The caveat on feedback: AI can structure and phrase it, but the substance — the observation, the judgment, the nuance of the relationship — has to come from you.
Analysts benefit most from summarisation and explanation. Converting a dense 60-page research report into a structured briefing. Drafting an executive summary from a set of findings. Explaining technical analysis in plain language for a non-technical stakeholder. Where analysts get into trouble is using AI to interpret the data itself — AI can structure what the data shows, but judgment about what it means is yours.
Marketers have arguably the widest range of genuinely useful applications: drafting copy, brainstorming campaign angles, adapting a message for different audiences and channels, generating headline variations, repurposing long-form content into social posts. The creative judgment — what's good, what's on-brand, what will land — stays with you. AI gives you the volume of options to react to.
Engineers and developers have access to one of the most mature AI applications: code generation. Boilerplate, documentation, debugging with a clear error message, translating between languages — these work well. The essential discipline is that you read and understand every line before it ships. AI-generated code can be subtly wrong, and you're accountable for what you deploy.
Operations teams find the most value in document work: writing SOPs, creating templates, summarising policy documents, drafting process documentation. These tasks are high-volume, follow recognisable patterns, and are relatively easy to verify — a strong fit for AI assistance.
HR and L&D professionals use AI most effectively for drafting: job descriptions, training materials, interview question sets, onboarding documents. The legal and cultural nuance — what language is appropriate, what questions are compliant, what tone fits the organisation — requires your expertise overlaid on everything AI produces.
The prompt that underpins all of this is the same regardless of role: specify who you are, who you're writing for, what good looks like, what format you need, and what to avoid. The role provides the context. The five modes provide the structure. The prompt engineering from the previous module connects them.
Who does what — the handoff
One of the most useful ways to think about AI-assisted work is as a structured handoff. You own the beginning. AI handles the middle. You own the end.
flowchart LR
A["You\nOwn the brief"] --> B["Clear task\nContext · Goal · Constraints"]
B --> C["AI\nProduces the draft"]
C --> D["First output\nFast · Structured · Needs review"]
D --> E["You\nReview and refine"]
E --> F{"Good enough?"}
F -->|"Yes"| G["Final output\nYour name on it"]
F -->|"No"| H["Re-prompt or edit"]
H --> C
What this looks like at each stage:
| Stage | Owner | What it actually involves |
|---|---|---|
| Define the task | You | Write a clear, specific prompt with context — mode, audience, format, constraints |
| Produce the draft | AI | Fast first version — never the final version |
| Review the draft | You | Check accuracy, tone, completeness, anything that could cause problems |
| Refine or re-prompt | You + AI | Iterate until it's right — this is the loop, not the exception |
| Approve the output | You | Your name goes on it — your responsibility |
The most important stage is the one people most often skip: review.
I've watched people copy AI output directly into an email and send it, or paste an AI-generated summary into a board presentation without reading it first. Sometimes it works out. Sometimes a statistic that doesn't exist ends up in front of a room full of executives.
The review habit doesn't need to be elaborate. For a quick internal email — read it once, does it say what you intended? For a client document — read it for accuracy, for tone, for anything that could be misread. For anything consequential — treat it the way you'd treat a draft from a junior colleague you'd just hired. Capable starting point. Needs your eyes before it goes anywhere.
The mistakes that cost people most
Most workplace AI failures aren't random. They follow predictable patterns. Knowing them in advance is the fastest way to avoid them.
Submitting without reviewing. This is the most costly mistake and the most common. AI output that hasn't been reviewed is not finished work — it's a draft with someone's name on it. The review doesn't take long. Fixing the fallout takes much longer.
Vague prompts. "Help me with this report" tells the model almost nothing. What mode? What audience? What format? What are the constraints? Every gap gets filled with the model's default — which is rarely what you wanted. Spend thirty seconds making the prompt specific and you'll spend far less time editing the output.
Wrong task fit. Using AI for tasks in the low-fit zone and treating the output as reliable. Asking AI for live data, precise calculations, or current regulatory requirements and acting on what it gives you without independent verification. The fit map exists precisely to prevent this.
No context given. AI doesn't know your role, your audience, your organisation's tone, or the background to this specific situation — unless you tell it. The single most reliable improvement to any prompt is adding the context the model needs to stop guessing.
