Generative AI for Business Tasks: Practical Uses in Drafting, Summarisation, Translation and Research

Drafting emails and request notes

The easiest place to begin is drafting emails and request notes.

It comes in handy in situations such as these:

• An email to a client to arrange a meeting
• A request note to colleagues
• A report to your manager
• A first reply to a customer
• Wording for an apology or a request to confirm something

When you have generative AI draft these, specify your relationship with the recipient, the points you want to get across, the tone, and the length.

You are drafting an external-facing email.

# Recipient
A contact at an existing client. We have corresponded a few times before.

# Purpose
To arrange a meeting time next week.

# Points to convey
- Three candidate dates
- Around 30 minutes is expected
- We would like to hold it online

# Tone
Polite but not overly stiff.

# Constraints
- Within 300 characters
- Suggest a subject line too
- Give three options

Asking for three options rather than one lets you compare the directions the wording might take. The person then picks from them and, where needed, recasts it in their own words.

Be careful with emails involving apologies, contract terms, sums of money or deadlines. Rather than sending the AI’s output as it stands, a person must always give such wording a final check.

To measure the effect, it is worth looking at the time taken per email, the number of edits from managers or colleagues, and how many clarifying questions come back after sending.

Building outlines for proposals and reports

For proposals and reports, a useful approach is to shorten the time spent staring at a blank page.

If you are writing an internal improvement proposal, say, do not have it write the body straight off—ask for an outline first.

Create an outline for an internal proposal on the following theme.

# Theme
Building a mechanism to make first-line responses to internal enquiries more efficient

# Audience
Department heads, HR and General Affairs, IT department

# Output format
Five H2 headings, each with two or three lines of explanation

# Constraints
- Avoid being too assertive
- Mark cost-effectiveness as an “illustrative example”
- Include the risks of adoption

Asked this way, it produces the skeleton of a proposal in short order. The person then adds the actual internal circumstances, the figures and the views of those involved.

The thing to watch is that generative AI will happily invent plausible-looking figures and examples. Anything that is not an actual result should be marked as an “illustrative example”, and where you do treat something as a result, confirm the period, the number of cases and the conditions before using it.

For measuring the effect, look at the time to a first draft, how complete it is before review, the number of times it is sent back, and how many people are involved in producing the document.

Using it for rewriting and proofreading

In document work, generative AI is useful not only when writing from scratch but also when tidying up text you already have.

You are an editor of BtoB business documents.
Tidy up the following text to read more clearly, without changing the meaning.

# Points to revise
1. Make the subject of each sentence clear
2. Keep sentences short
3. Add a brief gloss to technical terms
4. Make emotive phrasing neutral
5. After revising, explain the changes in at most five points

# Original text
{paste the text here}

The crucial thing in rewriting is to state explicitly that the meaning must not change. In the course of smoothing the prose, generative AI can quietly alter facts or nuance.

Dates, sums, customer names, product names, contract terms and the names of schemes in particular must always be checked against the original.

Summarising meeting notes and minutes

Writing up minutes after a meeting is a task that suits generative AI well.

From meeting notes or a transcript, for instance, you can organise things into a format like this:

Turn the following meeting notes into minutes for internal sharing.


# Output format

## Meeting overview
- Date and time
- Attendees
- Purpose


## Decisions
Set out as bullet points


## To-dos
| Owner | Item | Deadline |


## Key points of discussion
Three to five lines per theme


# Rules
- Leave out small talk
- Where a deadline is unclear, write “to be confirmed”
- Separate decisions from matters still under discussion

This way, whoever writes the minutes works not from a blank page but by checking and correcting what the AI has organised.

Bear in mind, though, that the context of a meeting and the intent behind a remark are not always things the AI reads accurately. Decisions, owners, deadlines and open items in particular must always be checked by a person.

To measure the effect, it is worth looking at the time taken to produce minutes, the time from the meeting ending to sharing them, missed to-dos, and the number of correction requests from attendees.

Reading long documents and reports quickly

The time spent reading internal documents, research reports, regulations and manuals can also be shortened through summarisation.

Before sharing with your manager, for example, you might “sum up the key points of this document in three”, “pull out only the risks bearing on the decision”, or “organise only the parts relevant to the finance team”.

Summarise the following document on the basis that a member of the planning team will read it.

# Output format
1. Overall key points within 300 characters
2. Five important issues
3. Risks affecting the decision
4. Points needing further checking

# Constraints
- Quote figures exactly as in the original
- Add no interpretation not in the original
- Mark anything unclear as “to be confirmed”

Summarising is convenient, but the shorter you make something, the more falls away. For material used in important decisions, do not judge from the summary alone—check the relevant passage in the original.

