AI in business: the options, the dangers and how to get it right
TL;DR
Difficulty: Beginner
There are three ways to use AI, and they differ in one thing: where your data goes.
- Cloud AI is the most capable and easiest, but every prompt leaves your building, and staff on consumer plans have no processing agreement.
- Self hosted AI keeps data in house, at the price of hardware, upkeep and usually a smaller model.
- A hybrid router sends anything with personal data, client data or secrets to the local model and the rest to the cloud.
- The law: the GDPR applies the moment personal data goes in. The EU AI Act applies in stages, with high risk duties moved to December 2027.
- Start with a written policy, training and a simple router. Put a number on the risk with expected value.
What to read first
Most businesses start using AI without ever deciding to. I walk through the three ways to put it to work (cloud, self hosted and a hybrid of both), the real incidents that show what careless use costs, what the GDPR and the EU AI Act require with their actual dates, and how one invented Leeds bookkeeping practice could route its sensitive work so it never leaves the building.
Three ways to use AI, and one question that decides between them
Most businesses meet AI the same way. Someone signs up for a chatbot, it writes a decent email in four seconds, and within a month half the office is using it for everything. Nobody decided anything. It just happened.
That is the moment to step back, because there are really only three ways to put AI to work, and each one answers a single question differently: where does your data go when somebody types into the box?
- Cloud AI. You use a provider's model over the internet. Your prompt leaves your building, is processed on their servers and comes back as an answer.
- Self hosted AI. You run an open weight model on hardware you control. The prompt never leaves your network.
- Hybrid AI. A routing layer sends each request to one or the other, depending on what is in it.
This guide walks through all three, what goes wrong when people get careless, and what the law says in the UK and the EU.
First, a reality check from the Office for National Statistics (ONS) Business Insights and Conditions Survey. Among businesses with 10 or more employees, self reported AI use rose from around 12% in late 2023 to around 35% by June 2026. Use grows with size: 28% of businesses with up to nine employees reported using at least one AI technology, against 49% of those with 250 or more.1
The vertical axis is the share of businesses reporting at least one AI technology, the horizontal axis the size band and date. Use has roughly tripled in under three years, yet about two thirds of businesses with 10 or more staff still report none. You are not late. You are early enough to do it properly.
Throughout, I use one invented business. Northfold Bookkeeping is a 14 person bookkeeping practice in Leeds, run by Hannah, at northfold.example. It does payroll, VAT and year end accounts for about 200 small clients. Nobody real, every number made up.
Cloud AI: the most capable option with the least setup
These are the names you have heard: ChatGPT from OpenAI, Claude from Anthropic, Gemini from Google and Copilot from Microsoft. Your request is processed in the provider's cloud and the answer comes back in seconds. No hardware, nothing to install, no model to update. You pay a subscription or per token (a token is a chunk of text, roughly a short word).
They draft copy, summarise reports, analyse documents, translate, and write and explain code. The providers update them constantly, which is why I do not list what any particular model can do: claims like that age in weeks. Check the provider's current documentation for the model you actually pay for.
The strongest argument is hardware. A large model needs serious memory on a graphics card (GPU) just to load, and cloud providers spread that cost across millions of users.
Worked example: what a cloud seat really costs Northfold
Hannah wants all 14 staff on a business plan of a cloud chatbot. Assume £25 per seat per month (the example's assumption, not any provider's price list).
Here is the number of seats, the assumed price per seat per month in pounds, the months in a year, and the yearly cost in pounds.
For £4,200 a year Northfold gets the most capable models with no maintenance. The rest of this guide is about what that does not cover.
Self hosted AI: open weight models on your own hardware
The alternative is to run a model on infrastructure you control. "Open weight" means the trained model itself (its weights, the billions of numbers that make it work) is published for anyone to download and run. Some well known families, each with its own licence:
- Llama from Meta. Commercial use is allowed under Meta's licence, with a condition: a business with more than 700 million monthly active users must ask Meta for a separate licence.2
- Mistral. A French company. When it released Mistral 7B in September 2023, it reported that the 7.3 billion parameter model beat Meta's Llama 2 13B, nearly twice its size, on all the benchmarks it tested, and published it under the Apache 2.0 licence.3 Size is not everything.
- DeepSeek. The DeepSeek R1 reasoning model is published under the MIT licence.4
- Qwen from Alibaba, Gemma from Google and Falcon from the Technology Innovation Institute in Abu Dhabi round out the familiar names.
