AI for Business: Where to Start, What to Avoid and What It Really Takes

The short answer: Start with one repetitive, measurable and low-risk task, such as answering common customer questions, drafting routine messages or sorting incoming enquiries. Keep a person responsible for the result, protect your customers' data, measure the time saved or the enquiries won, and only expand once the first project works. AI is a tool for specific jobs, not a strategy on its own.

If you run a business, you have probably asked yourself some version of the same question: "We should be using AI, but where do we start?" Maybe a competitor has added a chat assistant. Maybe your team already pastes things into ChatGPT, Claude or Gemini and you are not sure that is wise. Maybe you are tired of hearing that AI will change everything and still cannot see what to do on Monday morning.

This guide answers that in plain words. It covers what AI can and cannot do for a business, eight practical uses that tend to pay off first, how to choose and run a first project, what it costs, the risks to manage and the mistakes to avoid. It is written for owners and managers, not engineers, and it avoids hype. Along the way you get three things to copy and use: a 20-question test for any assistant before customers see it, a sheet for measuring whether the project paid off, and a one-page AI rule for your team.

What Can AI Actually Do for a Business?

The AI tools most businesses meet today are built on large language models. In simple terms, they are very good at working with language and documents. They can:

• Draft and rewrite text in a given tone
• Summarise long documents, calls and email threads
• Answer questions in plain language, including from your own documents when you connect them
• Sort and label things, such as enquiries, tickets and invoices
• Pull structured information out of messy text, such as names, dates and amounts from a PDF

They also have real limits. They can state wrong things with complete confidence (often called "hallucinating"). They do not know your business, prices or policies unless you give them that information. They have no judgement about consequences. And they can be tricked by instructions hidden in the text they read.

That leads to a simple rule of thumb.

Eight Practical Uses of AI in a Business

1. Answering customer questions on your website

A chat assistant that answers from your own content (services, prices, opening hours, policies, FAQs) can handle the repetitive questions at any hour and collect the details of people who are ready to talk. This is especially useful when your customers are in other time zones.

Good first step: Connect the assistant only to content you have approved, let it say "I don't know" and hand over to a person, and capture the visitor's contact details.

Watch out: An assistant that invents prices or promises is worse than no assistant. Test it with the hardest questions your customers ask.

2. Drafting and editing routine writing

Quotes, follow-up emails, product descriptions, job adverts and social posts all start faster with a draft. The value is the time saved on the first version, not the final wording.

Good first step: Give the tool examples of your best past writing so the tone matches, and always have a person read and send.

Watch out: Generic, bland text. Add your real details, numbers and examples.

3. Searching and summarising your own documents

Policies, manuals, contracts, past proposals and meeting notes pile up. An AI tool connected to them can answer "What did we agree with this client?" or "What is our refund policy for orders over 30 days?" in seconds, and point to the source.

Good first step: Start with one set of documents that is current and well organised.

Watch out: Outdated or contradictory documents produce outdated, contradictory answers. Clean the source first.

4. Sorting and routing incoming work

Enquiries, support emails, invoices and forms can be labelled, prioritised and sent to the right person automatically. A person still handles the work. The AI just removes the sorting.

Good first step: Begin with labelling only, and compare its labels with a person's for a few weeks before letting it route anything.

Watch out: Rare but important cases, such as a complaint or a legal notice, must always reach a human.

5. Extracting data from documents

Typing details from invoices, receipts, application forms or delivery notes into a system is slow and error-prone. AI can read them and fill the fields for a person to confirm.

Good first step: Pick one document type with a consistent layout, and have a person approve each entry until the error rate is clearly low.

Watch out: Money, dates and identity numbers need verification rules, not blind trust.

6. Meeting notes and follow-ups

Tools can transcribe a call, summarise it and list the decisions and action items. That saves the "who was going to do what?" email.

Good first step: Use it for internal meetings first.

Watch out: Tell people when a call is being recorded and follow your local rules on recording and consent.

7. Qualifying leads and helping sales

AI can read an enquiry, tell you what the person wants, suggest which service fits and draft a reply for you to edit. It can also keep your customer records up to date.

Good first step: Use it to prepare replies and summaries, not to decide who deserves attention. The quality of what it summarises depends on what you ask people to send, so look at how a good quote request form is built first.

