Key facts
AI for small firms: 17 checked facts and figures, with sources
Each entry is the fact one of our guides rests on: a rule, a date, a price or a figure we worked out, with where it came from. Quote any of them; a link to the guide lets your reader check the working.
Choosing an AI assistant
We wrote this questionnaire of 5 questions and 7 outcomes from the rules set out in 4 of our guides and the sources each outcome names. Where the assistant should work decides most routes: 12 of the 20 routes through the five questions end on a plan tied to an office suite. Only the 8 routes that choose a separate app reach the two-seat minimum or the price gap between monthly and annual billing.
Source: Microsoft UK, Microsoft 365 Copilot pricing page and Business Standard plan page, read 2 October 2026; Google Workspace UK pricing page, read 2 October 2026; Google, Generative AI in Google Workspace Privacy Hub (14 August 2026); Google Workspace UK pricing page and Microsoft UK Microsoft 365 Copilot pricing page, read 2 October 2026; chatgpt.com/pricing and its UK price data; claude.com/pricing, UK version; read 2 October 2026; ICO, When is a contract needed and why is it important?; chatgpt.com/pricing and claude.com/pricing (UK version), read 2 October 2026; each outcome names its source, and the linked guide shows the working.
On 1 October 2026 five seats for a year on annual billing cost £708 before VAT on Google Workspace Business Standard with Gemini (5 × £11.80 × 12), £828 on the Microsoft 365 Copilot Business add-on at its offer price (5 × £13.80 × 12, with a Microsoft 365 plan on top), and £900 on both ChatGPT Business (5 × £15 × 12) and Claude Team (5 × £18 × 12 = £1,080 including 20% VAT). All four vendors say business-plan content is not used to train their models by default.
Source: ChatGPT pricing page and the UK price data it loads; OpenAI "Enterprise privacy" (updated 8 January 2026) and "How your data is used to improve model performance"; Claude pricing page (UK version) and Claude help centre "What is the Enterprise plan?"; Anthropic Privacy Center articles on model training (19 August 2026 and 16 March 2026) and on the data processing addendum (16 March 2026); Microsoft 365 Copilot UK pricing page and Microsoft Learn "Data, Privacy, and Security for Microsoft 365 Copilot" (18 August 2026); Google Workspace UK pricing page and "Generative AI in Google Workspace Privacy Hub" (14 August 2026); UK GDPR Article 28 (legislation.gov.uk); GOV.UK "VAT rates". All read in a browser on 2026-10-01. Desk check only: we did not buy or test any plan.
What it costs
On 1 October 2026 no vendor published a figure for a full working day of use. OpenAI gave the most numbers (a ChatGPT Business standard seat has 15 Pro-model messages a month and an estimated 5 to 2,000 messages per five-hour window in its Work and Codex tools, depending on the model); Anthropic says Claude has "no fixed message count" and gives multipliers only, with a Team standard seat at 1.25 times the Pro plan per five-hour session; Microsoft gives no number for Copilot; Google gives counts for Workspace features, such as 100 generated slides a month.
Source: OpenAI Help Center "ChatGPT Business models and limits" and "GPT-5.6 and GPT-6 Pro in ChatGPT" and the ChatGPT pricing page (Business tab); Anthropic help pages "What is the Pro plan?", "What is the Max plan?", "What is the Team plan?", "How do usage and length limits work?", "Usage limit best practices" and "Manage usage credits for Team and seat-based Enterprise plans", and the Claude pricing page; Microsoft Learn "Overview of Microsoft Copilot Chat" (updated 10 September 2026) and the Microsoft 365 Copilot UK pricing page; Google Workspace Help "About AI usage limits" (updated 1 October 2026) and Gemini Apps Help "Gemini Apps limits and upgrades for Google AI subscribers". All read in a browser on 2026-10-01. Desk check only: we did not run any plan to its limit.
GOV.UK lists "computer software if your business makes regular payments to renew the licence" among the costs a self-employed person can claim, which covers a monthly AI subscription used for the business. A £20-a-month subscription is £240 a year and saves a basic-rate sole trader £62.40 in Income Tax and Class 4 National Insurance in the 2026 to 2027 tax year, unless they use the £1,000 trading allowance, in which case no expenses can be claimed.
Source: GOV.UK "Expenses if you're self-employed" (Overview; Office, property and equipment; Subscriptions; How to claim), "Tax-free allowances on property and trading income" (updated 8 May 2019), "Corporation Tax rates, expenses and reliefs", "Income Tax rates and Personal Allowances", "Self-employed National Insurance rates", "Register for VAT", "Charge, reclaim and record VAT" and "VAT rules for supplies of digital services to consumers" (updated 28 March 2022); HMRC Business Income Manual BIM37007 and BIM35810; HMRC VAT Notice 741A, section 5 (updated 29 September 2022); ChatGPT pricing page; Claude pricing page (UK version) and Anthropic help pages on tax and VAT IDs; OpenAI help page on tax and VAT IDs. All read in a browser on 2026-10-01.
