What Generative AI Can and Cannot Do for Business

Producing a first draft

A classic use of generative AI is producing a first draft of a piece of writing.

It suits work such as the following, for instance:

• Drafts of external emails
• Internal announcements
• Training notices
• Meeting requests
• Thank-you emails
• First passes at letters of apology
• First drafts of reports
• Draft answers for an FAQ

Much of what makes writing slow comes down to “the first sentence won’t come”, “the structure won’t settle”, and “I keep second-guessing the phrasing”. Generative AI helps you get off the mark faster.

Please draft an internal email letting new joiners know about a generative-AI training course. The audience is all staff; keep the tone polite and positive. Please also suggest three subject lines.

Ask in this way and you get a working draft far sooner than starting from a blank page.

That said, you should avoid using AI’s wording as it stands. The relationship with the recipient, the company’s standing, the history between you, and the warmth or weight of particular words are things AI does not necessarily grasp with any precision.

I use AI to draft my emails, but I always review them myself before sending. I check how the message will land with the recipient and whether any wording might weaken their trust. That final human check is what protects the relationship and the tone we have built up so far.

Leave the first draft to generative AI; let a person shape the final wording. That is the basic principle.

Summarising long passages

Generative AI is also well suited to condensing long passages or meeting notes into something shorter and tidier.

Take the following situations, for example:

• Pulling decisions and to-dos out of meeting notes
• Boiling a long document down to something you can read in three minutes
• Tidying up the relevant passages of internal regulations
• Sorting points raised in a customer interview by theme
• Summarising the content of a training video

Writing up minutes after a meeting, sorting decisions, points of debate, owners, and deadlines by re-reading the recording or notes from the top, is laborious. With generative AI, you can at least get a rough structure in place first.

In Kanata, the AI summary feature lets you organise information from documents, images, audio, URLs, text, and more. It handles documents such as CSV, PDF, PowerPoint, Word, and Excel, so it lends itself to summarising meeting materials, internal documents, and training resources. Note that documents are subject to a 10MB limit, so for larger files it is best to split them or to pull out only the part you need.

An AI summary is an aid to the memory of a meeting. It does not take on the responsibility for the meeting.

Organising information into comparison tables and checklists

Generative AI is also good at tidying up scattered information.

You can ask it to do things like the following, for example:

• Put the pros and cons of Option A and Option B into a table
• Order the points at issue for an initiative by importance
• Draw up a checklist for before a project kicks off
• Split a customer’s issues into “apparent” and “latent”
• Tease out the points that need deciding in a meeting

This use is especially handy for front-line managers and planning staff. Rather than leaving the judgement itself to AI, handing it the “organising” that comes before the judgement frees people to concentrate on the decision.

We are weighing up whether to bring in a new internal training tool. Please give five axes of comparison, and set out in a table the pros and cons of carrying on with the existing arrangement versus adopting the new tool. Finally, please list the things we should check before deciding.

Ask in this way and you have a starting point for the discussion.

This is also the format I reach for when first sorting out a project. It is not yet the moment to reach a conclusion. But the points in everyone’s heads are scattered. Putting them into a table once with AI at such a moment makes it easier to see what needs discussing and what is already settled.

That said, what goes into a comparison table is heavily shaped by what you put in.If the premise conditions are thin, the table will tend towards generalities. When asking generative AI to organise information, it is important to be clear about the scope, the purpose, the conditions, and the criteria for judgement.

Broadening ideas

Generative AI can also be used to generate ideas.

Take situations such as these, for example:

• Suggesting themes for internal training
• Drawing up several article-title options
• Thinking through angles for a sales email
• Putting forward hypotheses for a new initiative
• Drafting an agenda for a meeting
• Drawing up a list of questions for a customer

Left to think alone, a person inevitably leans on their usual ideas. Having generative AI put forward several options lets you broaden the range of your thinking.

That said, AI’s ideas are not necessarily usable as they stand. Whether to adopt one is something a person must judge, weighing the realities on the ground, the customer’s mood, internal constraints, the budget, and the resources available.

Generative AI becomes easier to use once you think of it not as “a tool that gives the right answer” but as “a tool that adds to your options”.

I think of generative AI’s idea-generation as close to a sounding board. Put a question to the AI and, at the very least, a different angle comes back. Of course, you don’t adopt it as is. But as a way of multiplying the prompts to think, it is a thoroughly practical use.

Reusing prompts and knowledge

When generative-AI use is left to individuals, the gap widens between those who use it well and those who don’t. To close that gap, it matters to share good prompts and reference materials across the team.

