Key takeaways
- A large language model (LLM) predicts likely text. It does not look facts up in a database, and it does not “know” when it is wrong.
- It shines at drafting, summarising, rewriting and classifying — and is weak at exact figures, recent events and anything it was never shown.
- Start small: pick 2–3 repetitive tasks, run a short pilot, and measure the time actually saved.
- Write a short AI policy before staff paste customer data into free tools.
What generative AI actually does
Tools like ChatGPT, Claude and Gemini are built on large language models. A model is trained on a very large amount of text, and from it learns the patterns of language: which words and ideas tend to follow which. When you ask it something, it produces an answer one small piece at a time, each time choosing a likely next piece based on your question and everything written so far.
That simple mechanism produces surprisingly capable results — fluent emails, working code, clear summaries. But it explains the weaknesses too. The model is not searching a reliable database of facts, and it is not reasoning the way a trained accountant does. If the most “likely-sounding” answer is wrong, it will still write it, fluently and confidently. This is often called a hallucination.
What it is good and bad at
- Good at: first drafts, summarising long documents or meetings, rewriting text in a different tone or language, extracting fields from messy text, categorising requests, brainstorming, explaining technical topics, and writing or reviewing simple code.
- Bad at: exact calculations without a tool, facts after its training date (unless it can search the web), your internal data (unless you give it access), legal or financial judgement, and knowing the limits of its own knowledge.
A useful mental model: treat it like a fast, well-read new hire on day one. Excellent at producing a draft; needs checking before anything goes to a customer.
Practical uses by department
| Department | Use case | Human check needed |
|---|---|---|
| Sales | Draft follow-up emails, summarise call notes, prepare quotation descriptions | Yes — prices, terms and promises |
| Customer service | Suggest replies, classify and route tickets, build an FAQ from past answers | Yes before sending; light review for routing |
| Accounting | Explain variances in plain language, draft reminder letters, extract data from invoices | Always — every number must be verified |
| HR | Draft job descriptions, policies and onboarding checklists | Yes — fairness, local labour law, tone |
| IT | Write scripts, explain error messages, draft documentation | Yes — test code before running it on live systems |
| Marketing | Social posts, product descriptions, translations, campaign ideas | Yes — accuracy, brand voice, claims |
For tool-specific advice, see our guides to using ChatGPT at work and Claude for business.
How to start
- List repetitive tasks. Ask each team where they spend hours writing, copying, summarising or answering the same questions.
- Pick 2–3. Choose tasks that are frequent, low-risk and easy to check — internal summaries before customer-facing replies.
- Pilot with a small group for a few weeks, using a business-grade tool rather than personal free accounts.
- Measure. Record how long the task took before and after, and how much correction the AI output needed. If the checking takes as long as the work, drop it.
- Scale what works, document how it is done, and consider automation for the steps that no longer need a person.
The real risks
- Wrong answers. Fluent text can hide invented facts, figures or references. Anything that leaves the company must be checked.
- Data privacy. Text pasted into a consumer AI tool may be stored and, depending on the provider's settings, used to improve its models. Customer records, contracts, passwords and financial data should only go into tools your company has approved and configured.
- Over-reliance. Staff can lose the habit of thinking a problem through, or stop noticing errors because the output usually looks right.
- Bias. Models learn from human-written text and can repeat its biases — a real concern in hiring, credit or customer decisions.
A simple company AI policy
One page is enough to start. It should answer:
- Approved tools — which AI services staff may use, and with which (company) accounts.
- Data rules — what may never be entered: personal data, client confidential information, credentials, unreleased financials.
- Human review — which outputs must be checked, and by whom, before they are sent or acted on.
- Transparency — when customers or colleagues should be told that AI was used.
- Ownership — the person using the tool is responsible for the result, not the tool.
- Reporting — who to tell if sensitive data was shared by mistake or the AI produced something harmful.
The bottom line
Generative AI is a powerful writing and thinking assistant, not an oracle. Used on the right tasks, with clear rules and a human checking the output, it can give your team back hours every week. Used carelessly, it leaks data and spreads confident mistakes.
If you'd like help choosing the right tools, setting up a safe pilot and writing a policy that fits your business, talk to our team.
