In practice · From Chat to Delegation

For your work

Artificial intelligence in your work

The manual applied to your work: for each field, the most useful cases with the request to copy and what to check. Every case is tested in the tools before it is published.

Your field

Trade unions

For people who do union work in a trade union, a staff association or an advice service: answering members, reading what the employer sends, preparing meetings with the employer, members' meetings and agreements. Eleven cases, each with the request to copy and adapt and with what to check before the result leaves the room.

The examples use your own collective agreement and fictional documents from a delivery company: if you work in another sector, change the agreement and the rest still applies. Each case lists the stages of the manual that teach the move. The cases were tested in Claude in the Italian edition of this guide, and the cases that report what the model did in a test were tested again in English on these documents, in November 2026. Where a test showed a typical mistake, the case says so.

Where the test used a fictional document, the case offers it for download: you can redo the test and compare the answer with what the case describes.

Before you start: members' data

Union work handles sensitive data: trade union membership itself is one of the special categories of personal data in the UK GDPR (Article 9), like health, which comes into many questions (time off for appointments, sickness absence, fitness for work). Three rules apply to every case below.

  • No names, staff numbers, National Insurance numbers or addresses in the requests. "A delivery worker at the depot" is enough for the model to work with.
  • With a free or personal tool, only documents without personal data: the collective agreement, local agreements, the employer's general announcements. If your organisation gives you its own tool, ask what you are allowed to upload (fact sheet "If the tool comes from your office").
  • No answer leaves the room until someone has checked it against the text it comes from.

1. A member's question about the agreement

The situation. A member sends a question about the collective agreement on WhatsApp. Usually you search the PDF, ask a colleague and reply the next day.

Setting it up, once. Open a project (stage 12) called "Agreement" and upload your collective agreement as a PDF, with the local agreements you use. In the project instructions write: "Answer only on the basis of the project documents. Always cite the clause and paragraph. If the agreement does not cover the case explicitly, say so and do not infer."

The request. Paste the question as it arrived, then:

"Reply to the member in five lines, in plain language, citing the clause and paragraph of the agreement. If the agreement does not provide for it, say so."

An example. The question: "I have a hospital appointment on Tuesday morning, how many hours of paid time off do I get a year for appointments?". Before you try, find the answer in your agreement yourself: you need it to judge the model's. Many agreements cap this time off, or treat it as unpaid or to be made up later. The point to watch is whether the model corrects the member when the agreement says something different from what they assume.

What to check. Open the PDF at the clause cited and read it next to the answer: thirty seconds. Then try a question the agreement does not fully cover, for example training a part-time worker attends on a day they do not normally work. If your agreement only says that training takes place "normally" during working hours, the model must answer that the agreement does not say so explicitly; if it cites a clause that settles it, it has invented it.

One more check, before anything else. A collective agreement is long, often a few hundred pages. A long PDF uploaded to a project can be read only part of the way: it happened in the test of the Italian edition, on a national agreement, and the model said so itself. Before you rely on it, ask: "Up to which clause can you read the agreement?". If it does not reach the end, upload the clauses you use most often as a separate file, or split the PDF into several files.

Stages. 3 (a document of your own), 4 (the check), 12 (the first project).

2. The employer's announcement

The situation. A letter from management arrives about a reorganisation. It has to be understood quickly, discussed by the branch committee and taken to the meeting with the employer.

The request, in three steps in the same chat.

"Read the announcement. Tell me in a table what changes, for whom, from when. Then list what the letter does not say and what we should ask at the meeting."

"Prepare an agenda for the meeting with the employer in six points and a draft message to members of 150 words, firm in tone, without adjectives."

"The message is too soft on the point about the sites: rewrite it saying that we ask for a written commitment not to close them."

To redo the test. The fictional document used: case2_reorganisation_letter.txt.

What to check. The list of what is missing is the most useful part: an employer's letter is also written so as not to say certain things. In the test, the practice letter left out three things (the number of staff affected, what the "review" of three delivery offices means and the date of the meeting), and the model found all three: with your own letters, ask yourself what you expect to find in that list before you read it. In the message to members, every fact must be traceable to the letter; the line to take, what to ask for and how hard to push, is for the branch committee to decide. In the test, asked to be firmer on the sites, the model added on its own "Without it, there is no agreement": a position nobody had taken.

Stages. 2 (the same request, done well), 3, 5 (the finished work).

3. The staffing figures

The situation. The employer sends a table with staffing levels before and after a reorganisation.