Build up, by department or by project, things such as output formats for minutes, draft-prompt templates for sales emails, and FAQ data on internal regulations, and you spare yourself designing the same work from scratch every time.

In Kanata, you can register AI settings, prompts, and training data by project, and reuse them from the chat or summary apps. Keep the prompts and internal materials you use day to day in order, and it becomes easier to turn one person’s ingenuity into a team asset.

What matters here is not to think tool-first. Shared folders, knowledge bases, internal wikis, AI platforms; there are plenty of options. Among them, where you want AI chat, summarising, and training-data management on the same line of work, an integrated environment such as Kanata can be a candidate.

In organisation-wide reskilling, an individual merely putting one-off questions to AI does not amount to organisational learning. By building up the approaches that worked and putting them into a form the team can reuse, generative-AI use gradually becomes standardised.

Guaranteeing factual accuracy

Generative AI is good at producing plausible-sounding prose. But sounding plausible and being correct are not the same thing.

The following, in particular, always need checking against the source:

• Figures
• Dates
• Laws and regulations
• Internal regulations
• Contract terms
• Customer names
• Company names
• Quotations
• The latest information
• Prices and specifications
• Market data

Even if generative AI answers that “the market is worth around one trillion yen”, without a source you cannot use it in external materials. Even for an internal explanation, any figure bearing on a management or budget decision must be checked back against a reliable source.

When asking generative AI to research, it is safer to ask it to “set out the points I should look into”, “list the primary sources I should confirm”, or “raise the angles easily overlooked on this theme” than to say “tell me the facts”.

Have AI draw up the map for the research, and let people pin down the facts. That division of labour matters.

Standing in for the final judgement

Generative AI can organise the material for a judgement. But it is no substitute for the final judgement itself.

The following kinds of judgement, for instance, are ones people and organisations should own:

• Whether to hire someone
• How to handle a performance review
• Whether to halt an initiative
• Whether to submit materials outside the company
• Whether to accept the terms of a contract
• How to apologise to a customer
• Whether to make an investment

Asking generative AI “which is better, Option A or Option B” is in itself perfectly fine. If anything, it helps sort out the points at issue.

But AI’s answer rests on the input and on general inference. It cannot adequately reflect the organisation’s policy, the history behind things, the feelings of those involved, or the appetite for risk.

Generative AI should be used not as “the boss who decides” but as “the aide who organises the material for the decision”.

In my view, rather than simply going along with the conclusion AI reaches, it is worth far more in practice to put questions back to it: “why can you say that?”, “what is the counter-argument?”, “what happens if the premise changes?”.

Carrying through a conversation grounded in the relationship with the other person

Generative AI can produce letters of apology, explanations, negotiating emails, and draft questions for a one-to-one. But to leave the conversation itself, the one that rests on the relationship with the other person, entirely to it is risky.

In an apology email to a customer, for instance, more than the polish of the wording, points such as these matter:

• The history so far
• What the customer is unhappy about
• How far your company will admit responsibility
• What response you can promise going forward
• Whose name it should be sent under
• Whether a call or a visit is needed

These are not things AI can judge on its own.

Internal one-to-ones and performance reviews are the same. You can have AI draft questions or comments, but which words to use with a team member, in what order to speak, and how far to press are things a manager needs to own.

Where a relationship of trust between people is involved, AI should be kept to a tool that helps with the preparation.

My sense is that AI is at its most powerful “before the conversation”. Sorting out what needs saying. Anticipating the points the other person may find worrying. Putting forward several ways of phrasing it. AI helps as far as that. But actually watching the other person’s face,reading the other person’s feelings and emotions, and choosing your words is a human job, and choosing your words is a human job.

Handling confidential or personal information unconditionally

What calls for the most caution in using generative AI for work is the handling of input information.

Japan’s Ministry of Internal Affairs and Ministry of Economy guidelines on AI businesses (March 2026) sets out that those who deploy AI should recognise the risks and carry out the necessary measures across the whole lifecycle. For business use too, you need to think about classifying input information, access rights, log management, training, and review arrangements as a set.

Information calling for particular care includes the following:

• Names
• Addresses
• Phone numbers
• Email addresses
• Employee numbers
• Bank details
• Health information
• National ID numbers
• Unpublished financial results
• M&A information
• Personnel-move information
• Customers’ contract terms
• Customers’ employee information

Such information must be handled in line with company rules, not on the basis of “a little won’t hurt” or “it’s internal, so it’s fine”.