Pasting sensitive data into public tools. This one has legal and reputational consequences that dwarf any time saving. Free-tier consumer AI tools are not appropriate for confidential client data, personally identifiable information, or commercially sensitive content. Know your organisation's AI data policy before you paste anything.
Expecting the first draft to be the final draft. People try AI once, get output that needs editing, and conclude that AI "doesn't work for them." The editing is the process. A first draft that's 70% right and takes two minutes is worth more than a blank page. The remaining 30% is where your expertise earns its keep.
flowchart TD
M["Common AI Mistakes at Work"]
M --> M1["Submitting without reviewing"]
M --> M2["Vague prompts with no context"]
M --> M3["Wrong task fit — using AI where it doesn't belong"]
M --> M4["No role, audience, or goal given"]
M --> M5["Pasting sensitive data into public tools"]
M --> M6["Expecting perfection from a first draft"]
The fix for all six is the same underlying discipline: slow down by thirty seconds before you start, and slow down by thirty seconds before you submit.
Building a personal system, not just a habit
Individual AI use cases are useful. But the people who get the most from AI at work over time are the ones who build a system — not just the habit of occasionally asking AI for help.
The difference is consistency. A habit is something you sometimes remember to do. A system is something that runs automatically because the decisions have already been made.
A personal AI workflow has three layers:
flowchart TD
W["Your Personal AI Workflow"] --> L1["Layer 1: Trigger List\nWhich tasks will I always try AI for first?"]
W --> L2["Layer 2: Prompt Library\nWhich prompts work well for my common tasks?"]
W --> L3["Layer 3: Review Habit\nHow do I check AI output before using it?"]
L1 --> EX1["Example: Every email draft over three paragraphs\nEvery meeting summary\nEvery first report draft"]
L2 --> EX2["Example: Saved prompts for weekly reports,\nstakeholder updates, and project briefs"]
L3 --> EX3["Example: Read once for accuracy,\nonce for tone, once for anything that could cause problems"]
Layer 1 — Your Trigger List
Write down five tasks you do every week that are high-fit for AI. Not what you might use AI for one day. What you will try AI for first, every time, starting now.
Common ones: any email draft over three paragraphs, any meeting summary, any first draft of a report or proposal, any first attempt at a job description, any explanation of a technical concept for a non-technical audience.
The trigger list removes the decision. Instead of "should I use AI for this?" — you already know the answer.
Layer 2 — Your Prompt Library
For each trigger task, build a prompt that works well. Test it, refine it, save it. A library of ten good prompts — one for each common task, with your role, format, and constraints already baked in — will save more time than a hundred one-off experiments.
What a saved prompt looks like in practice:
You are a professional executive assistant.
Summarise the following meeting notes into:
1. A 2-sentence overview of what was discussed
2. Key decisions made (bullet list)
3. Action items with owner and deadline, if stated
Tone: professional and neutral.
Do not infer or speculate — only summarise what was explicitly said.
[INSERT MEETING NOTES HERE]
That prompt took me twenty minutes to get right the first time. Now it takes zero time. That's the compounding value of a prompt library.
Layer 3 — Your Review Habit
Decide, for each output type, what review looks like before you use it. The standard should match the stakes.
Internal working documents: a quick read for obvious errors. External documents: read for accuracy, tone, and anything that could be misread. Client-facing or executive-facing outputs: treat as you would a draft from someone you'd just hired — assume it needs work until you've confirmed it doesn't.
The review habit is not about distrust. It's about ownership. AI produces the draft. You produce the output.
The workplace AI workflow in one diagram
flowchart TD
TASK["Work Task Arrives"]
TASK --> FIT{"AI fit?"}
FIT -->|"High fit"| PROMPT["Write a clear prompt\nTask + Context + Role + Format"]
FIT -->|"Medium fit"| CAREFUL["Prompt carefully\nPlan to verify output"]
FIT -->|"Low fit"| SKIP["Do it yourself\nAI can support, not decide"]
PROMPT --> DRAFT["AI produces first draft"]
CAREFUL --> DRAFT
DRAFT --> REVIEW["You review:\nAccurate? Complete? On-brand?"]
