AI at work · Useful habits, better processes
How to use AI effectively at work
Choose a job worth improving. Give AI the right information. Check what comes back. Then find out whether you have actually saved any time.
Buying an AI subscription takes a few minutes. Working out what to do with it takes a little more thought. This is inconvenient for anyone hoping that a new icon on the desktop would sort out the business.
My advice is to start with a job you already understand. The weekly update. The enquiry that always needs the same three questions answered. The notes that have to become a useful brief. Something ordinary enough that you can recognise a good result when you see one.
This guide shows you how to choose that job, build a small AI workflow and test whether it helps. There are six copyable prompts, a fictional customer-enquiry example and a simple time calculation. No grand transformation required before Thursday.
What is the best way to start using AI at work?
Look for a task that repeats, has reasonably clear inputs and produces an output you can judge. Drafting a routine update is a better first experiment than asking AI to decide which supplier should receive a substantial contract.
Pick something small enough to finish
Write down what you do, how often you do it and what makes it annoying. Then look for the part that involves organising, drafting, comparing or reformatting information. You do not need to hand over the whole job.
A useful first-task filter
| Question | A promising answer |
|---|---|
| Does it repeat? | Often enough to make learning the method worthwhile. |
| Are the inputs clear? | I can identify the facts and current source. |
| Can I check the output? | I know what correct and useful look like. |
| Can a mistake be caught? | A person reviews it before anyone relies on it. |
| Is the information suitable? | The account and data use are approved for this task. |
Try this prompt to choose a useful starting task
Prompt 1: find a sensible first AI task
Help me choose one small task to improve with AI. Here is a generic description of my recurring work, with no personal or confidential information: [tasks, frequency, approximate time and common problems]. Compare the tasks by clarity of inputs, ease of checking, likely consequences of an error and effort needed to try a first version. Do not invent time savings. Recommend one task, explain which part AI could help with and identify the human review needed. Ask about any missing detail that would change your recommendation.
Check it: Choose a task you understand well. An attractive suggestion is no use if you cannot tell whether the result is right.
Which AI tools should you use at work?
Start with what your organisation has approved and what the task needs. A writing assistant may suit a draft. An assistant connected to the right workplace files may suit a sourced update. A spreadsheet formula may solve the problem without a conversation at all.
Check the account and the information
Do not assume that a personal account and a work account have the same arrangements. Ask which tools you can use, what information you can provide and who is responsible for the result. Payment alone does not answer those questions.
The NCSC’s guidance explains risks around information submitted to language models and unreliable output. Those are useful reasons to set practical boundaries before uploading work, rather than discover them after the event. [1]
Write a short operating rule
For a first trial, record the approved tool, permitted inputs, required review and actions that need a person. Keep it specific enough to use. A policy saying “be responsible with AI” is an admirable sentiment with very little to say about Tuesday’s spreadsheet.
Try this prompt to prepare the trial rules
Prompt 2: draft a clear rule for this trial
Draft a short operating note for this AI trial using only these agreed details. Task: [task]. Approved tool and account: [details]. Permitted information: [details]. Information excluded: [details]. Reviewer: [role]. Actions requiring human approval: [actions]. Use headings for purpose, inputs, review, stop conditions and owner. Leave missing decisions marked “to confirm”. Do not claim that this note establishes legal compliance or approves a tool. Keep it under 200 words for the responsible person to review.
Check it: Check the note with the person responsible for the trial. AI can help write down a decision; it cannot supply approval you have not been given.
A practical AI workflow: turn an enquiry into a checked reply
Here is a fictional example you can try without using a real customer’s details. A training business receives an enquiry about a workshop. The aim is to draft a helpful reply and expose the gaps before anyone sends it.
Give it a small, complete source
- Customer request: an introductory AI workshop for eight people, ideally on 12 November, delivered online.
- Confirmed offer: a 90-minute introductory workshop can be delivered online.
- Still unknown: availability on that date, the price and the participants’ current experience.
- Required next step: ask about experience and say that availability and the quote will be checked.
These are practice facts, not a description of a current Oxford AI School offer or an available booking.
Separate drafting from committing
AI can arrange these facts into a reply. It cannot make the diary available, authorise a discount or decide what the customer has agreed. Keep those decisions outside the drafting step.