Organising enquiry histories and chat logs

In HR, finance and support, similar enquiries tend to come round again and again.

Expense claims, attendance, leave, application procedures, how to use internal tools, and so on. Summarising and categorising the enquiry history makes the common questions and the procedures worth improving easier to see.

Categorise the following enquiry history.

# Output format
| Category | Count | Representative question | Action needed |

# Categories
- Expenses
- Attendance
- Application procedures
- System use
- Other

# Constraints
- Do not include individuals' names.
- Put anything you cannot judge into “Other”
- At the end, suggest questions worth adding to the FAQ

With this use, take care over how personal and customer information is handled. Names, email addresses, employee numbers, addresses, account details and the like should be masked before input.

Getting the gist of overseas material and English emails

When reading English material, trying to translate the whole thing accurately from the start takes time. It is more efficient first to pull out the gist, the key figures and the points worth checking.

Summarise the following English text in Japanese.

# Purpose
For a planning team member to grasp the gist

# Output format
1. A summary within 200 characters
2. A list of important figures, dates and proper nouns
3. Issues likely to bear on the work
4. Points where the original needs checking

# Constraints
- Give technical terms in English as well
- Mark anything whose meaning is uncertain as “to be confirmed”

Used this way, you produce not a full translation but a “map to read by” beforehand. It is handy for going through overseas news, industry reports, competitor information and outside vendors’ material.

The thing to watch is figures, dates, proper nouns and contract terms. However natural the AI’s translation reads, it does not necessarily reflect the original accurately. For passages used in important decisions, go back to the original and check.

Turning Japanese documents into English or plain Japanese

Translation is not only English into Japanese. There is also rendering an internal Japanese document into English, turning it into plain Japanese for staff who are not native speakers, or adjusting the tone for an overseas office.

Rewrite the following internal notice into Japanese that staff who are not native speakers can follow easily.

# Audience
Staff who use Japanese at work but are not used to formal honorific language

# Output conditions
- Keep sentences short
- Add a brief gloss to difficult words
- Set out important procedures as bullet points
- Do not change the original meaning

# Original text
{paste the text here}

This kind of use suits HR, General Affairs, IT and training staff.

For text touching on work rules, contracts, personal data or the law, however, checking the nuance after translation is indispensable. Where it is to serve as a formal internal document, treat confirmation by the relevant department or a specialist as a given.

How to measure the effect in translation

The effect of translation work is hard to gauge by character count alone.

Indicators worth watching include:

• Time taken to a first understanding of overseas material
• Time to produce a first-draft translation
• Number of terms corrected in review
• Number of clarifying questions from readers
• Time spent organising before sending out for external translation

As an illustrative example: where grasping the gist of a ten-page overseas report had taken sixty minutes, having the AI pull out the key points, figures and items to check first might bring the first read down to roughly twenty to thirty minutes. That said, it depends on how difficult the material is, how specialised the English, and how wide the final check.

This figure is not a general benchmark but a placeholder to start from when you begin measuring the effect internally. If you mean to use it in an article or an internal report, you need to record the page count, the number of people involved, the comparison period and the scope of review before doing so.

Drawing out the issues for a research topic

When looking into a new topic, the first difficulty is knowing where to begin.

It is worth having generative AI lay out the issues for the research.

The theme is “making internal enquiry handling more efficient”.
Before writing an internal report, organise the issues I should investigate.

# Output format
1. The seven issues to nail down, in order of importance
2. The primary sources to check for each
3. Items needing numerical confirmation
4. Counter-arguments easy to overlook
5. A proposed H2 structure for the final report

# Constraints
- Do not assert specific figures
- Mark anything needing a source as “source required”

Asked this way, you build the starting point of the research in short order. The person then goes on to check official material, internal data and primary sources.

Building comparison tables and evaluation criteria

In choosing tools, comparing outsourcers or weighing up schemes, building the comparison criteria takes time.

You can ask generative AI as follows:

Build evaluation criteria for comparing the following three options.

# Options
Option A: carry on with the existing tool
Option B: bring in a new SaaS
Option C: build part of it in-house

# Output format
| Evaluation criterion | Why it matters | Information to check |

# Constraints
- Do not judge on cost alone
- Include operational burden, security and training cost
- Write on the basis that the final call rests with a person

This gives you a footing that those involved can debate from more easily.

Prices, features, contract terms and legal compliance, though, must always be confirmed against the latest official information.