Before using generative AI, it is safer to hold to at least the following tests:

1. Is the information public?
2. Is it information anyone inside the company may see?
3. Does it contain anything that could identify an individual?
4. Does it touch on a customer contract or an NDA?
5. Is it information that would cause trouble if it got outside the company?

If in doubt, don’t put it in. That is the basic rule.

On the ground with AI, there is always a moment of “it’s handy, so let me just paste it in”. I too have watched, again and again on development and sales-support projects, people tempted to paste a document straight in. But never skipping that small extra step is what underpins the trust an organisation needs to keep using AI.

Writing emails

In writing emails, generative AI is a relatively easy area to use.

For writing emails, “have AI produce three options, then let a person choose and revise” is a realistic way to work. Comparing several options makes it easier to settle on the wording than using a single one as it stands.

If anything, I think the more you have AI write an email, the more firmly you should hold to the premise of “not using it as it stands”. The wording may be tidy, yet a sense of unease can linger that these are not your own words.Having a person check it at the end — that small extra step is, surprisingly, important in business writing.

Minutes and meeting notes

In drawing up minutes, generative AI is suited to summarising and organising.

Some AI summary tools let you switch between automatic generation and custom generation. Where you want the same format every time, as with a recurring meeting, specifying headings under custom generation, such as “agenda, decisions, to-dos, homework for next time”, makes it easier to keep the shared format consistent.

Kanata too lets you switch between automatic and custom generation for AI summaries. In organisations with many tasks they want summarised in the same format, such as meetings and training, this sits well with an approach of setting up a summary format and reusing it.

What I sense watching companies run their meetings is that the trouble with minutes is not only “the time it takes to write them”. The way they are written differs from meeting to meeting. Decisions are vague. The owners of to-dos are unclear. Let these pile up and follow-through after the meeting slows. An AI summary can be used not merely to save time, but to bring everyone to a shared understanding after the meeting.

Research work

In research work, care is needed in how generative AI is used.

In research, it is safer to ask AI “how to look into it” than to ask it for “the answer”.

I’m writing a report on bringing generative AI into the company. Please give seven points I should research, and for each set out the primary sources to confirm and the people to interview internally. Don’t assert figures; where needed, write “source required”.

Used this way, you can quicken the early stages of the research while keeping the risk of AI’s misinformation in check.

In my view, AI’s value in research lies less in the answer itself than in “raising the quality of the questions”. What should be looked into. Whom to ask. Which premise has been left out. Just sorting that out changes the precision of the research considerably.

Producing materials

In producing materials, generative AI can be put to use for structuring and tidying the prose.

In producing materials, the important thing is not to try to have it produce a finished version from the outset. First draw up a structure, then organise the key points under each heading, and last produce the body. Asking in stages makes the quality easier to manage.

When I use AI to produce materials, I rarely ask in the first instruction for “a finished document”. I am more likely to use it in ways such as “raise the points a decision-maker will care about in this document”, “build the flow for a twelve-slide structure”, or “anticipate the counter-arguments”. Before a document can be made to look tidy, you need to settle what it should convey.

HR, appraisal, and management

In the HR and management arena, particular care is needed in how generative AI is used.

In appraisals and one-to-ones, using AI’s output as it stands can leave the words tidy yet failing to land with the person. Appraisal comments in particular need to rest on the person’s actual conduct.

Leave AI to “organise the facts” and “suggest ways of putting it”, and let the manager themselves choose the final words.

In my view, there is great potential in using AI in the management arena. But that is precisely why it should be used to sharpen one’s thinking and to free up more time to face the other person, not to take judgement off one’s own hands.

Line 1: AI does the first draft, people do the final

As a rule, generative AI’s output is a first draft.

Emails, minutes, materials, FAQs, appraisal comments; it is the same for every task. Rather than treating what AI produces as a finished article, use it on the premise that a person will check, revise, and finalise it.

Just having this rule does a good deal to ease the worry on the ground.

• You may have AI produce it, but you must not send it out as it stands
• Before anything goes outside the company, a person must always look at it
• Figures, proper nouns, and dates are checked against the source

Make it clear, in this way, that it is used on the premise of review.

When drawing up rules for generative-AI use, I recommend placing this one line first.

AI’s output is not a finished product but a draft awaiting checking.

That single line alone changes the tool’s standing. Rather than over-trusting AI, it becomes easier to treat it as an aid for moving work forward.

Line 2: AI makes the material for a judgement, people make the judgement

Generative AI is suited to producing comparison tables and organising the points at issue. But the final judgement is made by people.