Watch out: Never let it contact people on its own at first, and never let it make up claims about your service.

8. Asking questions of your own numbers

Sales, stock and website data can be explored in plain language: "Which products sold best last quarter?" or "Where do our enquiries come from?" This turns reports that nobody reads into answers people can ask for.

Good first step: Check every answer against the original report for the first few weeks.

Watch out: The tool can misread a table or mix up periods. Treat early answers as leads to verify.

A ninth use sits outside your own operations: being found when customers ask AI tools for recommendations. Google's own guide to generative AI search says the same foundations apply as for ordinary search, and that you do not need llms.txt files or special markup. In practice it comes down to clear, accurate, original pages that search engines can index. I cover the how-to in SEO for business websites and what Google itself says in why SEO is important for business.

How to Choose Your First AI Project

Most failed AI projects share the same cause: they began with a tool and went looking for a problem. Begin with the problem. A good first project passes most of these tests:

1. It is repetitive. It happens many times a week, so a small saving per task adds up.
2. It has a clear definition of "good". You can say what a correct result looks like.
3. Mistakes are cheap. A wrong draft costs minutes, not customers.
4. A person can check it quickly. Review should be faster than doing the work.
5. The information exists. The documents, data or examples are available and reasonably tidy.
6. Someone owns it. One named person is responsible for results.
7. You can measure it. You know today's time, cost or volume, so you can see a change.

If a project fails more than two of these tests, pick a different one first.

A 30-Day Plan for Your First AI Project

1. Week 1, choose and measure. Pick one task that passes the tests above. Write down what a good result looks like. Measure today's baseline: how long it takes, how often it happens, how many mistakes occur.
2. Week 2, test with real examples. Try a ready-made tool on 20 to 30 real past cases, without confidential data. Compare its output with what your team actually produced. Note where it is right, wrong and unsure.
3. Week 3, run a small pilot. Let one or two people use it on live work, with a person reviewing every result. Write down simple rules: what it may do, what it may never do and when it must hand over.
4. Week 4, measure and decide. Compare time, quality and cost against your baseline. Then make one of three decisions: stop, adjust and repeat, or expand and connect it properly to your systems.

A Worked Example: A Customer-Questions Assistant

This is an illustration, not a client case. Imagine a small online shop that keeps getting the same questions on WhatsApp and by email: delivery times, returns, sizes, opening hours and "is this in stock?" Here is how I would take that from idea to live, in order.

1. Collect the real questions. Copy two weeks of messages into one sheet and group them by topic. Count how often each one appears. You will see which few topics come up again and again, and those are your starting list.
2. Write the approved answer once. For each topic, write one short, current page: the policy, the price, the exception and who to contact. This page, not the AI's guesswork, is the source of every answer.
3. Decide what it must hand over. Complaints, refunds above a set amount, legal or medical questions, and anything the pages do not cover. For these it says so plainly and passes the conversation to a person.
4. Build the smallest version. An assistant that answers only from those pages, says which page it used and offers "talk to a person" at every step.
5. Test it with the hardest questions from the list below before any customer sees it.
6. Run it as a drafting aid first. For a week or two, let it suggest replies that a team member reads, edits and sends. Compare its drafts with what your team would have written.
7. Go live in one place, clearly labelled. Tell visitors it is an AI assistant, keep the route to a person visible and read a sample of conversations every week.
8. Measure against your baseline with the sheet further down, and decide whether to expand.

Test an Assistant Before Your Customers Do

Before launch, ask your assistant these ten kinds of question, twice each, so twenty tests in all. Save the answers, and run the same list again whenever you change its instructions or its documents.

If it fails a test, fix the instructions or the source page first, then run all twenty again. Do not launch while it still invents prices or promises.

How to Tell Whether It Worked

In week 1 of the plan you write down a baseline. In week 4 you compare it with what you see now. Keep the sheet simple:

Here is an example with made-up numbers. Forty enquiries a week take 6 minutes each to answer, which is 240 minutes. With AI drafting and a person checking, each takes 3 minutes, which is 120. The saving is 120 minutes, or two hours a week. Whether that is worth it depends on what the tool and the upkeep cost, and on whether the replies are as good as before. If your team spends as long rewriting the drafts as they used to spend writing the replies, the project has not paid, however modern it feels.

Ready-Made Tool, No-Code Automation or Custom Build?

