What does AI really cost your business, and where does your company data go?

AI tools can save your team real time, but the licence is the smallest part of what AI costs a business. The larger costs come from the company information staff paste into AI tools you don't control. A supplier's promise not to train on your data doesn't stop that data being stored, read by people or handed over in a lawsuit. The safe approach is to pay for business AI, set clear rules, tidy up who can see what, and check the contract for retention and human review, not just training.

The short version

• Business AI licences cost money per user, per month. Microsoft 365 Copilot Business lists at £16.10 per user per month on an annual commitment, ex VAT, on top of your Microsoft 365 plan (checked 3 November 2026; Microsoft is offering £13.80 for the first year on annual commitments made by 31 December 2026) [1].
• The hidden costs are data clean-up, permissions, training and governance—and a breach. IBM found breaches cost an average of US$670,000 more at organisations with high levels of unapproved "shadow AI" [2].
• Staff already paste company information into AI tools, mostly through personal accounts [3]. Without a clear AI policy and a secure, approved AI tool, that data is likely going into AI services you don't control.
• "We don't use your data for training" is not the same as "we don't keep your data". AI providers can still store your conversations, have people review them and be ordered by a court to keep them [4][5][6].
• Microsoft 365 Copilot only shows a user what that user can already open [7]—so if your file permissions are too open, Copilot makes overshared files easy to find.

What does AI cost a business?

There are four kinds of cost to plan for. The first is on the price list; the others are the ones that catch businesses out.

Is your company data already in AI tools?

Probably. When AI tools are free and quick, staff use them to get work done whether or not the business has approved them. In Microsoft and LinkedIn's 2024 survey, 78% of people using AI at work brought their own tools [9].

What they put in is the problem. LayerX's 2025 browser data found that 77% of employees using AI had pasted data into chatbot queries, 82% of those pastes came from unmanaged personal accounts, and 22% included personal or payment card data [3]. IBM's 2025 research found that one in five organisations had a breach due to shadow AI, and only 37% had policies to manage AI or detect it [2].

If your business hasn't written an AI policy, hasn't talked to staff about AI and hasn't given them a secure AI tool to use, assume company information is already going into AI tools that aren't secure. Staff aren't being careless—they're trying to work faster with the tools they can find, usually free consumer apps on personal accounts. Without a clear rule and a safe, approved alternative, nothing stops client details, pricing, contracts or internal plans ending up in services your business has no agreement with and no control over.

A well-known example: in 2023 Samsung engineers pasted internal source code into ChatGPT while asking it for help. Samsung responded by banning generative AI tools on company devices while it put proper controls in place [10].

Why isn't "we don't train on your data" enough?

Training is only one of the things that can happen to your data after you send it. A promise about training says nothing about whether your data is kept, who can read it, or who else can be given it. Four gaps matter:

1. 1.Retention. Providers keep data for their own reasons. OpenAI's API platform for businesses, for example, keeps abuse-monitoring logs for up to 30 days by default, even though that data is not used for training [4].
2. 2.Human review. On consumer AI apps, people may read your conversations. Google's Gemini Apps privacy notice says human reviewers, including trained reviewers from its service providers, read some chats, and that reviewed chats are kept for up to three years even if you delete your activity [5].
3. 3.Legal orders. In 2025 a US court ordered OpenAI to preserve ChatGPT conversations that would otherwise have been deleted, as evidence in a copyright case [6]. The broad order was lifted for new chats from 26 September 2025, but conversations already preserved stayed available to the case [11].
4. 4.The law still applies. If you put personal data into an AI tool, UK GDPR applies: the ICO's guidance covers using AI lawfully and transparently, and is under review following the Data (Use and Access) Act [12].

None of these depend on whether the provider trains its models on your data. So the question to ask a supplier isn't just "do you train on our data?" but "what do you keep, for how long, who can read it, and where is it held?"

Is Microsoft 365 Copilot safer?

For most Microsoft 365 businesses, Copilot is the safer place to start. Microsoft states that prompts, responses and the data Copilot reads are not used to train its foundation models, that Copilot has opted out of the human-review abuse monitoring available in Azure OpenAI, and that interactions are stored under the same contractual commitments as your other Microsoft 365 content [7]. One detail for UK businesses: Microsoft notes that customers outside the EU may have Copilot queries processed in the US, the EU or other regions [7].

The catch is your own permissions. Copilot can surface anything a user is already allowed to open [7], and Microsoft notes that generative AI amplifies oversharing [8]. If SharePoint sites, Teams and shared links in Microsoft 365 are too open, find the overshared content, remove "anyone" links, restrict access and label sensitive files before and during a roll-out [8].

Can you use AI without your data leaving your control?

Yes. The safest AI is the one where you decide where your data lives and who can read it. There are three ways to do that, depending on how sensitive your work is:

What do private AI options cost?

It depends on how many people use it and how large a model the work needs, and the figures below are examples rather than quotes. Renting a dedicated cloud GPU for a private model costs from roughly US$800 a month for one mid-range GPU running around the clock, before storage, networking and support [14]. On-demand top-end GPUs in the UK cost a little over £2 per GPU-hour [15].

Buying your own hardware spans a wide range. Smaller language models can run on a single graphics card costing roughly £240 to £320 [16], while a four-GPU server for the largest open models costs around £158,000, plus about £1,400 a month in electricity [15]. Most small businesses need neither extreme, and the running, security and support of the system usually cost more over time than the hardware.

What should a business owner do this month?

1. 1.Choose approved AI tools on business accounts, and stop personal accounts being used for work data. Free consumer plans are where human review and long retention are most likely.
2. 2.Read the contract for retention, human review, data location and legal holds—not just training. Ask whether zero-retention options exist for sensitive work [4].
3. 3.Tidy permissions before switching on Copilot or any AI that searches your files: remove "anyone" links, fix site access and label sensitive data [8].
4. 4.Write a one-page AI policy: what staff can put into AI, what they must never put in (customer personal data, credentials, financials, unreleased plans), and which tools are approved.
5. 5.Train your team and keep the basics strong. Most businesses haven't yet: of the 31% of UK businesses using or considering AI, only 24% had cyber security practices to manage its risks [13].

How does BTNI help with AI?

BTNI designs, builds and manages all three options as part of AI Automation. New to workplace AI? Start with how AI is changing business technology.