Try the complete enquiry prompt
Prompt 3: draft the reply and expose the gaps
Use this fictional training enquiry only. The customer wants an introductory AI workshop for eight people, ideally on 12 November, online. A 90-minute introductory workshop can be delivered online. Availability on that date, the price and the participants’ experience are unknown. Draft a warm, direct reply in British English, under 140 words. Acknowledge the request, ask about participants’ experience and say that date availability and the quote will be checked. Do not confirm a booking, price or discount. After the draft, list the facts used and the checks needed before sending. Treat the customer’s preferences as requests, not confirmed arrangements. Do not send anything.
Check it: The reply should leave the date and price unconfirmed. Check all three unknowns remain visible, and that “eight people” has not become eight booked places.
How do you check AI-generated work?
Check the parts that could change what someone believes or does: facts, figures, commitments, omissions and the requested next step. Then check the writing. A graceful sentence promising the wrong thing is still the wrong thing.
Use a known error to test the checking method
Take the practice enquiry above. Imagine the draft says:
“Your workshop for eight people is confirmed for 12 November at £450. We will send the joining link shortly.”
That sounds efficient. It has also confirmed an unconfirmed date, invented a price and implied that the booking is settled. It fails to ask about experience. This is why “it reads well” is not a quality check.
Give the review a fixed job
Compare claims with the source. Mark unsupported statements. Identify missing questions. Do not ask a second AI response to certify the first one as correct and assume the matter is closed. You still need to inspect the evidence yourself.
Try this AI quality-check prompt
Prompt 4: audit the draft against the source
Source facts: eight people want an introductory online AI workshop, ideally on 12 November. A 90-minute online workshop is possible. Date availability, price and participants’ experience are unknown. The reply should ask about experience and explain that availability and a quote will be checked. Draft: “Your workshop for eight people is confirmed for 12 November at £450. We will send the joining link shortly.” Compare every factual claim and commitment with the source. Return a table of claim, source support, problem and suggested correction. Identify required questions or next steps omitted from the draft. Do not introduce new facts. Then provide a corrected short reply for human review.
Check it: The review must catch the unsupported date confirmation, invented £450 price, implied settled booking and missing experience question. If it misses one, the review process needs work.
How can you measure AI productivity without fooling yourself?
Time the complete task. Include preparing the input, drafting, checking and correcting. If AI produces a draft in ten seconds and you spend twenty minutes repairing it, the ten seconds are not the interesting number.
A simple fictional time comparison
| Activity | Manual method | AI-assisted method |
|---|---|---|
| Prepare and draft | 9 minutes | 2 minutes |
| Check | 3 minutes | 5 minutes |
| Additional corrections | Included above | 1 minute |
| Total | 12 minutes | 8 minutes |
In this invented example, the saving is four minutes per task. At ten comparable tasks, that would be 40 minutes. It is an illustration of the calculation, not a result we are claiming for a customer or a promise for your team.
Count setup separately and judge quality too
Record the time spent learning the tool and creating the process. Then measure repeated use over several similar tasks. Note errors, rework and whether the result actually helped its reader. A quick answer nobody can use is not a productivity gain.
Try this prompt to design a small trial
Prompt 5: measure the whole task
Help me design a small comparison for this recurring task: [task]. Current method: [steps]. Proposed AI-assisted method: [steps]. Quality standard: [what a usable result must meet]. Create a simple recording table for preparation time, drafting time, checking time, corrections, total time, errors found and whether the result met the standard. Record initial setup and training time separately. Suggest how to compare several similar examples fairly. Do not predict savings, invent results or convert minutes into financial savings without an agreed basis. Explain what would count as a reason to stop or revise the trial.
Check it: Use comparable tasks and the same quality standard. Count the awkward examples as well as the one that looked brilliant in the demonstration.
Build an AI workflow you can repeat
Once a method works, write it down. Keep the prompt, required inputs, an approved example, known mistakes and the checks. Give it an owner and a review date. Otherwise the team will develop six versions of the same prompt and a vigorous debate about which one was the good one.
Save the method, not just the magic words
The prompt is only part of the process. Someone must select the right file, provide the current facts, inspect the output and decide what happens next. A reusable workflow makes those responsibilities visible.
Keep actions and approvals explicit
Drafting a reply is different from sending it. Suggesting a spreadsheet change is different from applying it. If you later connect tools that can take action, define the permissions and review steps for those actions before enabling them.
