Most "AI for business" advice is either a vendor pitch or a list of tools with no sense of which ones matter. This is the version we would give a friend running a company: where AI actually pays back, where it does not, and what to do first.
Where it reliably pays back
Writing that already exists in draft form. Rewriting, shortening, adjusting tone, turning notes into prose. This is the highest-confidence use — the model is editing rather than inventing, so the failure modes are small.
Summarising volume. Long threads, meeting transcripts, support tickets, research. Anything where the value is "tell me what's in here" rather than "tell me something new."
First drafts of structured documents. Job descriptions, SOPs, briefs, proposals. You will rewrite half of it, and half is still a big saving from zero.
Customer support triage. Categorising and routing incoming messages, and drafting replies a human then approves. The approval step is not optional.
Code. For teams with developers, this is often the single largest measurable gain — see AI coding tools.
Where it does not
Anything where being wrong is expensive and hard to notice. Financial figures, legal interpretation, medical guidance, compliance. Models produce confident, plausible, wrong answers, and "plausible" is exactly what makes them dangerous in these contexts.
Work that needs current facts, without web access. A model without live search will answer from a training snapshot with no signal that it is out of date.
Anything requiring exact arithmetic. Use a spreadsheet.
Fully unsupervised customer contact. The reputational downside of one bad automated reply usually exceeds the labour saved.
What to actually do first
Start with one workflow, not a strategy.
- Pick a task your team does weekly that involves writing or reading a lot of text.
- Run it manually with AI for two weeks. No integration, no automation — just people using a chat window for that one job.
- Measure the honest thing: time saved minus time spent fixing output.
- Only then automate, and only if step 3 came out positive.
Most failed AI projects skipped straight to step 4.
Data, before anyone pastes anything
This is the part that gets skipped and later becomes a problem.
- Free tiers are the most likely to train on your conversations. Paid tiers usually default to not doing so; business tiers normally exclude it contractually.
- Assume anything in a free tier could be read by a human reviewer. Fine for a blog draft. Not fine for a contract, a customer list, or unreleased financials.
- Write one rule your team can remember. "Nothing under NDA, nothing with customer names, nothing not yet public" beats a policy document nobody reads.
What it costs
Consumer plans run about $20 per person per month. Business and team tiers cost more per seat and add admin controls, usage reporting, and data-handling terms.
The honest arithmetic: at $20/seat, AI pays for itself if it saves roughly twenty minutes a month per person. That bar is low enough that the real question is not cost but whether people actually adopt it.
Adoption, not price, is where this fails. Buying seats nobody opens is the most common waste in this category.
Which model for which job
- Customer-facing writing → Claude, for prose that needs less editing
- Research and current information → Gemini, for live search
- Code → Claude for existing codebases, GPT for ecosystem breadth
- High-volume simple tasks → the smallest fast model, not the flagship
Using a flagship model to reformat spreadsheet headers is the most common way teams overspend. More detail: which AI model to use for what.
Common questions
How can AI help my business? Most reliably: editing writing that already exists, summarising long documents, drafting structured documents, and triaging support. Least reliably: anything requiring exact figures or current facts without web access.
How much does AI cost for a business? Around $20 per person per month for consumer plans, more for team tiers with admin controls and data guarantees.
Is it safe to put company data into AI tools? Depends on the tier. Free tiers commonly train on your input; paid and business tiers usually do not. Check the setting and set one simple internal rule.
What is the best AI for business? No single one. Claude for customer-facing writing, Gemini for research, and a small fast model for high-volume tasks.
Do I need a custom AI or is a subscription enough? For almost every small and mid-sized company, a subscription is enough. Custom builds make sense when you have proprietary data the model must know and enough volume to justify the engineering.
How do I know if it is working? Time saved minus time spent correcting output, measured on one workflow for two weeks. If that number is not clearly positive, do not scale it.