Which one is good enough for your work is not something a list can tell you. Test two or three candidates on 20 real, anonymised tasks from your own business and score the answers yourself.
The tools developers use
If you build AI into your own software, these tools come up again and again:
- Ollama downloads and runs a model with a single command, the easiest way to try local AI.
- llama.cpp runs models directly on your hardware with fine control over every setting, and vLLM serves models to many users at once.
- Hugging Face Transformers is the standard Python library for working with pretrained models.
- LangChain chains prompts and tools into multi step workflows, and LlamaIndex connects a model to your own documents and databases, the pattern called retrieval augmented generation (RAG).
Worked example: how much memory a model needs
Memory is roughly parameters times bytes per parameter. Hugging Face's documentation: models usually load at 2 bytes per parameter, and quantisation (storing each weight in fewer bits) cuts that, so an 8 billion parameter model needs about 16GB at full size, 8GB at 8 bits and 4 to 5GB at 4 bits, at some cost in accuracy.5
Hannah's developer looks at a 70 billion parameter model, quantised to 4 bits. Four bits is half a byte:
Here is the number of parameters, converts 4 bits into half a byte per parameter, and bytes is 35GB for the weights alone. The model also needs working memory for the conversation itself, so in practice it wants around 40GB of GPU memory. That is professional hardware or several cards working together.
A 7 billion parameter model at 4 bits needs billion bytes, 3.5GB, which fits on a consumer graphics card with 8GB. That is the trade: smaller models run on modest hardware, larger ones handle harder reasoning.
The privacy problem: when convenience becomes a liability
Every prompt sent to a cloud AI leaves your infrastructure. It is processed on servers you do not control and, depending on the plan and the settings, may be stored or used to train future models.
That last point depends on the contract, and it differs between consumer and business plans. Anthropic, for example, changed its consumer terms in August 2025 to let users of its Free, Pro and Max plans choose whether their chats are used to improve its models, and said the change did not apply to its commercial terms or to API use.6 The person in accounts who signed up with a personal email is on the consumer side of that line.
In May 2023 Samsung restricted the use of generative AI tools such as ChatGPT on company computers after staff uploaded sensitive code to ChatGPT, and said it was looking for ways to use generative AI safely.7
In August 2024 the Dutch data protection authority, the Autoriteit Persoonsgegevens (AP), warned employers after receiving breach reports caused by AI chatbots: an employee of a GP practice had entered patients' medical data into a chatbot, and an employee of a telecoms company had entered a file that included customer addresses.8 By December 2025 the AP said it had received dozens of such reports across 2024 and 2025, with more in 2025 than the year before.9
The case that prompted that warning was the Dutch city of Eindhoven. A check covering 30 days, from 23 September to 23 October 2025, found staff had uploaded files containing personal data to public AI websites: documents under the Youth Act, internal reports and job applicants' CVs. The city could not say how many files were affected in total, because the AI sites kept uploads for only 30 days. It blocked public AI websites, moved staff to a tool inside its own environment and asked OpenAI to delete the files.10
What careless use looks like in practice
Nobody means any harm.
Finance. Someone asks a chatbot to tidy a profit and loss statement or draft investor notes and pastes in the real figures. That transfer may break internal policy, client contracts or market rules on inside information.
Data and client confidential information. A data scientist pastes a sample of a client's dataset, with names, emails and purchase histories, into a chatbot. Personal data has just gone to a third party without a proper processing agreement, without telling the people it is about and very likely without a lawful basis under Article 6 of the GDPR.11
Developers and keys. A developer pastes code into an AI assistant for help with a bug, with API keys, database passwords or payment gateway credentials buried in it. Those secrets have been transmitted and possibly logged, and a leaked key works for anyone who holds it. Who answers for code an assistant helped write is a question of its own, and I take it apart in AI writes the code. Who is responsible for it?, and what these assistants can and cannot do for a developer in Will AI like ChatGPT ever replace developers?
I've seen this kind of carelessness first hand. In my work with clients I came across a developer who had uploaded over 14,000 customer records to GitHub, where they stayed publicly accessible until I found them. The same habit that leads to mistakes in a code repository happens with AI tools, often with even less awareness of the risk.
Worked example: what one careless paste puts at risk at Northfold
A junior bookkeeper at Northfold is reconciling a client's payroll. The spreadsheet will not balance, so she pastes 40 rows into a free chatbot she signed up for with her personal email, and asks what is wrong.
Count what just left the building:
- 40 employees' names, so 40 people's personal data.
