Skip to content
3 role hubs · 8 tasksData privacy

AI tools for businesses: choose by task and data, not by top-ten list

Pick an AI tool by the task you need done and the data it will touch, not by a ranked list. For most finance, HR and admin work, a general chatbot or the assistant in your existing Microsoft or Google account is enough, provided your organisation has approved it.

Short answer: write down the task, then the data it touches, then who signs it off. Those three answers rule most tools in or out before you read a single comparison table. They usually point to something your organisation already pays for.

Why top-ten lists mislead finance, HR and admin teams

Search for AI tools for businesses and you mostly get ranked lists of nine, fifteen or twenty-five tools. Many are written by vendors or carry affiliate links, and most are aimed at sales, marketing and software teams.

Those lists rarely say anything about the question that matters most in a finance, HR or admin team: what happens to confidential data. A ranking cannot tell whether you are about to paste in a supplier invoice or an employee’s grievance letter, and the risks are very different. The grievance letter is personal data about a named person, so UK GDPR applies.

The ICO publishes guidance on AI. It includes a detailed overview of how to apply the principles of the UK GDPR to the use of information in AI systems, and an AI and data protection risk toolkit. A ranked list has no view on any of it.

Three questions before you compare any tool

1. What task? One job in one plain sentence. “Help with finance” is too broad to choose a tool for. “Draft first-pass commentary on why costs moved this month” is specific enough.

2. What data? Sort it into three bands: public, internal (commercially sensitive but not about people), or personal. Names, salaries, absence records and anything else that identifies someone are personal data. That means UK GDPR applies and your data protection lead needs to know.

3. Who approves? Someone in your organisation has to agree both the tool and the use. That is usually IT, the data protection lead or a manager. If nobody has, the answer for now is no.

Behind all three sits the rule this site uses for every task: numbers and decisions come from the system of record (the ledger, the HR system, the bank). AI drafts words, structure and formulas. A person checks and approves. If a tool choice would break that rule, it is the wrong choice.

Four kinds of tool and what each suits

Most products fall into one of four kinds. Comparing kinds instead of brands keeps the list short.

General chatbot. Good for drafting, rewording, explaining a concept or suggesting a structure for a document. It knows nothing about your business unless you tell it. On a personal account, keep it to tasks with little or no confidential data. A business account your organisation has approved can change that, within whatever limits the approval sets.

Assistant inside Microsoft 365 or Google Workspace. It sits inside software your organisation already licenses and runs under your work account, so part of the approval conversation may already have happened. For most office teams this is the first place to look. What each plan includes changes often, so check the current version with IT.

Specialist add-on. Software built for one job, such as reconciliation or recruitment. It can do that job well, and it also means another supplier, another contract and possibly another data protection impact assessment.

Automation. Something that runs a sequence of steps without a person clicking through each one. It is useful, but an error can repeat on every run before anyone notices. That is why it needs the most careful checks and the clearest sign-off. Leave it until the first two kinds are working and approved.

A worked example: choosing a tool for monthly variance commentary (fictional figures)

Everything in this example is made up to show how the test works. None of it comes from real data.

Each month a finance team writes commentary explaining why actual spend differed from budget. In this fictional case a department budgeted £50,000 and spent £56,000: £6,000 over budget, or 12 per cent.

  • Task: draft the written explanation of the variance.
  • Data: internal management accounts. Commercially sensitive, with no personal data.
  • Who approves: the finance director signs off the commentary; IT and whoever owns data protection sign off the tool.

With those answers the shortlist is small. The assistant inside your existing Microsoft 365 or Google Workspace account is the first candidate, because the data is internal and the approval route is shortest. A general chatbot is only an option if your organisation has approved a business account for internal financial data.

The house rule then fixes how the tool is used. The £6,000 and the 12 per cent come from the ledger. You give the tool the figures and the reasons you have already gathered from budget holders and the ledger detail. It drafts sentences. Before anything goes to the director, you check every number and every claim in the draft against the source. The tool helps phrase the variance; it never decides what the variance is or why it happened.