On 1 October 2026 OpenAI's pricing page listed a ChatGPT Business standard seat at £15 a month billed annually or £18 billed monthly, for teams of 2 to 200, and Plus at £20 a month. Five Business seats for a year cost £900 on annual billing (5 × £15 × 12); the page does not mention VAT, and its UK price data marks Business prices as before tax, which would make that £1,080 with 20% VAT.
Source: ChatGPT pricing page, Individual and Business and Enterprise tabs, and the UK price data the page loads; OpenAI "Enterprise privacy" (updated 8 January 2026) and help article "How your data is used to improve model performance"; Claude pricing page for the one comparison; GOV.UK "VAT rates". All read in a browser on 2026-10-01. Desk check only: we did not buy a plan or reach a checkout.
On 1 October 2026 the UK version of Anthropic's pricing page listed a Claude Team standard seat at £18 a month billed annually or £22.50 billed monthly, including 20% VAT, for teams of 2 to 150. Five people for a year cost £1,080 on annual billing (5 × £18 × 12), which is £900 before VAT, and ten cost £2,160; one person on Pro pays £180 for a year up front or £18 a month.
Source: Claude pricing page, UK version and first-load US dollar version; Claude help centre "What is the Team plan?", "How is my Team plan bill calculated?" (6 August 2026), "What is the Pro plan?", "What is the Max plan?" and "What is the Enterprise plan?"; Anthropic Privacy Center articles on model training for commercial products (19 August 2026) and consumer products (16 March 2026) and on the data processing addendum (16 March 2026); ChatGPT pricing page for the one comparison. All read in a browser on 2026-10-01. Desk check only: we did not buy a plan.
On 1 October 2026 Microsoft's UK pages priced Microsoft 365 Business Standard with Copilot at £18.10 a user a month paid yearly, before VAT, which is £1,086 a year for five people; buying Business Standard (£10.80) and the Copilot Business add-on separately came to £1,476 in the first year at the £13.80 offer price and £1,614 at the £16.10 standard price.
Source: Microsoft UK pages "Microsoft 365 Copilot plans for business", "Compare all Microsoft 365 business plans" and the Business Basic, Business Standard and Business Premium plan pages; Microsoft Learn "License options for Microsoft Copilot" (updated 25 September 2026), "Data, Privacy, and Security for Microsoft Copilot" (updated 18 August 2026), "Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat" (updated 18 August 2026), "Overview of Microsoft Copilot Chat" (updated 10 September 2026) and "Anthropic models in Microsoft Online Services" (updated 18 September 2026). All read in a browser on 2026-10-01. Desk check only: we did not buy or test any plan.
Data, privacy and the rules
We wrote this questionnaire of 5 questions and 8 outcomes from the rules set out in 3 of our guides and the sources each outcome names. Only one of the eight outcomes says the customer must be told that AI was involved: a significant decision taken with no meaningful human involvement. One says the content cannot lawfully be posted at all. The other six require no label on the work itself.
Source: Digital Markets, Competition and Consumers Act 2024, Schedule 20, paragraph 13, legislation.gov.uk, read 2 October 2026; UK GDPR Articles 22A and 22C, legislation.gov.uk; ICO draft guidance, Automated decision-making, including profiling; read 2 October 2026; UK GDPR Article 22A, legislation.gov.uk; ICO draft guidance, Automated decision-making, including profiling; read 2 October 2026; UK GDPR Article 22A, legislation.gov.uk; ICO draft guidance, What does the UK GDPR say about ADM?; read 2 October 2026; ASA and CAP, Disclosure of AI in Advertising (29 May 2025) and AI and Deepfakes (11 June 2026), read 2 October 2026; ASA and CAP, Disclosure of AI in Advertising: Striking the Balance Between Creativity and Responsibility (29 May 2025), read 2 October 2026; ICO, What privacy information should we provide?, read 2 October 2026; Digital Markets, Competition and Consumers Act 2024, section 226, legislation.gov.uk, read 2 October 2026; ICO and ASA pages named in the linked guide; each outcome names its source, and the linked guide shows the working.
None of the ICO pages we read on 1 October 2026 requires a document called an AI policy, but ten lines can state what the law does require of a firm using AI: a written contract with the supplier, a person meaningfully involved in any significant decision about someone (Articles 22A to 22D of UK GDPR, fully in force since 5 February 2026), no fake reviews (banned since 6 April 2025) and a notifiable data breach reported to the ICO within 72 hours.
Source: ICO, "When is a contract needed and why is it important?", "When do we need to do a DPIA?", "Guidance on AI and data protection" (updated 15 March 2023), "Data minimisation", "What privacy information should we provide?", "Personal data breaches: a guide", "How to deal with data protection complaints" and the draft guidance "Automated decision-making, including profiling" (updated 31 March 2026); UK GDPR Articles 22A and 22C and the Digital Markets, Competition and Consumers Act 2024, sections 225 and 226 and Schedule 20, on legislation.gov.uk; ASA, "AI and Deepfakes: Four Things Advertisers Need to Know Before They Hit Run" (11 June 2026); OpenAI "Enterprise privacy" (8 January 2026); Anthropic privacy centre (19 August 2026, 16 March 2026); all read in a browser on 2026-10-01. Not legal advice.
