AI is getting cheaper, but the bills and risks remain. What it means for your business
Meta has released a cheap model for AI agents, while experts warn about chatbots given too much access. What this means for a small business and what to check before adopting an AI assistant.

In this article
You put an AI chatbot on your website, and a month later an invoice arrives that nobody expected. Or it turns out that an assistant with access to company email answers anyone who simply asks it the right way. Early AI enthusiasm is often expensive and sometimes uncomfortable. This week brought two stories that touch both problems: artificial intelligence is getting cheaper, and security experts are pointing out where things can go wrong.
What happened this week
A cheaper model for repetitive work
On Monday, 9 November 2026, Meta released a model called Muse Ember. According to TechCrunch, it is the company's lightest and most cost-efficient model so far, built for the work agents repeat all day: tool calls, extracting information from documents, routing requests and reading long texts. Input costs 10 cents per million tokens, which the article says is significantly less than comparable small models from competitors. Ember also handles images and audio, so an agent can read screenshots, charts or scanned documents.
The way such models are used is interesting too. A larger model sets the rules and a lightweight one applies them across thousands of cases. Meta sums it up this way: the larger model plans, the lighter one handles the volume.
Cheap does not mean risk-free
The Czech outlet Lupa.cz reported from a conference in Prague. Spanish security researcher Chema Alonso showed that a company chatbot on a website can be pushed, with a plain but cleverly worded question, to do something other than what it was set up for. The company may then pay for computing power that strangers use. If the assistant also has access to documents, email or databases, it could in the worse case become a route to sensitive data. The author adds that this does not work with every chatbot. It depends on the permissions the operator granted and on how its activity is monitored.
What this means for a small business
For a company of a few people, three simple conclusions follow.
- The price per task is falling. Reading receipts, sorting mail or summarising a long document will get cheaper. You do not need to build your own agent. What matters is whether AI helps with work you do every day.
- Permissions matter more than the model. The biggest risk is not that the AI does not know something, but that it is allowed to do too much. Before adopting an assistant, ask what it can see, what it can write and what it can send.
- A person has the last word. An AI suggestion is a draft. Saving, sending and paying should stay with people.
Questions worth asking a vendor
- Which records does the assistant read, and who allowed it?
- Can it save, send or delete anything on its own?
- Where do my questions and data go, and how long are they kept?
- What does it do when it is unsure?
Where Andromia helps
Andromia is not a model or a developer tool. It is one system for clients, invoices, tasks, email and chat, with AI built in where it saves time. According to andromia.com:
- The AI assistant answers questions about your data, for example which invoices are overdue, and adds links to the specific records. It summarises mail and chat and suggests a reply. You say what you need, and the form opens with pre-filled fields.
- Scanned documents: you photograph a receipt with your phone or upload a PDF, and the program reads the seller, date, amounts, VAT and items. You see the data next to the photo, check it and confirm.
Both features are included in the Team and Complete plans. See the pricing page for the exact scope and prices.
How permissions work
For the same questions listed above, the website says this:
- The assistant does not save or send anything on its own. It only pre-fills a form and only suggests an email or message. Only a person can save or send.
- It reads only records that you have access to. The owner and people the owner allows can use it.
- The questions and data the assistant reads are processed by the Claude language model from Anthropic. An employee's personal ID number and identity card number never reach it.
- Conversations are kept for 30 days after the last message, and you can delete them.
- The assistant can make mistakes, so check important data.

This does not rule out every risk, and nobody can promise that AI never fails. The principle is sound, though: AI prepares, a person decides. That is exactly what you should demand from any assistant you let near company data.
What a day looks like when this is solved
In the morning you ask the assistant what is overdue and what is waiting for you. You get a list with links instead of a long search through menus. You have your mail summarised and edit the suggested reply to an important message. You photograph receipts from the road straight away, so they do not sit in your wallet until the end of the month. At the end of the day you know what was saved, because you saved it.
If you would like to see how it works with your own data, try the AI assistant in Andromia. The trial lasts 14 days.
Sources: TechCrunch, Lupa.cz.


