Which task should we start with? Ideally one that comes up often, takes time and is easy to check. A new tool on its own changes little. The benefit only shows once it is applied to work that recurs every week.
In many businesses, there are five such tasks: quotes, minutes, reports, research and onboarding. In all five, AI produces drafts. Decisions and approval stay with people.
How to recognise a good first task
Choose the task first, then the tool. Check each candidate against five criteria:
- It comes up every week.
- It currently takes a noticeable amount of time.
- It follows a recognisable pattern.
- A result can be checked in a few minutes.
- It needs no personal or confidential data, or there is a suitable tool for it.
A simple exercise: ask three or four people in the team to note down, for one week, every task that keeps recurring, with a rough estimate of how long it takes. At the end of the week, put the lists side by side and cross out everything that meets fewer than three of the five criteria. What remains are your candidates.
On the last criterion:
- If the task works without names, customer data or internal figures, a general AI assistant with a business account is enough.
- If personal details can be replaced beforehand, replace names and anything that could identify the person with placeholders such as “Customer A”.
- If the task cannot be done without personal data, you need a tool covered by a data processing agreement (DPA) under Article 28 GDPR. More on this in the article AI and the GDPR: which data you may enter.
- If the task needs your own knowledge, such as a staff handbook, and comes up very often, consider an application of your own. See ChatGPT, Copilot or your own AI?
Preparing quotes
Many enquiries go through the same steps: read them, spot missing details, assemble suitable building blocks, insert prices, write the text. AI can read the enquiry, identify missing details and write a draft from your building blocks. Prices are calculated using your stored rates, and the quote only goes out once it has been approved.
A worked example from a trades business that receives enquiries by email:
- Before: The enquiry comes in. Someone reads it, searches the folder for a similar old quote, copies standard text and realises while writing that measurements are missing. The follow-up question goes out, and the quote is only written once the reply arrives, often in several attempts.
- After: The enquiry goes to the AI without contact details. It first lists which details are missing and drafts the follow-up question. Once everything is available, it writes a draft from the approved building blocks. The employee inserts the prices from the price list, checks the line items, adjusts the tone and approves the quote.
The work shifts from gathering material to checking it. How much difference that makes varies from business to business. Measure it yourself, as described in the last section.
Template to try: “Here is a customer enquiry and our list of service building blocks. First list which details are missing for a quote, and write a friendly follow-up question asking for them. Then write a quote using only the building blocks from the list. Do not insert any prices. Mark those places with [PRICE].”
Minutes and summaries
Notes become minutes with tasks and responsibilities. Long emails and documents are cut down to the essentials. This task is well suited to getting started, because the person who sat in the meeting can judge the result straight away.
Specify a fixed structure. Then all minutes look the same and missing points stand out. A template you can adopt:
Template: “Turn the following notes into minutes with these sections: purpose and participants (roles only, no names), results, decisions, tasks with owner and deadline, open questions. Do not add anything that is not in the notes. If an owner or a deadline is missing, write ‘open’.”
The last sentence of the template is the most important. AI writes convincingly, even when something is wrong or missing. The instruction to mark missing information as open keeps gaps visible. Whoever checks the minutes can then see at a glance where questions still need to be asked.
Reports and documentation
Short notes, photos and master data become a traceable report. Open points stay visible, and a person approves the result. This suits service call-outs, site inspections or checks, for example, where someone currently writes up the report at the end of the day.
Clarify three things in advance: which data is used, where the limits are and who approves. In the example of a service report, that means notes, photos and master data from the job, and no assessment of people. The draft does not smooth anything over: open points stay open. Approval lies with the person who was on site. The reports example shows what this can look like.
To get started, a structure you already use is enough: job, condition found, work carried out, open points, next step. Include this structure in every prompt, and all drafts will follow the same layout.
Research and groundwork
Sorting topics, collecting questions, getting an overview: AI is well suited to this kind of groundwork. It helps you structure a new topic before you go into depth, for example before a meeting with a new supplier or before an internal decision.
Always check facts and sources yourself. Use the answers as a starting point, not as evidence. A template that takes this into account:
Template: “I am preparing for a conversation about [topic]. Break the topic down into five to seven subtopics and give me two questions I should ask for each one. Flag any statements you are unsure about and tell me where I can check them.”
Onboarding new colleagues
An assistant answers questions from your staff handbook and shows where the answer can be found. It only answers from approved documents, such as the handbook, works agreements and forms. Anything not covered there stays open. Personnel files and salary data do not belong in it. Questions about contracts and pay still go to HR.
Unlike the other four tasks, this one needs an application of your own, because a general assistant does not know your handbook. Whether that pays off depends on how often you bring new people on board and how well your documents are maintained.
You can do some groundwork even without an application: collect the questions new colleagues ask in their first weeks and check whether the answers exist in writing. This list later forms the basis for any assistant.
Next steps
Do not start with decisions about people, or with tasks whose results are hard to check. AI that screens or assesses job applications counts as high-risk AI under the EU AI Act. More on this in the article High-risk AI: which deadlines now apply.
To start with one of the five tasks:
- Choose one task and one person who is responsible for it.
- Estimate how long the task takes today and write it down.
- Try the matching template from this article on five to ten real cases, without personal data.
- Adjust the template until the drafts are usable, and save it for the team.
- After a few weeks, compare time spent, quality and satisfaction in the team.
If you would first like to see which areas take up the most time in your business, the AI check can help. Which task takes up the most time in your team every week?
Frequently asked questions
What should we avoid starting with?
Decisions about people, such as job applications, and tasks whose results are hard to check. AI that screens or assesses job applications counts as high-risk AI under the EU AI Act.
How do we measure the benefit?
Estimate beforehand how long the task takes and compare after a few weeks. Also pay attention to quality and satisfaction in the team.
Which tool do we need to start with?
For quotes, minutes, reports and research, a general AI assistant with a business account and a suitable contract is often enough. Onboarding needs an application of your own that answers only from your approved documents.
Can customer data go into the templates?
Only into a tool covered by a data processing agreement (DPA) under Article 28 GDPR, and only with a legal basis for the processing. It is simpler to replace names and anything that could identify the person with placeholders beforehand. Many drafts work well that way too.
What if the results are not right at first?
Improve the prompt first: specify a structure, include a good example, have missing information marked as open. If the results are still unusable after several attempts, the task is probably not a good starting point.
Sources
Updated: . This article is not legal advice.