The request. Upload the file (Excel or CSV), then:

"Analyse the table. Where are most staff lost, in absolute terms and as a percentage? Make a bar chart of the changes by site and tell me three things I would ask the employer, looking at these figures."

To redo the test. The fictional document used: case3_staffing_table.csv.

What to check. In Claude and ChatGPT the model usually writes and runs a small program to do the calculations. Not always: in the test the model drew the chart and gave the right figures, but had worked them out without a program and did not say so; asked, it admitted it and redid them with a script. Ask how the figures were calculated, and check at least one total anyway. Then ask: "How many workers will lose their jobs?". The table is about staffing levels, not redundancies: the right answer is that the data does not allow it to be worked out. If the model gives a number, it has derived it from an assumption it has not told you. In the test it answered that the table does not say, and gave only a range with its assumptions.

Stages. 7 (Excel: formulas, data and charts).

4. The written statement for a disciplinary meeting

The situation. A worker receives a letter inviting them to a disciplinary meeting and setting out the allegations, and has a few days to prepare. The Acas Code of Practice on disciplinary and grievance procedures says the letter should give enough information to prepare an answer. The rep has collected the worker's account in notes.

The request. Paste the letter, without name and staff number, and the notes, then:

"Read the letter and the worker's account below. Prepare a draft written statement for the disciplinary meeting, in formal terms, using only the facts the worker has given. Do not admit anything the worker has not said and do not add arguments of your own. At the end, list separately the arguments you think are missing, so that we can decide whether to use them."

What to check. The model tends to add arguments the worker has not given, because they make the defence more convincing: that is why the request asks for them in a separate list. Read the draft line by line against the notes. The draft is a starting point: the signature and the responsibility stay with the worker and whoever accompanies them.

Stages. 3, 4.

5. The note of the meeting with the employer

The situation. After a meeting, what was said has to be put in writing, and above all what the employer has committed to do.

The request. Paste the notes taken during the meeting, or dictated straight after, then:

"From these notes of the meeting of [date] with the employer, write the note of the meeting: those present by role, points discussed, the position of each side, commitments made, by whom and by when, points still open. Keep what the employer said out loud separate from what it put in writing. Do not add anything that is not in the notes."

What to check. The commitments and dates, one by one, against the notes: they are the part you will use at the next meeting. Recording a meeting and having it transcribed (stage 17) is only possible if everyone present agrees.

Stages. 2, 17 (meetings as a source).

6. The week's news review

The situation. Following the news about the employer, the sector and employment law takes time, and what you miss you hear about from members.

The request. First try it in a chat with web search turned on (stage 15); when the result suits you, the same request becomes a task that starts on its own every week (stage 20).

"Search for news from the last seven days on: [your employer] (business plan, deliveries, depots and counters, staff); agreements and disputes in the sector; rules on work and artificial intelligence. Use only national newspapers, legislation.gov.uk, GOV.UK and the websites of Acas, the TUC and the unions. For each item: headline, date, two lines, link. If you cannot find a link that leads to the item, leave it out."

What to check. Open the links, at least for the items you will use: a wrong date or an old story presented as new are the most common mistakes. If you allow only a few sources, the review comes out short, and the model may set aside items from news agencies and specialist sites. In the test, on a postal employer, it found two items in the week, both from the union's website, and listed what it had left out because it came from a regional paper or another site. Decide once which sources you accept, agencies included, and write them in the request. Scheduled tasks are not available in every tool and every plan: stage 20 says where.

Stages. 15 (when it searches instead of remembering), 20 (the task that comes back on its own).

7. Bringing a policy into line with the rules

The situation. A rule changes, or the employer introduces something new, and an internal text has to be revised: the branch's own rules, the policy on members' data, or a company policy on which you submit your proposed changes.

The request. Upload the policy in Word and the text of the rules (downloaded from legislation.gov.uk or the regulator's website, not quoted from memory), then:

"Here are the policy and the text of the rules. Point out every part of the policy that has to change to comply with the rules. Propose each change as a tracked change, and for each change write a comment with the article and paragraph of the rules that requires it. Do not touch the parts the rules do not concern. At the end, list the points where the rules leave room for interpretation."

How it comes out. Add to the request "Give me back the policy in Word with tracked changes and comments". In the test, Claude delivered a Word file from the chat with the changes as tracked changes to accept or reject one by one, and a comment for each change. With the add-in inside Word (stage 6) the same work is done on the open file. If the tool you use does not create files, ask for a table with three columns: current text, proposed text, article of the rules.

An example. The company policy on the handheld delivery device, compared with the UK GDPR principles (Article 5), the information workers must be given (Article 13), the safeguards for automated decisions (Articles 22A to 22D) and the ICO guidance on monitoring workers. The practice documents include the extracts with their sources.