How to measure the effect in research

For research work, indicators such as these are useful:

• Time spent designing the research
• Gaps and omissions in the items investigated
• Time to a first draft of the report
• Number of further checks from managers and those involved
• Time to a decision

The effect of generative AI is not that “research itself becomes unnecessary”. Rather, it lies in setting the entry point in order and getting you to the information worth checking more quickly.

Check figures, proper nouns, dates and quotations

Generative AI is good at producing natural prose. It can, on the other hand, get figures, proper nouns, dates and the source of a quotation wrong.

For documents that leave the building, material a manager uses to make decisions, or information sent to customers, always check against the original source.

Representative things to check are:

• Figures such as sales, costs, headcounts and ratios
• Customer names, company names, product names
• Contract dates, deadlines, application cut-offs
• Names of laws, schemes and regulations
• Quotations and their sources

It is safer to treat the AI’s output not as finished copy but as a draft to be checked.

Do not feed in personal or sensitive information

What calls for particular care in generative AI use is how information is handled.

Kanata’s everyday-work best-practice guide treats public and general internal information as something to handle within a dedicated internal space, sets personal data as off limits in principle, and forbids sensitive information and undisclosed financial data.

The following sorts of information, for instance, need handling with care:

• Names, addresses, phone numbers, email addresses
• Employee numbers, account details, national ID numbers
• Health information, beliefs, family information
• Undisclosed financial results
• M&A, personnel changes, contract terms
• Customers’ confidential information

Even where you do need to use such information, you should mask it before input, check the internal rules, and consult the relevant department.

Approaches to AI governance differ by country, and the rules and guidance continue to evolve, so it is worth deciding the scope of use with reference to your own information-management policy. For an international reference point on the responsible development and use of AI, see the OECD AI Principles.

Assume specialist checking for legal, contractual, labour and financial matters

Generative AI can also be used to explain contracts, regulations, laws and schemes. It cannot, however, be left with the final judgement.

It can, for instance, do a first-pass review of a contract and “pull out clauses that look risky”. But whether to accept a clause, whether it should be amended, and whether it poses a legal problem all need checking by legal staff or a specialist.

The same goes for matters touching HR schemes, labour, accounting, financial products, medicine and safety.

Use outputs only after a person has reviewed them

Generative AI speeds up writing. But responsibility for the output rests with the person using it.

Text that leaves the building, wording sent to customers, and material used in management decisions must always be reviewed by a person.

In review, check the following:

• Whether the facts are correct
• Whether it could mislead the reader
• Whether there is excessive assertion or exaggeration
• Whether it breaches internal rules
• Whether sources and grounds can be confirmed
• Whether it contains personal or confidential information

Kanata’s best practice likewise holds that, because AI can produce “plausible falsehoods”, documents that leave the building, figures and quotations must always be verified by a person.

Start with one department, one task, one month

Aim for a company-wide rollout from the start and the rule-making and training balloon out of all proportion.

It is more realistic to try it small first—one department, one task, about a month.

You might try it in ways like these:

• In HR, for building an FAQ on internal enquiries
• In finance, for tidying up expense-claim guidance notes
• In planning, for building outlines of research reports
• In support, for categorising enquiry histories

Narrowing the focus makes both the effect and the snags easier to see.

Share your prompts

What generative AI produces turns on how the instruction is framed.

A request that worked well is easier to reuse if, rather than keeping it to yourself, you share it with the team.

Kanata describes a design in which your frequently used instructions are saved to a prompt library and reused within a project. There is also a reference-data library to which you can register internal material you want the AI to draw on.

The following sorts of prompts, for instance, are worth sharing:

• A minutes template
• An email-draft template
• An FAQ-building template
• A research-issue-organising template
• A translation-checking template

There are several options when it comes to generative AI tools. One is to use the AI features built into your existing chat tools or office software. Where you want to organise and reuse prompts and reference data by department or project, on the other hand, a business-support platform such as Kanata—which lets you handle apps and libraries on a per-project basis—is also an option. Which tool you use is an important choice, to be made to suit your internal security requirements, your existing systems and the burden on whoever runs it.

Decide what to delegate to AI and what people check

In your operating rules, set out not only “the tasks AI may be used for” but also “the points people check”.

Divided up this way, people on the ground can use it with greater confidence.

Review the effect and the failures once a month

In generative AI use, the failures matter as much as the successes.

Looking back once a month from angles like these makes the operation easier to improve:

• Which task saw time saved
• Which prompt was easy to use
• Which output contained errors
• Where there was uncertainty over what information could be entered
• What task to try next

Generative AI is not a case of make the rules once and you are done. It needs revisiting to suit the work, the internal rules, the tools in use and how practised the members are.