The P&L guide covers the steps for that kind of work, and the accountants hub lists the other finance tasks.

What to ask your IT team before anyone signs up

Ask these before anyone creates an account:

  1. Is this tool already licensed or approved here, and for which kinds of data?
  2. Are we using a business account or a personal one, and is what we type used to train the provider’s models?
  3. Does this use need a data protection impact assessment (DPIA)?
  4. Who signs off the tool, and who signs off each output?

On question 2, the published answers differ by product and by account type. At the time of writing, the vendors’ own pages say:

  • Microsoft 365 Copilot: Microsoft says prompts, responses and data accessed through Microsoft Graph aren’t used to train foundation LLMs, including those used by Microsoft Copilot.
  • OpenAI: for ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu and the API, OpenAI says it does not use inputs or outputs to improve its models by default. For services for individuals, such as a personal ChatGPT account, it says it may use your content to train its models unless you turn off “Improve the model for everyone” under Settings > Data controls.
  • Google Workspace with Gemini: Google says your content is not human reviewed or used for generative AI model training outside your domain without permission.

These statements change, so check the current version of each page and ask IT to confirm what applies to your organisation’s setup. Whatever the answer, keep personal data and anything confidential out of personal accounts.

On question 3, the ICO’s DPIA guidance says you must carry one out where processing is likely to result in a high risk to individuals’ rights and freedoms. The ICO also keeps a list of high-risk processing. On that list, a DPIA is required where innovative technology is combined with another criterion, such as evaluation or scoring, or sensitive data. The ICO says innovative technology includes the novel application of existing technologies, such as AI. It also notes this guidance is under review following the Data (Use and Access) Act, so read the live version and let your data protection lead make the call.

The “30% rule” and other rules of thumb

People search for “the 30% rule for AI”. We looked for it in official and primary sources (regulators, government, standards bodies and the main AI providers) and found no official rule. The blog posts that use the phrase define it in different and sometimes opposite ways, so we will not define it here.

A rule of thumb is only as good as its source. If you cannot trace one to someone accountable, write your own as a short policy your organisation owns. It should say which tasks AI may help with, which data never goes in, who checks the output and who signs it off. A rule you wrote and can defend is worth more than a number nobody stands behind.

Where to go next: task guides by role

If you would rather start from the job than the tool, pick your role. Finance: the accountants hub and bank reconciliation. People work: the HR hub and HR policies. Running the office: the office managers hub.

Questions

Questions people ask. Every answer open.

Short answers, drawn from this post. The sources are listed above.

Something missing, or out of date? Tell the author.

Email hello@usingaias.com

This post 5

What are the best AI tools for a small business?
No single list is honest here, because the right tool depends on the task and the data it touches. Start with what your organisation already licenses, such as the assistant in your Microsoft 365 or Google Workspace account. Only look further if it can't do the job.
Which AI tool should a finance or HR team start with?
Usually the one your organisation already has and has approved. Pick one task, decide whether it touches public, internal or personal data, and get sign-off before you start. The numbers and decisions still come from your ledger or HR system, and a person checks the output.
Is it safe to put business data into ChatGPT or Copilot?
It depends on the account type, the data and whether your organisation has approved it. At the time of writing, Microsoft says Microsoft 365 Copilot prompts and responses aren't used to train foundation models. OpenAI says it does not train on ChatGPT Business or Enterprise data by default, but it may use content from personal accounts unless you opt out. Check the current version of each vendor page, ask IT, and keep personal data out of personal accounts.
What is the '30% rule for AI' and is it a real rule?
We found no official or primary-source rule. Blogs that use the phrase define it in different ways. If your team wants a rule of thumb, write it into your own AI use policy so someone accountable owns it.
Do I need my employer's approval before using an AI tool at work?
Check your organisation's policy first. If the tool will touch business data, and especially personal data, get approval from IT or your data protection lead before you use it. Some uses may also need a data protection impact assessment.