You do not need a custom system to begin, and you often should not start with one.

A sensible path is to prove the idea with a ready-made tool, automate it with a no-code platform if the volume justifies it, and build custom only when you need control, scale or deep integration with your own software.

What Does AI for Business Cost?

There is no single price, because the cost has several parts:

• Tool subscriptions, usually per person per month.
• Usage costs, when you use a provider's API: you pay according to how much text you process, so a busy chat assistant costs more than a quiet one.
• Build and integration time, if you connect AI to your website, software or data.
• People's time, because someone must review results and maintain the system.
• Ongoing care, such as updating the documents it relies on and checking its answers over time.

Two habits keep costs under control. First, start with a fixed-scope pilot with a clear question to answer ("does this save at least X hours a week?"). Second, set a spending cap and watch the usage reports from the first week. When I scope AI work for a client, milestones and pricing are agreed up front, and you can see how I work before you commit to anything.

The Risks, and How to Manage Them

Wrong answers

Because AI can be confidently wrong, design for it. Let it answer from your approved documents, show its sources, allow "I don't know", and keep a person in the loop for anything involving money, health, legal matters or promises to customers.

Customer and business data

Plans differ in whether your inputs may be kept or used to improve the provider's models. Read the data terms of the exact plan you use, and prefer business or API plans over free personal accounts for anything involving customers. Share the minimum data needed, and check the privacy rules that apply to you, such as GDPR in Europe or your local data protection law. If you collect personal data on your website, your privacy policy should match what actually happens to it.

Security

If an AI assistant can read emails, web pages or documents and also take actions, such as sending messages, issuing refunds or editing records, then hidden instructions in that content can try to steer it. This is called prompt injection, and it is the first entry (LLM01) in the OWASP Top 10 for LLM applications, 2025 edition. The defences OWASP lists are practical: give the assistant the smallest set of permissions it needs, keep outside content separate from your instructions, and require human approval for any action that is hard to undo.

Telling people they are talking to AI

In the EU this is now a legal matter. The AI Act's transparency rules apply from August 2026, and the European Commission describes the aim for chatbots simply: people should be made aware that they are interacting with a machine. Even where no law requires it, it is good practice, and it builds trust. I am not a lawyer, so check the rules in each market you sell to.

Decisions about people

Using AI to screen job applicants, set prices for individuals or decide who gets credit is a much higher-risk area. The EU treats uses like employment as "high-risk", with rules now due from December 2027. Keep people decisive in anything that affects someone's job, money or rights.

Brand and consistency

Anything a customer sees reflects on you. Test the assistant's tone and answers before launch and review samples of its conversations regularly.

If you want a structured way to think about all of this, the NIST AI Risk Management Framework offers a free, widely used model built around four ideas: govern, map, measure and manage. Version 1.0 dates from January 2023, and NIST says it is being revised, so check the current version.

A One-Page AI Rule for Your Team

Most of the risks above come down to people not knowing what is allowed. A short written rule fixes that. Copy this, fill in the blanks and give every team member a copy.

How we use AI

• We use AI for drafting, summarising, sorting, translating and answering from our own approved documents.
• We never paste in passwords, card or bank details, identity numbers, health information, private customer details or confidential contracts.
• We use only these tools and plans for work: [list them]. Free personal accounts are not for customer data.
• A person checks every AI output before it reaches a customer. Any exception needs [owner's name] to approve it in writing.
• AI does not decide anything about hiring, pay, credit, or legal or medical matters, or refunds above [amount]. A person decides.
• We check the names, numbers, dates and prices in anything AI wrote.
• Customers are told when they are talking to AI, and they can always reach a person.
• If an AI mistake reaches a customer, we tell [owner's name] the same day and fix the source it used.
• This rule is owned by [name] and reviewed every [three months]. Last reviewed: [date].

Ten Mistakes to Avoid

• Starting with the tool instead of the problem. Define the task first.
• Automating a broken process. AI makes a messy process faster, not better. Fix it first.
• Giving it no owner. Without a named person, nobody checks the results.
• Skipping human review. Remove review only after the numbers prove it is safe.
• Measuring usage instead of outcomes. "Our team used it 500 times" tells you nothing. Hours saved and enquiries won do.
• Giving it too much access. Start read-only and add permissions one by one.
• Pasting confidential data into free consumer tools. Set a clear team rule.
• Ignoring your team. People worry about their jobs. Explain what changes, train them and ask for their ideas.
• Building a chatbot nobody asked for. Look at your real customer questions first.
• Letting the source material go stale. An assistant is only as good as the documents behind it.