Try this prompt to create a workflow card
Prompt 6: document the tested workflow
Create a one-page workflow card from this tested method. Task and purpose: [details]. Owner: [role]. Approved tool: [tool/account]. Required inputs: [inputs and source/version]. Prompt that worked: [prompt]. Accepted example: [approved example]. Known failures: [failures]. Human checks and approval: [steps]. Review date: [date]. Organise it into inputs, steps, expected output, quality checks, stop conditions and owner. Clearly separate drafting from actions such as sending, publishing or changing records. Mark missing details for confirmation. Do not add integrations or automatic actions.
Check it: Give the card to a colleague with a fresh practice example. If they need to guess what you meant, improve the instructions before calling the process reusable.
How should a team introduce AI into everyday work?
Ask each person to bring one recurring task and one example of a good result. Compare the tasks, choose a small trial and agree who checks the output. Keep the discussion close to the work instead of holding a competition to name the largest number of AI tools.
A manageable first week
- Choose: select one task and record the current method.
- Prepare: confirm the tool, inputs and review responsibilities.
- Try: use a fictional or suitable approved example.
- Compare: check accuracy, usefulness and total time.
- Decide: keep, revise or stop the method based on the evidence.
Make it easy to report a failure
If people are expected to produce impressive AI success stories every Friday, they may become rather good at producing impressive AI success stories. Ask what failed, what took longer and what should remain manual. That is useful information.
Some tasks should stay with a person
Use particular care where a mistake could materially affect someone’s rights, money, health or employment. For a beginner trial, choose a lower-consequence drafting or organising task with clear review. Do not treat a fluent output as the authority to make a consequential decision.
Make one piece of work better before adding another tool
The useful question is not “how much AI are we using?”. It is “which job is now easier, clearer or more reliable, and how do we know?”. That keeps the conversation grounded when the next remarkable product announcement arrives.
Choose the guide that fits your next task
If you are developing copy or comparing documents, start with how to use Claude effectively. If your work sits in Outlook, Word, Excel, Teams or PowerPoint, read how to use Copilot effectively.
For practical support, explore our AI training programmes or basic AI training. Bring a task and an example of good work. We can use those to make the learning relevant.
Start small. Check properly. Keep what works. You do not need to automate the office before you can improve the weekly update.
Keep the exercises: Download all 6 prompts.
Keep learning: more ways to use your AI effectively
Sources and further reading
Sources checked on 16 October 2026. Product features depend on account, plan and settings. The prompts and practice examples are original editorial exercises; fictional figures are not reported customer results.
How to use AI effectively at work: frequently asked questions
How do I start using AI effectively at work?
Choose one familiar, repeated task with clear inputs and an output you can check. Use an approved tool, try suitable practice information and review the result. Measure the complete task before deciding whether to repeat it.
What are good examples of AI use at work?
Useful starting tasks include drafting a routine update, organising notes into a brief, comparing approved documents and preparing questions for a meeting. Keep the task narrow and retain human review before anyone relies on the output.
What are the best AI prompts for work?
Useful prompts describe the task, source, audience, output and boundaries. Include the facts the assistant may use, explain what must remain unconfirmed and state how the result will be checked. A specific brief usually helps more than a long list of impressive roles.
Which AI tool is best for workplace productivity?
Choose an approved tool that fits the task and can use the relevant information appropriately. Compare results on the same example, including checking and correction time. There is no single best tool for every job.
Can I put confidential work into an AI tool?
Only use information that your organisation has approved for the specific tool, account and purpose. Check the actual arrangements with the responsible person. A paid account alone does not establish permission to upload confidential material.
How do I measure whether AI saves time?
Record preparation, drafting, checking and correction time for comparable tasks. Track quality and errors as well as total minutes. Record setup and training separately, and avoid drawing conclusions from one unusually good example.
Should I let AI send emails or update records automatically?
Start with drafts that a person reviews. Any later ability to send, publish or change records needs clearly defined permissions, review steps and responsibility. A successful drafting trial does not by itself justify automatic actions.
Do I need coding skills to use AI at work?
You can practise drafting, organising and comparing information without coding. Clear instructions, appropriate source material and good checking habits are the main starting skills. More complex integrations may require additional expertise.
How can a manager help a team use AI well?
Agree suitable tasks, approved tools and clear quality standards. Give people time to practise, appoint a reviewer and discuss failures as openly as successes. Keep a shared record of tested methods rather than collecting unverified prompt tips.
What should I do if AI makes the task slower?
Check whether the inputs, task scope or instructions can be improved. If preparation and correction still outweigh the benefit at the required quality, stop that trial or choose another task. Using AI is optional; producing useful work is the objective.