To redo the test. The documents used: case7_handheld_device_policy.docx, case9_reference_uk_gdpr_ico.txt.

What to check. Every comment must cite an article that exists and says that: open it. Reject the changes the rules do not require, even if they look like improvements. The points of interpretation in the final list go to whoever handles legal matters, and are not settled with the model.

Stages. 3, 4, 6 (Word and Outlook).

8. Comparing two versions of an agreement

The situation. The agreement in force and the proposed renewal, or two drafts of the same agreement exchanged between the parties. Reading the new text is not enough: what is missing does not show.

The request. Upload the two versions, then:

"Compare the two versions clause by clause. In a table: clause, what changes, in whose favour, text removed. Put first the paragraphs removed and the words that change the meaning, such as 'normally', 'by agreement with', 'after consulting the trade unions'."

To redo the test. The fictional documents used: case8_hybrid_working_agreement_2025.docx, case8_hybrid_working_agreement_2027_draft.docx.

What to check. The "text removed" rows, one by one, against the old version: they are the ones that count at the table. If a document is a scan, the model can misread a number or skip a line; with scanned PDFs, compare one page at a time.

Stages. 3, 6.

9. The employer introduces an automated system

The situation. A notice arrives about a new system: automatic allocation of delivery zones, measurement of delivery times, performance assessment. The data protection rules say what workers must be told, and the ICO guidance on monitoring workers expects the employer to seek the views of workers or their representatives before introducing monitoring. Your collective agreement may give the union further rights to information or consultation.

The request. Upload the notice and the text of the rules (the UK GDPR articles and the ICO guidance), then:

"Compare the notice with what the rules I have uploaded say workers must be told: who is responsible for the data and how to contact them, the purposes and the lawful basis, the categories of data, how long it is kept, the workers' rights including complaints, and, for automated decisions, meaningful information about the logic involved, the significance and the envisaged consequences, and how to obtain human review and contest a decision. For each item tell me whether the notice covers it, covers it in part or does not cover it, quoting the sentence. Then draft a formal letter to the employer asking for the missing information."

To redo the test. The documents used: case9_reference_uk_gdpr_ico.txt, case9_zone_system_notice.docx.

What to check. Check the list of items against the text of the rules you uploaded, not against what the model remembers: the UK rules on automated decisions changed on 5 February 2026, when Articles 22A to 22D replaced Article 22, and the ICO guidance of 2023 still refers to the old Article 22. A model may quote the old article. Whether the system decides "solely" by automated means, when a manager "may confirm" its proposal, is a point for whoever handles legal matters: ask the model to put it among the open points. The sentences quoted from the notice must really be there.

Stages. 3, 4, 15.

10. Listening to members

The situation. You have the open answers to a survey, or the questions received in a month, and you need to understand what really matters before a members' meeting or a claim.

The request. Upload the file of answers, without names, then:

"Group the answers by theme. For each theme: how many answers, two or three typical sentences quoted word for word, and in which sites or roles it comes up most. Order the themes from the most frequent. Put separately the answers that do not fit any theme."

To redo the test. The fictional document used: case10_survey_answers.csv.

What to check. That the quoted sentences really are in the answers: search for one or two in the file. The counts: if the file is a spreadsheet, the model counts with a program and the numbers can be redone; if the answers are pasted as text, the counts are estimates. The "no theme" group often contains the most serious report.

Stages. 3, 7.

11. Slides for the members' meeting

The situation. The reorganisation in cases 2 and 3 has to be explained to the workers at a members' meeting, in a quarter of an hour.

The request. In the same chat as case 2, or uploading the message to members and the staffing table:

"Prepare eight slides for a meeting of workers: what changes, who is affected, the figures by site with a chart, what we have asked the employer, the next steps. One sentence per slide, no long lists. In the notes under each slide, write what to say out loud."

What to check. Every number on the slides against the table, and the meaning the sentence gives it: the people whose work changes are those at the sites that change, not the whole headcount of the table. In the test the slides got this right (three offices, 47 of 125 posts), but "lose 47 jobs" reads as redundancies when the hubs gain 40 posts: words count as much as figures. The slide "what we have asked": if you have not given the model the real requests, it invents them. Working in the same chat as case 2, the slide used the request already made, and the model added a new one in the speaker notes and said it was still to be confirmed. The chart: axes, units and title.

Stages. 8 (PowerPoint: presentations and diagrams).

Cases tested on 7 November 2026.

Back to the manual: From Chat to Delegation