A Quick AI Readiness Check

• We can name one repetitive task that costs us real time every week
• We know how long that task takes today and how often it happens
• We can say what a correct result looks like
• A person can review the output faster than doing the work
• The documents or data it needs are current and organised
• One person owns the project and its results
• We know which data must never go into an AI tool
• Our team has a simple written rule about using AI
• Anything customer-facing will tell people when they are talking to AI
• We have a budget cap and a date to decide whether it worked

Frequently Asked Questions

Where should a small business start with AI?

Start with one repetitive task where a mistake is cheap and a person can check the result quickly. Common first choices are answering frequent customer questions from your own content, drafting routine emails, and sorting incoming enquiries. Measure the time or results before and after, and expand only if it works.

Is it safe to put customer data into AI tools?

It depends on the tool and the plan. Free personal accounts often have different data terms from business plans and API access, so read the terms of the exact plan you use. Share only the data that is necessary, keep sensitive data out of casual tools, and follow the data protection laws that apply to you. When in doubt, do not paste it.

Will AI replace my staff?

In my view, the best early uses of AI remove tasks, not people: the sorting, drafting and copying that fills a working day. Your team's judgement, relationships and responsibility still matter, and decisions that affect people should stay with people. The healthier approach is to use the time saved for work that grows the business, and to involve your team from the start. If you are a freelancer or a student rather than an employer, see how AI is changing freelancing.

How much does it cost to add AI to a business?

It ranges from almost nothing, when you use a ready-made tool for drafting, to a real project when you build an assistant into your own website or software. The costs are tool subscriptions, usage fees, build time, review time and maintenance. Start with a fixed-scope pilot and a spending cap, and judge it by hours saved or enquiries won.

Should I build custom AI or use ready-made tools?

Begin with ready-made tools to learn what works. Move to no-code automation when the volume justifies it. Build custom only when you need your own branding, tight control over data and cost, or deep integration with your software, such as a customer assistant on your own website.

What is the difference between a chatbot, an automation and an AI agent?

A chatbot answers questions in a conversation. An automation moves information between your tools by fixed rules, and may use AI for a step such as sorting or summarising. An agent goes further and chooses and performs several steps on its own, such as looking things up and updating records. The more freedom you give it, the more careful you must be with permissions and human approval.

Do I have to tell customers when I use AI?

Where a customer is talking to an AI system, such as a chatbot, the EU AI Act's transparency rules (which apply from August 2026) expect them to be made aware that it is a machine. Elsewhere the law varies by country and by use, so check the rules where you sell. Even when nothing requires it, telling people builds trust and costs you nothing. I am not a lawyer, so treat this as a starting point.

What should a business never use AI for?

Do not let AI make the final call on hiring, firing, pay, credit, or medical or legal matters, or on refunds and promises that are hard to undo. Do not paste passwords, card details, identity numbers or confidential customer data into casual tools. And do not let it send messages to customers without a person checking, at least until your tests and your measurements show it is safe.

How do I stop an AI assistant making things up about my business?

You cannot remove the risk completely, but you can shrink it. Let it answer only from approved, current pages. Tell it to say "I don't know" and hand over when the answer is not in them. Test it with the hardest questions your customers ask, including ones it should refuse. Then read a sample of real conversations every week and fix the source page whenever it gets something wrong.

How long before AI shows results?

A first pilot can give you a clear signal in about 30 days, if you chose a repetitive task and recorded the baseline first. Real savings tend to come after a few rounds of adjustment, so judge the project over a few months, not a week. Be wary of anyone who promises a fixed result before they have seen your process.

Ready to Put AI to Work in Your Business?

The businesses that benefit most from AI are usually not the ones with the biggest budgets. They pick a clear problem, start small, check the results and build from there.

I build chat assistants and workflow automations for businesses at home and abroad, along with the websites and software they connect to. If you have a task in mind, or you are not sure which one to start with, write a short brief on my contact page and I will reply with an honest view and next steps. You can also see examples of live client sites I have built.