- Their salaries and National Insurance numbers, which identify them and could enable fraud.
- The client's bank sort code and account number from the header row.
- The client's name, which Northfold's engagement letter promises to keep confidential.
Now the obligations: no processing agreement with the provider, 40 employees never told, a client contract broken, and Hannah has to decide whether to report a breach. Nobody had told the bookkeeper where the line was.
Putting a number on the risk with expected value
"It might go wrong" is hard to budget for. Expected value turns a risk into a figure you can weigh against the cost of preventing it. It is the sum of every outcome times its probability:
Here is the expected cost per year, the probability of outcome in a year, the cost of that outcome in pounds, and means add up across all outcomes. The idea is set out from scratch in Foundations: expected value.
Worked example: is a policy and a routing layer worth it for Northfold?
The figures are the example's assumptions, not measured rates. Hannah assumes that without any controls there is a 10% chance in a year of an incident like the payroll paste that becomes serious, and that a serious incident costs £30,000 (time spent, legal advice, notifying people and one lost client). There is a 90% chance of no serious incident, which costs nothing.
Here is the assumed probability of a serious incident, its assumed cost in pounds, the probability of no serious incident, its cost, and the expected cost per year in pounds.
With a written policy, training and a routing layer that catches the obvious personal data, she assumes the probability drops to 2%:
Here is the assumed probability with controls and the expected cost per year in pounds. The controls are worth up to pounds a year in expected terms.
Expected value averages over many years. If losing the one big client would sink the practice, the average does not help in the year it happens. That is an argument for the controls, not against them.
The legal reality: GDPR, the AI Act and the UK regulator
When staff enter anything that identifies a person into a cloud AI tool, that is processing of personal data, and the GDPR applies (in the UK, the UK GDPR and the Data Protection Act 2018).
Three articles do most of the work. Article 6 requires a lawful basis for any processing. Article 28 requires a contract with any processor handling personal data on your behalf, with set terms. Article 44 allows transfers of personal data to a country outside the European Economic Area only under the conditions in that chapter of the regulation.11
The transfer point is less dramatic than it was. In July 2023 the European Commission decided that the EU US Data Privacy Framework gives adequate protection for personal data sent to US organisations certified under it.12 In September 2025 the EU General Court dismissed a challenge to that decision.13 The UK's equivalent, the UK Extension to the framework, came into effect on 12 October 2023.14 A free account opened with a personal email still comes with no processing agreement at all.
The fines, worked through for a small firm
Under the EU GDPR the most serious breaches, including those of the basic principles and the transfer rules, carry fines of up to €20 million or 4% of worldwide annual turnover, whichever is higher.11 The UK GDPR sets the same tier at £17.5 million or 4%.15
Northfold's turnover is £1.2 million (an invented figure):
Here is the 4% turnover cap, Northfold's annual turnover in pounds, and the result in pounds. Because the rule says "whichever is higher", the ceiling for Northfold is the fixed £17.5 million, not £48,000. The 4% figure only becomes the larger one above a turnover of pounds. Nobody expects a bookkeeper to be fined £17.5 million, but the ceiling shows how seriously the law takes it.
Fines are also less settled than headlines suggest. In December 2024 Italy's data protection authority, the Garante, fined OpenAI €15 million over ChatGPT, citing among other things the training of the model on personal data without an adequate legal basis. A court in Rome upheld OpenAI's appeal against that decision in a ruling published on 18 March 2026.16 The law is still being tested, case by case.
The EU AI Act, with its real dates
The EU AI Act is Regulation (EU) 2024/1689. It entered into force on 1 August 2024, the twentieth day after its publication, and applies in stages: the general date is 2 August 2026, the prohibited practices and the AI literacy duty applied from 2 February 2025, and the rules for general purpose AI models from 2 August 2025.17
Processing sensitive data does not make a system high risk. High risk is defined by what a system is used for, listed in Annex III: among others biometrics, critical infrastructure, education, employment (for example tools that filter job applications or evaluate candidates), access to essential services, and law enforcement.17 A chatbot that drafts emails is not high risk because someone pastes personal data into it. That is a GDPR problem.
The dates for high risk systems have since moved. The Digital Omnibus on AI, Regulation (EU) 2026/1744, published on 24 July 2026, delays the high risk obligations to 2 December 2027 for the systems listed in Annex III and to 2 August 2028 for AI built into products already covered by EU safety law. It also softened the AI literacy duty: providers and deployers must now "take measures to support the development of AI literacy of their staff", and the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual".18 I cover what this means for software builders in the EU AI Act for SaaS builders, and the German and Austrian angle in AI compliance for the DACH market.