Split it, for instance, as follows:

Hold firmly to this division of labour and generative AI works not as “a tool that blurs responsibility” but as “an aide that raises the quality of judgement”.

Looking at what AI has organised, people sometimes feel a sense of unease. That unease is valuable. Unease means the on-the-ground premises and hard-won experience may not have made it into AI’s tidy version. Take AI’s output as a starting point and move on to a better judgement. That stance matters in practice.

Line 3: decide in advance what information may be entered

In generative-AI use, the input matters more than the output.

However handy the use, put in information that must not be entered and the risk outweighs it.

At a minimum, share the following classification across the company:

When in doubt, don’t enter it. If needed, check with IT, legal, or the information-security lead before deciding. Sharing this basic stance with every employee matters.

In my view, rules for AI use should be put into words the floor can judge by in three seconds, rather than into “difficult regulations”. For example: “if in doubt, don’t enter it”, “don’t enter anything that would be a problem if it got out”, “mask anything that could identify an individual”. Rules as plain as that are the ones that work on the ground.

“What it can be used for” is left abstract

Told to “make use of generative AI”, staff on the ground cannot move straight away. Even in explaining it to all employees, abstract notes such as “it can be used for writing” or “it can be used for summarising” are not enough.

• For sales: tidying meeting notes, structuring proposals, drafting customer emails.
• For HR: training notices, building out the FAQ, draft one-to-one questions.
• For IT: first-line responses to queries, putting together procedure documents, summarising incident reports.

You need, in this way, to translate it into business tasks close to the department and the role.

When I design AI training, I rarely begin by explaining advanced features. I am more likely to start by working out together “where, in your job tomorrow, might you use this”. AI training that is not connected to the work on the ground inevitably ends as mere knowledge.

The prohibitions are what registers first

Generative AI carries risks. Warnings about data leakage and misinformation are therefore indispensable.

But when the prohibitions are all that gets stressed, the ground takes it as “safest not to use it after all”.

What matters is to convey “prohibited” and “encouraged” as a set.

• Don’t enter personal information. But, masked into a form that cannot identify an individual, it can be used to tidy up the wording of a query.
• Don’t leave contract judgements to AI. But it can be used to surface the points to check with legal.
• Don’t take figures on trust. But it can be used to organise the research items and the report structure.

By showing not only “what is not allowed” but “how to use it well”, the ground finds it easier to act.

In driving AI adoption, fear alone does not make it stick. In driving AI adoption, fear alone does not make it stick. Putting into words the “realistically usable range” that sits in between is, in my view, a large part of the role of AI-adoption support.

The good ways of using it stay locked in individuals

Even where there are staff using generative AI well, if those ways stay locked inside the individual, they don’t spread across the organisation.

Good prompts, handy summary formats, and usable checklists need to be shared with the team.

The way of sharing need not be a dedicated tool. You can start with an internal wiki, a knowledge base, a shared folder, or a pinned chat post. That said, where you want to manage AI chat, summaries, training data, and prompts together by unit of work, an environment that lets you build a library by project, as Kanata does, is useful.

To spread generative-AI use through an organisation, it matters not only to “increase the number who can use it” but to “leave a record of how to use it”.

I find the companies where AI use goes well have something in common. It is that they don’t depend on one outstanding individual. A good way someone has found is turned into a team template. It is shared at the regular meeting. It is left in the library. This unglamorous accumulation is, in the end, what makes the big difference.

What generative AI is good and bad at

First, make clear what generative AI is good at and what it is not.

Sharing this distinction alone does much to reduce hesitation on the ground.

In my experience, what first gets people excited in generative-AI training is the “look, it can even do this”. But what really leads to it sticking is the shared understanding of “this we don’t leave to it”. Set what it can do alongside what it cannot, and only then do you have a footing from which people can use it with confidence.

Information you may and may not enter

Next, teach the rules on input information.

In generative-AI training, the handling of information should be explained before how to write prompts. Learn the handy uses, but enter information that must not be entered, and as business use it is dangerous.

In training, exercises such as these work well:

• Is this information fit to enter?
• Which parts should be masked?
• Whom should you check with?
• Should the handling differ between an external AI and an in-house environment?

Using concrete examples helps staff think in terms of their own work.

Customer information for a salesperson, employee information for an HR person, and system-configuration information for an IT person, for instance, each carry different points to watch. Set out the company-wide principles while preparing concrete examples by department.

The basic prompt frame

Instructions to generative AI that are vague produce vague output.