REVIEW -->|"Good"| OUTPUT["Final output — your name on it"]
REVIEW -->|"Needs work"| ITERATE["Refine prompt or edit draft"]
ITERATE --> DRAFT
What to carry forward:
AI at work is a colleague, not a vending machine. Brief it well, review its work, take responsibility for what goes out. That relationship — clear brief, active review, your name on the finish — is the whole model. Five modes cover almost everything: draft, summarise, research, explain, transform. Name the mode before you prompt. It focuses your brief and improves your output with no additional effort. Build a system, not just a habit. A trigger list of tasks you'll always try AI for first, a prompt library of saved prompts that work, and a review habit calibrated to the stakes — that combination produces consistent results. Individual experiments produce inconsistent ones.
Cheat Sheet — all the key terms
| Term | Plain English | Where it applies |
|---|---|---|
| AI as colleague | Fast, capable, always needs review — not always right | Every AI-assisted task |
| Draft mode | Let AI write the first version, you refine it | Emails, reports, proposals |
| Summarise mode | Compress content to key points | Meetings, documents, research |
| Research mode | Surface options and background — you verify | Planning, strategy, analysis |
| Explain mode | Translate complexity to clarity | Stakeholder comms, training |
| Transform mode | Convert content from one form to another | Slides ↔ prose, bullets ↔ narrative |
| Task fit | High / Medium / Low — match before prompting | Every task |
| High fit | Fluency is enough, easy to verify, low cost of error | Drafting, summarising, reformatting |
| Medium fit | AI helps, but facts and judgment need checking | Research, analysis, contracts |
| Low fit | AI assists thinking — it doesn't make the decision | Legal calls, live data, hiring |
| Human-in-the-loop | You own the brief and the finish | Consequential outputs |
| Trigger list | The tasks you'll always try AI for first | Your personal workflow |
| Prompt library | Saved, refined prompts for recurring tasks | Workflow efficiency |
| Review habit | How you check output before it goes anywhere | All final outputs |
| Data sensitivity | Know your org's AI policy before pasting anything | Any confidential content |
How to know if this landed
You'll know this has landed when someone names the mode before they prompt — draft, summarise, research, explain, transform — without thinking about it. When they have a trigger list that exists somewhere and they consult it. When they've built at least three saved prompts for their most common tasks, and those prompts include role, format, and constraints. When they review AI output before it goes anywhere — not always in depth, but always once. When they know their organisation's AI data policy and have applied it in the last week. When a colleague shows them a bad AI output, they can identify what was missing from the prompt. And when they treat AI output as a draft until they've reviewed it — not as finished work.
What changes after the session
The trigger list exercise is the one that shifts behaviour most immediately, and the reason is slightly counterintuitive.
People know, in the abstract, which of their tasks are good fits for AI. But without the trigger list, they face a decision every time: should I use AI for this particular email? This particular summary? That decision costs cognitive energy, and under any time pressure at all, the answer defaults to "I'll just do it myself."
The trigger list eliminates the decision. The answer is already made. If it's on the list, AI goes first.
What shifts: the frequency of AI use goes up considerably in the week after the session, even among people who were already using AI occasionally. More regular use produces better prompts, which produces better outputs, which produces more confidence, which produces more regular use. The compounding is real.
The prompt library has a different effect. The first time someone builds a saved prompt for a task they do every week — really takes the time to get the role, format, and constraints right — they usually can't believe they didn't do it sooner. A prompt that took twenty minutes to develop saves two minutes every time it's used. For a task done three times a week, that's paid back in five weeks. After a year, it's a significant chunk of reclaimed time.
The review habit is the hardest behaviour to change, and the most important. It doesn't require much — most review is a two-minute read. What it requires is the mindset shift from "AI output" to "draft requiring my approval." That shift is simple to describe and genuinely difficult to embed without deliberate practice. The workshop helps, but the team norm — the expectation that review is non-negotiable — is what makes it stick.
Book a Workshop
Ready to help your team move from knowing about AI to using it every day — confidently and responsibly?
or
Half-day workshop includes task-fit mapping using your team's real work tasks, role-specific use case deep dives, a prompt library sprint where every participant leaves with five saved prompts, review habit design calibrated to your team's output types, and a personal AI workflow map as a takeaway.