AI in business: what to hand over without having to redo it

You asked artificial intelligence to write a reply to a customer.

Within seconds, it produced a polite message. Very polite, in fact. It promised fast delivery, a personalised discount and support at every stage.

All you had to do was remove the discount, clarify the delivery terms, rewrite half the sentences and check where it got the idea of free returns.

The reply was generated quickly. The work was not finished yet.

That distinction matters in business. Judge the value of AI by the usable finished result and the effort it takes to get there. A polished message in a chat window does not, on its own, tell you whether your work has become easier.

How do you choose a task where artificial intelligence will actually help?

“I want to introduce AI” is an understandable ambition, but too broad a task.

A more useful question is: what recurring work do we do, what do we receive as input, and what result do we need?

For example:

• turn meeting notes into a list of agreements;
• extract order details from a customer’s enquiry;
• draft a reply using agreed terms;
• turn verified product specifications into a description;
• group reviews by topic for further analysis.

The US Small Business Administration lists meeting summaries, repetitive tasks and content preparation among the possible uses of AI for small businesses. These are possible directions; you still need to test the value of a particular application in your own work.

For your first attempt, choose one small task. Ideally, you know what a good result looks like and can spot an error before it affects a customer.

Drafting a standard reply using verified terms is a more manageable starting point than letting AI conduct every negotiation on behalf of the company on its own.

Suppose you need to insert an order number, date and amount into a document each time. If those details already exist in a spreadsheet, a regular template or a configured data transfer may handle the task.

AI can be useful when you need to work with different wording: understand an enquiry, extract the relevant information or summarise a conversation.

But use a formula for an accurate calculation. Use accounting records to check a payment. Use up-to-date inventory records to confirm availability.

Sometimes the most useful way to “introduce AI” starts with getting your price list in order. Then you can see which parts of the process actually need its help.

Generative AI can produce incorrect statements and present them confidently. NIST describes this risk in its Generative AI Risk Management Profile, including fabricated explanations and references.

For a business owner, this can look quite ordinary. A reply suddenly includes a feature the product does not have. A meeting summary includes a deadline nobody agreed to. A service description includes a guarantee the business does not offer.

Asking it to “not make things up” is useful as part of the task, but does not replace checking.

So decide from the start which information the system should use, how it should flag gaps, and who checks the result.

If a deadline is unknown, the draft should retain a note saying “needs clarification”. It makes the text less smooth, but avoids creating an unapproved promise.

Here are a few options for a trial. Choose based on your work, rather than the number of available features.

• Meeting summary. What to provide: notes or a recording you are authorised to process. What to check: whether decisions, suggestions and open questions have been separated correctly.

• Draft customer reply. What to provide: the enquiry and current terms. What to check: prices, deadlines, commitments and whether the tone is appropriate.

• Extracting order details. What to provide: the enquiry text and a list of required fields. What to check: whether any details have been omitted or invented.

• Product description. What to provide: verified specifications. What to check: materials, included items, dimensions and stated properties.

• Grouping reviews. What to provide: texts without unnecessary personal data. What to check: whether topics have been identified correctly and conclusions can be traced back to the reviews.

Every task has limits.

A conversation recording may be incomplete. A customer may describe a size ambiguously. A review may contain irony. A good workflow allows cases like these to be handed over to a person.

Do not push for a confident answer when there is not enough information.

Imagine a fictional packaging manufacturer. A manager regularly rereads enquiries to extract the box type, dimensions, quantity and printing preferences.

You can try handing this preparatory work over to AI.

There is something concrete to check here. The system should not determine the size itself, choose a material or promise delivery.

This result helps the manager prepare for the next conversation. Responsibility for the proposal and order confirmation remains clear.

For our example, you can use this starting point:

This is a starting point for testing, not a guarantee of accuracy. Test it on different enquiries: short ones, incomplete ones, and ones with corrections or conflicting requests.

For example, a customer first writes “200 units”, then adds “perhaps we will start with a hundred”. The result needs to show that clarification is required. Confidently choosing a number would only hide the problem.

One successful answer shows that the system handled one task.

To decide whether to use it regularly, you need a range of examples. Take material you are authorised to process and add difficult cases:

• important information is missing;
• some information was corrected later;
• there are several possible interpretations;
• the text contains errors;
• the customer requests a term the company does not offer.

For each result, record whether you can use it, what corrections it needs and how long they take.

Pay particular attention to errors that are easy to miss. An awkward sentence stands out immediately. An invented deadline among otherwise correct details can look completely natural.

If these errors keep recurring, narrow the task, change the input or postpone this application. Not every piece of work should be handed over to AI simply because it agrees to do it.

Suppose preparing a reply manually takes 12 minutes in our fictional business.

With AI, it takes:

• preparing and inserting the data - 2 minutes;
• getting a draft - 1 minute;
• checking and correcting it - 5 minutes.

Total: 8 minutes. In this example, the saving is 4 minutes per reply.

If checking and correcting takes 10 minutes, the total becomes 13 minutes. In that case, this particular way of using AI does not save time.

Costs also include setup, maintenance, the tool itself and the consequences of errors. Other benefits might include a more convenient result format or less fatigue from repetitive work. But evaluate those separately too.

Look at how much work remains after generation. That is often where the real cost of a quick answer is hiding.

Before pasting work correspondence into a new service, check its data processing terms and the settings of the particular account.

For a trial task, it is often enough to remove names, phone numbers, addresses and other unnecessary details. If you need to extract box specifications, the customer’s phone number will not help.

Do not paste passwords, access keys or information you do not have the appropriate permission to process. Removing a name alone does not always make material safe to share: order details, prices or agreements may still be confidential.

For ongoing use, the team needs a simple rule: which tools are allowed, which data can be uploaded, and who is responsible for checking.

Preparing the text of a reply is one step. Sending it to the customer is the next.

While the result remains a draft, an error can be corrected before use. When the system sends messages, changes orders or publishes material on its own, consequences may arise before anyone notices the problem.

So start with preparing the result and having a person check it. For further automation, separately define the permitted actions, the conditions for stopping, and a way to see exactly what was done.

Having access to a work system does not mean AI needs all your permissions. The connection should fit the specific task.

We have a separate article on why businesses should understand their own access rights: “Contractor disappeared. Who controls your website?”.

Choose one task and describe the expected result. Prepare permitted input data, examples of correct work and a way to check it.

Then carry it out manually and with AI across several different cases. Compare quality, time spent and the difficulty of making corrections.

After that, you can decide whether to continue. Perhaps the tool helps. Perhaps you need better data. Or perhaps a regular template is enough for this work.

A good first result is fairly modest: one recurring task has become easier, you understand its limits and know who checks the work.

For a business, that is already more useful than a dozen impressive demonstrations after which nobody knows what to do tomorrow.

Sources checked on 9 October 2026. The packaging example and time calculation are illustrative; the proposed trial is an editorial recommendation.

If you want to work out which task in your business suits a first attempt, write to Grandma. We will start with a specific piece of work you already have to repeat.