I recommend being clear about “role”, “purpose”, “audience”, “background”, “output format”, and “constraints” when writing a prompt. It is a frame that works well in company-wide training too.

Even in asking for an email, simply putting in elements like the following steadies the output:

Role

You are in charge of drafting external emails

Purpose

To invite people to take part in training

Audience

Department heads

Background

From next month, generative-AI training will run for all staff

Output format

Three subject lines; body within 300 characters

Constraints

Don’t lay it on too strongly; let the business need come through

Writing a prompt calls for no special literary skill. It is the skill of conveying a request clearly.

I take the ability to write prompts as the “asking skill” of the AI age. Whom, what, on what terms, in what form you ask. This carries over not only to AI but to working with people.

Rules for reviewing output

Last, teach how to check AI’s output.

At a minimum, share the following check items:

• Do the figures match the source?
• Are the proper nouns correct?
• Are the dates correct?
• Are fact and supposition kept apart?
• Is there any exaggeration?
• Is it information fit to go outside the company?
• Does it match the company’s tone?
• Is it content a person can take responsibility for explaining?

Use generative AI without this review habit and the risk may come to outweigh the convenience.

To doubt AI’s output is not to reject AI. If anything, it is the premise for using it well. I find that the better someone is at using AI, the less they take its output on trust.They question it, check it, and apply their own judgement to it. That stance is what makes for real skill in practice.

Start with AI chat, for drafting and tidying everyday work

The first step is to use AI chat to lighten everyday work a little.

Take uses such as these, for example:

• Drafting emails
• Rephrasing text
• Drafting meeting agendas
• Structuring reports
• Writing training notices
• Drafting FAQ answers

Kanata’s AI chat lets you ask, consult, and request work in a conversational form. It lends itself to moments in daily work such as “I want to tidy my thinking a little”, “I’d like a first draft”, or “I want to recast this differently”.

For all staff, a small start of roughly “use it once a day on your usual writing” suits best. Rather than entrusting complex work from the outset, it matters to build small successes within the everyday.

Next, use AI summaries to ease the load of meetings and document-tidying

The next step is putting AI summaries to use.

Many departments have work that calls for summarising, such as meetings, training, sharing materials, and tidying queries.

You are likely to feel the benefit in work such as the following in particular:

• Minutes of regular meetings
• Shared notes from department meetings
• Tidying the key points of long PDF documents
• Summarising training videos
• Organising customer-interview notes

Kanata’s AI summaries let you switch between automatic and custom generation. For routine work, specifying the output format under custom generation makes it easier to summarise in the same structure every time.

For my part, I think AI summaries are a feature well suited to the early stage of a company-wide rollout. Reading time falls. Sharing speeds up. Tidying up after a meeting gets easier. The benefit is easy to feel.

Use e-learning to drive company-wide reskilling

Generative AI is partly learned by using it, but for a company-wide rollout, basic training is needed too.

Kanata’s e-learning lets you run internal training, new-joiner education, and self-study around video content. You can summarise a video’s content with AI, or ask AI about anything unclear while taking the course, so the tool supports the learning experience as the training proceeds.

You could prepare training content such as the following, for instance:

• What generative AI can and cannot do
• Information you may and may not enter
• How to write a basic prompt
• Use cases by department
• How to review AI output
• How to respond in an incident

Share the same grounding with all staff, then move on to practice department by department, and you can keep variation in use in check.

In my view, It is important not to make generative-AI training a one-off event. Rather than having people grasp everything in the first session, combine videos, FAQs, hands-on sessions, and an internal community to add to the ways of using it little by little. That tends to stick better on the ground.

Use the library to leave the good ways of using it within the organisation

Last — and no less important — is building up prompts and training data.

Generative-AI use begins, at first, with individual trial and error. But to turn it into results as an organisation, you need to keep, share, and improve the ways that worked.

You could build a library of things such as the following, for instance:

• Minutes-drafting prompts
• External-email draft prompts
• FAQ-answer prompts
• Proposal-structure prompts
• Training-content creation prompts
• Internal-regulation data
• Product-explanation materials
• Collections of common queries

In Kanata, you can manage prompts, training data, and AI settings by project, so you can keep them in a form that is easy to reuse by department and by task.

To establish generative AI as an organisational foundation — rather than leaving it as a personal convenience — this accumulation is indispensable.

In my view, the real difference in AI use shows not the moment a tool goes in, but three or six months later. Are the good prompts still there? Are the old materials kept up to date? Are the on-the-ground successes being shared? That accumulation of practice is what becomes an organisation’s capacity to use AI.