We wrote this questionnaire of 5 questions and 7 outcomes from the rules set out in 4 of our guides and the sources each outcome names. Three of the seven outcomes stop before any paperwork question is asked: no personal data in the prompt, a personal account, and not knowing which plan the account is on. Only a business plan reaches the questions about the vendor contract and the privacy notice.
Source: ICO, Guidance on AI and data protection, and Data minimisation, read 1 October 2026; ICO, When is a contract needed and why is it important?; OpenAI and Anthropic data pages; read 1 October 2026; OpenAI, Enterprise privacy (8 January 2026); Anthropic privacy centre (19 August 2026); read 1 October 2026; ICO, When do we need to do a DPIA?, and Guidance on AI and data protection, read 1 October 2026; ICO, When is a contract needed and why is it important?, and What needs to be included in the contract?, read 1 October 2026; ICO, What privacy information should we provide?, read 1 October 2026; ICO, A guide to lawful basis (2 April 2026), and A brief guide to international transfers (15 January 2026), read 1 October 2026; each outcome names its source, and the linked guide shows the working.
The ASA said on 29 May 2025 that there is "no blanket legal requirement in the UK to disclose the use of AI in ads". The firm duty to tell people comes from UK GDPR: since 5 February 2026, Articles 22A to 22D require a firm that takes a significant decision about someone with no meaningful human involvement to give them information about it and let them ask for a person to look at it again.
Source: ASA and CAP, "Disclosure of AI in Advertising: Striking the Balance Between Creativity and Responsibility" (29 May 2025) and "AI and Deepfakes: Four Things Advertisers Need to Know Before They Hit Run" (11 June 2026); ICO, draft guidance "Automated decision-making, including profiling" (updated 31 March 2026), "What privacy information should we provide?" and "How do we ensure transparency in AI?" (15 March 2023); UK GDPR Articles 22A and 22C, and the Digital Markets, Competition and Consumers Act 2024, sections 225 to 227 and Schedule 20 paragraph 13, on legislation.gov.uk; European Commission, "AI Act" policy page; all read in a browser on 2026-10-01. Not legal advice.
The ICO's guidance, read on 1 October 2026, says that every time a controller uses a processor there must be a written contract. OpenAI lists a data processing addendum for ChatGPT Business, Enterprise and the API, and Anthropic builds one into its commercial terms; neither page lists one for a personal plan, so customer data typed into a member of staff's own account sits outside any contract the firm holds.
Source: ICO, "A guide to lawful basis" (updated 2 April 2026), "When is a contract needed and why is it important?", "What needs to be included in the contract?", "A brief guide to international transfers" (15 January 2026), "What privacy information should we provide?", "When do we need to do a DPIA?", "Data minimisation" and "Guidance on AI and data protection" (updated 15 March 2023); OpenAI, "Enterprise privacy" (8 January 2026) and "How your data is used to improve model performance" (13 March 2026); Anthropic privacy centre articles on model training (19 August 2026 and 16 March 2026), the data processing addendum (16 March 2026) and server locations (15 June 2026); Microsoft Learn, "Data, Privacy, and Security for Microsoft Copilot" (18 August 2026); Google Workspace generative AI privacy hub (14 August 2026) and Gemini Apps privacy hub (24 September 2026); all read in a browser on 2026-10-01. Not legal advice.
Putting it to work
We wrote this list of 28 steps in 5 stages from the rules set out in 14 of our guides and the official pages each step names. 18 of the 28 steps come before any account is switched on. Two periods in the list are set by the ICO: 72 hours to report a notifiable breach and 30 days to acknowledge a complaint. The two-week trial and the six-month review are good practice.
Source: Our plan-fit questionnaire, from vendor pricing pages read 2 October 2026; OpenAI, Anthropic, Microsoft and Google pricing pages, read 1 October 2026; Microsoft UK, Microsoft 365 Copilot plans and Business Standard pages, read 1 October 2026; ChatGPT pricing page and its UK price data, read 1 October 2026; Anthropic, Microsoft, Google and OpenAI UK pricing pages, read 1 October 2026; OpenAI and Anthropic help pages on usage limits, read 1 October 2026; ICO, When is a contract needed and why is it important?, read 1 October 2026; OpenAI, Anthropic, Microsoft and Google data pages, read 1 October 2026; ICO, A brief guide to international transfers (15 January 2026); vendor data pages, read 1 October 2026; ICO, A guide to lawful basis (updated 2 April 2026); ICO, When do we need to do a DPIA?, and Guidance on AI and data protection; ICO, What privacy information should we provide?; ICO guidance pages and UK GDPR, read 1 October 2026; ICO, