AI for business: a practical guide for company leaders

In short: AI pays off in business where work is frequent, based on text or existing data, and checked by a person: drafting, reading documents, answering common questions, summarising and looking things up. It does not replace judgement, accountability or relationships. Pick one concrete task, measure it before you start, pilot with a small group and only then scale.

According to McKinsey (2025), 88% of survey respondents said their organisation uses AI regularly in at least one business function, up from 78% a year earlier. The same research found that most organisations are still experimenting rather than running AI across the whole company. The gap between “we tried it” and “we rely on it” is where leadership decisions matter most.

What does AI for business actually mean?

AI for business means putting artificial intelligence into everyday work across departments so that reading, writing, sorting, searching and forecasting get done faster. In business terms, think of it as a very fast assistant that never tires, produces good first drafts, and sometimes gets things wrong.

Three terms tend to get mixed up:

• Generative AI creates new content: text, images, summaries, email replies. ChatGPT, Claude and Gemini are all generative AI tools.
• Large language models (LLMs) are the engines inside those tools. They are trained on huge amounts of text, so they can read and write natural language, Vietnamese included.
• Traditional machine learning predicts and classifies from a company's own numbers, for example demand forecasts or lead scoring.

The next step beyond generative AI is the AI agent, an assistant that carries out a chain of tasks inside your systems instead of only answering. See What is an AI agent? for that topic.

What AI does well, and where it falls short

AI is strong at patterned, text-heavy work where a small mistake can be caught and fixed. It is weak wherever absolute accuracy is required and nobody checks the output.

AI does well	AI does poorly or is risky
Drafting emails, quotes, product descriptions, posts	Giving exact figures when it isn't connected to real data
Reading invoices, contracts and CVs and extracting fields	Making decisions with legal or financial consequences on its own
Summarising meetings, long documents, customer feedback	Understanding internal context that was never written down
Answering repeated questions from existing documents	Handling an angry customer or a situation with no precedent
Classifying emails, tickets and feedback	Keeping secrets when staff paste sensitive data into free tools
Forecasting when historical data is reasonably clean	Working on messy, duplicated input data

The biggest weakness of generative AI is hallucination: fluent, confident answers that invent numbers or policies. Every serious deployment therefore needs two things. The AI must be grounded in the company's own documents and data, and a person must review the steps that matter.

AI use cases by department

Each department has a few tasks AI can help with today using off-the-shelf tools, and a few that need a connection to company data first.

Department	Easy starting points	Needs connected data
Sales and marketing	Content, outreach emails, call summaries	Quotes drafted from the price list, lead scoring in the CRM
Customer service	Suggested replies for agents	Chatbot answering from documents, order look-ups
Finance and accounting	Explaining regulations, drafting letters	Reading supplier invoices, reconciling balances, management reports
HR	Job descriptions, interview questions	CV screening, an assistant for internal policy questions
Operations, production, warehouse	Writing procedures and work instructions	Demand forecasting, low-stock alerts, defect analysis
Leadership	Summarising reports, preparing meetings	Asking business questions about live data in plain language

The middle column mostly needs training and a clear rule about which data may go into which tool. The right-hand column needs clean data, access control and an owner, so it takes more work and makes a bigger difference. The three most requested areas have their own guides: AI in sales and marketing, AI customer service chatbots and AI in finance and accounting.

Buy, configure or build?

Most small and mid-sized companies should start with tools that already exist and customise only where they have something unique. Building from scratch makes sense only when the problem is tied to your own process and your own data.

Approach	Fits when	Effort	Watch out for
Business plans of AI assistants (ChatGPT, Claude, Gemini…)	Many people need drafting, summarising, document analysis	Low: choose a plan, set rules, train	Pick a plan that does not train on your data; manage accounts centrally
AI features in software you already use (CRM, accounting, sales)	Your current software already has a suitable feature	Low to medium	Same features for every customer, little differentiation
Custom solution connected to internal data (document chatbot, AI agent)	Your process is specific and data sits in several systems	Medium to high: design, API integration, testing	Needs a data owner and a maintenance plan
Training your own model	You have an engineering team and lots of data	Very high	Rarely needed for SMEs

Leaders often ask whether to choose Claude or ChatGPT. Choose by the job, the data terms and the office suite you already use (Microsoft 365 or Google Workspace), then let a small team try both on the same real documents. The difference between leading models is usually smaller than the difference between a well-designed process and one nobody thought through.

How ready does your data need to be?

Data does not have to be perfect, but it has to be good enough for the chosen task. A customer chatbot needs up-to-date policy documents and FAQs. A quoting assistant needs one consistent price list and product codes. Stock forecasting needs several seasons of sales history without duplicate codes.