Does it reach a business in Leeds? It can. The Act covers providers placing AI systems on the EU market wherever they are based, and providers and deployers outside the EU where the output of the system is used in the EU.17
Worked example: which rules touch Northfold's three AI ideas
- Staff using a chatbot to draft client emails. Not high risk under the AI Act. The question is data protection: which plan, which contract, and what staff are allowed to paste. The UK GDPR applies.
- A tool that ranks applicants for a trainee bookkeeper post. This is the kind of use Annex III lists under employment, so if the AI Act applies to it, it is high risk, with obligations from 2 December 2027.18 For a UK only hire the Act does not reach it, but the UK GDPR rules on automated decisions and fairness still do.
- A chatbot on northfold.example answering questions from clients in Ireland. The output is used in the EU, so the AI Act's reach is in play, and the transparency rules in Article 50 require AI systems that talk to people to be designed so that people know they are dealing with an AI, unless that is obvious.17
Three ideas, three sets of rules. "Is AI legal for us?" never has a one word answer.
The UK: no AI Act, principles and the existing regulators
The UK has no AI Act. The government's response to its AI regulation consultation, updated 6 February 2024, set out five principles for existing regulators to apply within their own areas: safety, security and robustness, appropriate transparency and explainability, fairness, accountability and governance, and contestability and redress.19 For personal data, the regulator that matters is the data protection one. Since 30 September 2026 that is the Information Commission (still the ICO on its website), because the Data (Use and Access) Act 2025 abolished the office of the Information Commissioner.20 The ICO's guidance on AI and data protection is the place to start, and it notes that it is under review after that Act.21
The reality, in my experience, is that enforcement is inconsistent. I once reported a significant data breach to the ICO, with full details of the customer data exposed, and received a generic reply five months later asking whether the issue still existed. After I confirmed it did, I never heard from them again. The legal framework exists. Active enforcement is another matter.
That is no licence to ignore it. Enforcement is unpredictable, the damage to your reputation from a breach can be severe, and people whose data is exposed have their own right to compensation from the controller under Article 82 of the UK GDPR.22 They do not need the regulator to act first.
Self hosting: what it solves and what it costs
Running models on your own infrastructure removes most of the privacy problems above. Data never leaves your network, and nothing you type trains someone else's model. For organisations with strict duties of confidentiality, such as healthcare, finance and legal services, it may be the only option they can defend.
Setting it up properly
Self hosting has become far easier. A basic setup has five parts:
- Hardware. A server with a GPU sized to the model, worked out as in the memory example above.
- A runtime. Ollama is the friendliest. vLLM or llama.cpp give more control.
- A model. A small one (7 billion parameters) is fast and needs little hardware, a large one (70 billion) handles harder reasoning.
- An API. Ollama exposes one automatically. By default it listens only on the machine itself, 127.0.0.1 on port 11434, and its FAQ shows how to change that and put it behind a proxy such as nginx.23
- Security. The server sits on its own network segment without internet access, behind a firewall, requires authentication and accepts connections only from approved systems. Keep logs of prompts and responses with a sensible retention period, so you have an audit trail. And staff still need clear rules on what may be processed.
A minimal check that the model answers, on the server itself:
ollama pull <model-name>
curl http://localhost:11434/api/generate -d '{"model": "<model-name>", "prompt": "Summarise VAT cash accounting in two sentences."}'
The trade offs
Self hosting is not a magic answer. Whether an open weight model is as good as a cloud model for your task is something to test, not assume. Updates, patches and troubleshooting are yours, with no support ticket to open. The hardware is a real upfront cost, electricity adds up, and keeping it running needs skills your team may not have.
Worked example: Northfold's self hosted server, costed
Every figure here is the example's assumption. Hannah's developer prices a server with a GPU at £6,000, written off over three years. It draws an average of 300 watts around the clock, at an assumed £0.25 per kilowatt hour, and needs about two days of a contractor's time a year at £450 a day for updates.
Electricity first. A year has hours:
Here is the draw in kilowatts, the hours in a year, the assumed price per kilowatt hour in pounds, and the yearly electricity cost in pounds.
Then the yearly total:
Here is the server cost spread over three years, the electricity, the maintenance days, and the yearly total in pounds.
