10 things AI still can't do in business (and what it's now good at)

AI is everywhere in business software now. It drafts emails, summarises meetings, answers customer questions and, increasingly, carries out whole tasks as an “agent”. So it’s worth being clear about what it can’t do, because that’s where businesses get caught out.

The short answer: AI is now good at drafting and summarising, and at routine tasks with a person checking. It still can’t take responsibility for what it produces, or promise its facts are right. Keep a person in the loop.

What AI is now genuinely good at

• It drafts emails and proposals from your notes.
• It summarises long threads and meetings.
• It searches across your files to find an answer.
• It handles routine multi-step tasks, such as processing invoices or sorting enquiries, as an AI agent with a person approving the result.

10 things AI still can’t do in business

1. Take responsibility. If an AI-drafted contract clause or email is wrong, the person who sent it is accountable, not the AI.
2. Guarantee facts. AI can produce answers that are confident and wrong. Anything important needs checking.
3. Understand your business without your data. It only knows what it has been given. Without your processes, clients and history, its advice is generic.
4. Build real relationships. Clients buy from people they trust. AI can help you respond faster; it can’t replace the relationship.
5. Make ethical and values-based judgements. Decisions about people and fairness need human judgement and accountability.
6. Read the room. Tone, politics, a client’s mood on a call: AI is improving, but it still misses the human signals that matter.
7. Handle the truly unexpected. AI works from patterns. A situation unlike anything it has seen can produce poor results without warning.
8. Fix bad data. If your records are incomplete or inconsistent, AI will reflect that, faster.
9. Keep your secrets for you. Paste confidential data into a personal AI tool and it may leave your control. Use business versions with data protection.
10. Replace security basics. AI helps attackers write convincing phishing and clone voices, and the NCSC expects it to make attacks more frequent (NCSC, May 2025). Multi-factor authentication, and a phone call before any unusual payment, matter more now.

What has changed since we first wrote this

• AI agents arrived. AI can now carry out tasks across systems as well as chat. Good uses keep a person approving the outcome.
• AI is built into Microsoft 365. Copilot works inside Outlook, Word, Excel and Teams. See Copilot for small business.
• Attackers use AI too. See phishing email examples.
• Mistakes come with the new. You’re replacing tried and tested ways of working, so errors are more likely at first. That’s why our agents start by drafting, not sending.

How to use AI safely in a small business

1. Use business versions of AI tools, signed in with work accounts.
2. Tidy permissions so AI only sees what each person is meant to see.
3. Keep a person in the loop for anything sent to clients or affecting money or people.
4. Write simple rules: what data AI can and can’t be used with.
5. Start small, with one or two processes where time savings are easy to measure.

Where an AI agent helps

Because of those limits, the agents we build keep people in control:

• The AI Data-Guardrail Agent ↗ watches for company data going into public AI tools.
• The AI Process-Audit Agent ↗ maps your processes and shows what to automate first.

See all our agents in the solutions library on jerait.ai ↗, or find the right agent for your business on jerait.ai ↗.

The catch: the AI an agent runs on is paid for by use. With Jera IT support, your first agent is built, supported and maintained free; we give you a rough idea of the usage cost before you commit, and a real figure from the trial.

How Jera can help

We help businesses across Edinburgh, Glasgow and Aberdeenshire get Microsoft 365 ready for AI, starting with who can see what. Bring us one job your team does every day, and book a call to find out whether it’s worth handing to AI at all.