AI tools for business, explained

What AI tools for business actually do

An AI tool earns its place when it does one named job faster than the person who used to do it, and when you can tell afterwards whether it worked. In practice that job is almost always one of four: reading unstructured text and turning it into structured data, drafting something a human then checks, answering questions over your own documents, or writing and reviewing code. Everything else in the category is a packaging decision. So the first question is not which model is best, it is which task you are buying, what it costs today in human time, and how you will measure the difference.

The four kinds of product in this category

General assistants. Claude, ChatGPT, Gemini, Mistral Le Chat. One subscription per person, a chat window, and increasingly a workspace: on 10 November 2026 Anthropic put spreadsheets, whiteboards and data analysis inside Claude and sent it into video calls. Right for drafting, analysis, research and ad hoc work. Wrong as the system of record for anything.

AI inside software you already pay for. Gemini is included in Google Workspace Business plans, Copilot sits in Microsoft 365, and inbox, accounting and POS products increasingly ship their own assistant. Right as the first thing to try, because the marginal cost is often zero and the data stays with a processor you already contracted. Wrong if the embedded feature is a demo rather than a product.

Task-specific tools. Fyxer for mail triage and drafting, document extraction for supplier invoices, transcription, translation, AI coding agents. One job, a measurable baseline, a price you can defend. Right where the task is repetitive and the error is cheap to catch. Wrong where a mistake reaches a customer unseen.

Models and agents you build on. The APIs underneath, priced per token rather than per seat, plus the connector standards that let an assistant read your systems. Anthropic's Fable 5.2 and Mythos 5.2, released on 9 November 2026, are priced at 8 dollars per million input tokens, down from 10. Right when the work is embedded in your own product or workflow.

What actually matters when you choose

1. Name the task and the baseline. Hours a week, error rate, backlog. Without a before figure, you cannot tell a productivity gain from a new subscription, and almost nobody writes one down.
2. Where does the data go, and is training off? The vendor is a processor and the model provider a sub-processor, both yours to account for under the GDPR. Ask for a dated sub-processor list, confirmation that your content is not used for training, and the retention period in writing.
3. Does it work in your languages? Test French, German and Portuguese output, not English. Quality drops unevenly between languages, and Luxembourgish is usually absent entirely.
4. The cost model, not the headline. Per seat, per token, or per credit. Credit systems are where budgets go wrong: the same subscription can cost three times as much in a heavy month. Ask what happens when credits run out.
5. Can it reach your actual systems? An assistant that reads your mailbox, your ledger or your catalogue is a different product from one you paste into. Integration is where most of the value sits, and most of the privacy risk.
6. What happens when it is wrong? Keep the human on the send button. Ask what audit trail exists, who can see the prompt history, and whether an output can be traced back to a source document.
7. Key and secret hygiene. If you build with APIs, treat the key like a bank card. Railway reported on 10 November 2026 that an attacker took AI provider keys belonging to 1 852 users and ran up an estimated 740 000 dollars of fraudulent usage.

What do AI tools cost in 2026?

Published prices, read on 11 November 2026, USD figures as published.

The useful read: a capable assistant costs 17 to 25 dollars per person per month, AI inside a suite you already buy often costs nothing extra, specialised tools run 30 to 40 dollars, and anything per token depends entirely on volume.

The mistakes buyers make

Buying seats before naming the task, which produces enthusiasm in month one and a renewal argument in month twelve. Paying twice, because the AI you just bought is already in the suite you pay for. Letting the tool send, post or pay rather than draft. Pasting client data into a personal consumer account with no processor contract behind it. Assuming a European interface means European processing. And measuring adoption, seats used, instead of the outcome you said you were buying.

Europe and Luxembourg specifics

• The AI Act applies in stages. The regulation entered into force on 1 August 2024, the prohibited practices and the AI literacy duty applied from 2 February 2025, and obligations for general purpose models from 2 August 2025. Later phases cover high-risk systems, and that timetable has been the subject of proposed amendment, so confirm the current position rather than quoting a date from a vendor slide.
• AI literacy is an obligation, not a nice idea. Providers and deployers have to take measures to ensure a sufficient level of AI literacy among the staff who use these systems. A one page internal policy and a short training record is the cheapest way to be able to show it.
• Transparency where a human might be fooled. Where people interact with an AI system or see generated content, that has to be disclosed. In practice: label the chatbot, label generated images, and do not let a customer believe they wrote to a person.
• The GDPR chain still decides most cases. Article 28 obliges you to contract your processors properly and to know the sub-processors, and a transfer outside the EEA needs its own mechanism. Mistral and the EU regions of the large providers are the usual routes to keeping the chain inside Europe.
• Professional secrecy and staff monitoring. In Luxembourg, lawyers, fiduciaires, healthcare and the financial sector carry duties beyond the GDPR, so client content in a model prompt is a compliance decision. If a tool measures how your employees work, employee representation has to be involved before it is switched on.
• Languages. A Luxembourg business writes in French, German, English and Portuguese, often in the same thread. Test the output in each one during the trial.

Questions to ask an AI vendor

• Which exact job does this do, and what should I measure before and after?
• Who processes my content, in which country, and which sub-processors are on the dated list?
• Is my data excluded from training by default, and what is the retention period?
• Is the price per seat, per token or per credit, and what happens when credits run out?
• What is your statement on the AI Act obligations that apply to your product?

Where to go next

For the two categories where we have already done the work: the best AI coding tools and the best AI email clients. The privacy checklist that applies to almost any AI tool handling customer text is in AI email and the GDPR, and the plain explanation of what an assistant does with your mail is in what does an AI email assistant actually do. The price of the models underneath moved again in Anthropic's Fable 5.2 and Mythos 5.2 release, and the cost of careless keys is in the Railway breach. If you sell software and want assistants to recommend it, read how to get your software recommended by AI assistants and then get listed.