So the server costs Northfold about £3,557 a year against £4,200 for the cloud seats. The money is close. The difference is what you get: the server keeps every client's payroll inside the building, but runs a smaller model and needs someone to look after it. Which is why a hybrid is worth a look.
The hybrid approach: route by sensitivity
A hybrid setup routes each request by how sensitive it is. Confidential data stays on the self hosted model. Requests with nothing sensitive in them, where the extra capability of a cloud model is worth having, go to the cloud.
Read it from the top. Every request passes the same check. Yes keeps sensitive content on your hardware, no sends it to the cloud, and both end in the log, which lets you show later what went where.
The check can be simple or clever. Simple versions match patterns (an email address, a National Insurance number, a sort code, something shaped like an API key). Better ones add a small classification model that judges what kind of request it is. Some systems redact the personal data, send the cleaned request to the cloud and put the original values back into the answer.
A few tools make this practical. A gateway such as LiteLLM presents one API in front of many models, so your applications integrate once and the gateway decides where each call goes.24 Frameworks such as LangChain also support several model backends, so routing can live in your own code.
Worked example: a first router for Northfold
The simplest useful router is a few patterns and one rule. This sketch in Python is a starting point, not a complete filter:
import re
PATTERNS = {
"email": re.compile(r"[\w.+-]+@[\w-]+\.[\w.]+"),
"ni_number": re.compile(r"\b[A-CEGHJ-PR-TW-Z]{2}\s?\d{2}\s?\d{2}\s?\d{2}\s?[A-D]\b", re.I),
"sort_code": re.compile(r"\b\d{2}-\d{2}-\d{2}\b"),
"api_key": re.compile(r"\b(sk|pk|api|key)[_-][A-Za-z0-9_-]{16,}\b", re.I),
}
def route(prompt: str) -> str:
hits = [name for name, rx in PATTERNS.items() if rx.search(prompt)]
if hits:
return f"local (found: {', '.join(hits)})"
return "cloud"
Run four of Northfold's real requests through it:
- "Rewrite this paragraph of our newsletter so it sounds less stiff." Nothing found, so it goes to the cloud.
- The payroll paste from earlier. It contains National Insurance numbers and a sort code, so it stays local.
- A developer's question with
api_key = "sk_live_..."in the snippet. The key pattern matches, so it stays local, and the log tells Hannah a key was nearly sent out, which is worth a quiet word. - "Explain the VAT flat rate scheme for a client who runs a café." Nothing found, so it goes to the cloud.
The patterns will miss things (a client's name on its own, for example), which is why the policy and the training still matter. The router catches the obvious mistakes that people make on a busy Friday.
A hybrid in practice: GrowCentric
I've worked on exactly this kind of hybrid architecture for GrowCentric.ai, with a self hosted AI layer between the cloud AI and the user interface. When users interact with GrowCentric's guided input, their requests reach the self hosted component first.
That layer does three jobs. It processes sensitive data locally without exposing it to cloud systems. It prepares and enriches requests before they reach the more capable cloud models. And it filters the responses, so what comes back to the user has been checked.
The guided input means users are not typing free form prompts. They choose from structured options the system understands, which sharply reduces the risk of accidental exposure, because nobody is pasting raw data into a text field. The work that needs company data happens locally, and only cleaned, structured requests go on to the cloud when more capability is needed.
It is more work to build than either pure approach. For organisations with genuine privacy duties that also want the most capable models, it is the best balance I know.
Getting AI right in your organisation
About a third of UK businesses with 10 or more staff now report using AI.1 Some of your staff may be using it already, whether or not you said they may. The question is how to do it without risks you would never accept elsewhere. Cloud AI gives the most capability for the least effort. Self hosting solves the privacy problem at the price of hardware and upkeep. A hybrid routes the sensitive work home and sends the rest out.
Whichever you choose, five things hold:
- Write a clear policy on what data may and may not go into which AI tool, and tell every member of staff.
- Add technical controls that stop sensitive data reaching cloud AI, by routing, redacting or blocking.
- Train your people. The best architecture does not help if staff do not know where the line is. In the EU it is also a legal duty to take measures that support it.18
- Monitor use. Know who uses which system, for what, and whether the answers are fit to use.
- Review it regularly. This field moves in months. A policy written today needs another look before the year is out.
AI can take tedious work off your people, surface what is hiding in your data and speed up decisions. Get the architecture right from the start and it becomes an advantage. Get it wrong and your clients' data ends up on someone else's servers, or in the news.