Before starting, answer four questions for your first use case:

1. Where does the data live: which system, which file, who owns it?
2. Is there one correct version, or does every team keep its own spreadsheet?
3. Who may see which data, and must the AI respect the same permissions?
4. Which data is personal data about customers or staff?

If the answer to question 2 is “every team has its own file”, the first job is cleaning up data, not AI.

Risks and how to control them

The three most common risks are wrong answers, data leaks and vendor lock-in. All three can be managed if the rules are set at the start.

Wrong answers

Hallucination drops sharply when the AI may only answer from documents and data you provide, cites its sources so users can check, and is allowed to say “I don't know”. For anything involving money, contracts or commitments to customers, a person approves the final step.

Data leaks

Staff pasting payroll, customer lists or contracts into free AI tools is the most common leak. Write a short policy: which tools are allowed, which data must never go in, and company-issued accounts only. In Vietnam, the Personal Data Protection Law (No. 91/2025/QH15), in force since 1 January 2026, sets clearer duties for anyone processing personal data, including with AI.

Lock-in and transparency duties

Design so you can swap the AI model without rebuilding everything: your documents, prompts and business rules should belong to you. Vietnam's Law on Artificial Intelligence (No. 134/2025/QH15) took effect on 1 March 2026; its Article 11 requires AI systems that interact directly with people to let users know they are dealing with AI. The full governance picture is in AI governance and data security.

How to choose your first AI use case

Your first use case should be small, frequent, measurable and owned by someone. These five steps keep the effort under control.

Step 1: List repetitive tasks by department

Each department head writes down the three most time-consuming tasks, how often they happen each week and how long each takes. Typical results: answering the same customer question, keying in invoices, preparing quotes, weekly reports.

Step 2: Score each task

Score frequency, data availability, the cost of an AI mistake and how easy it is to measure. A good first task is frequent, has data ready, is easy to correct when wrong and can be measured in time or volume.

Step 3: Measure the baseline

Record today's numbers: average time per case, error rate, backlog. Without a “before” you cannot prove an “after”.

Step 4: Pilot with a small group and a deadline

Pick 3–5 real users, fix a time box and log every wrong answer. The business owner, not IT, decides whether the pilot passed.

Step 5: Scale, fix or stop

Compare results with the baseline. If it worked, extend to other teams; if not, look for the cause in the data, the process or how people use the tool. Stopping is a good outcome when it saves you from investing in the wrong place.

If your chosen task is a chain of steps across several systems, read AI process automation vs RPA.

Costs, common mistakes and when to wait

Besides the AI tool's subscription, budget for these costs. The largest is usually internal staff time, not the licence.

• Usage-based fees: many AI services charge per user or per request, so cost grows with use. Ask vendors how pricing scales before signing.
• Design and data integration: for custom solutions, paid work to design, connect the AI to your documents and existing software, and test.
• Preparing documents and data: hours from experienced staff writing, fixing and checking the documents the AI will read.
• Process mapping and usage rules: managers' hours, even for off-the-shelf tools.
• Training: sessions plus trainees' hours.
• Monitoring and upkeep: an owner tracking wrong answers in the first weeks and updating documents when policies change, every year.

Common mistakes:

1. Buying the tool before choosing the job. Everyone gets an account, nobody is told what to use it for, and usage fades within weeks.
2. Picking a problem that is too big. “AI for all of customer service” can't be measured; “answer 20 delivery questions” can.
3. Skipping the reviewer. Letting AI send content straight to customers from day one turns a small error into an incident.
4. No business owner. AI projects owned by “IT” tend to stop at the demo.
5. Not logging errors. If wrong answers aren't tracked, the AI never improves and users lose trust.

Hold off on custom AI when processes aren't standardised and everyone works differently, because the AI will learn the mess. Also wait when the task happens only a few times a month, when one mistake is costly and nobody would catch it, or when the real problem is a lack of people or rules.

FAQ

Which department should start with AI? The one with lots of repetitive, text-based work and a manager willing to own the project. Usually customer service, sales or accounting.

Does a small company need its own IT team to use AI? Not for off-the-shelf tools. Connecting AI to internal software works best with an internal data owner plus an outside implementation partner.

Is it safe to use free ChatGPT for work? Not for customer, employee, contract or financial data. Use a business plan with data-protection terms and company-managed accounts.

Will AI replace employees? AI takes over repetitive parts of a role and rarely the whole role. People move to checking, handling exceptions and working with customers.

How much work is it to build a business chatbot? A chatbot that answers from existing documents is relatively quick; writing and maintaining those documents is the main effort. A chatbot that looks up orders or balances needs software integration and takes longer.

---

HAT Solutions helps leadership teams choose their first AI use case, measure the baseline and pilot with real data before a larger investment. Send us each department's list of repetitive tasks and we will help you rank them. Talk to us about applying AI, or read Where should digital transformation start? and Business process automation.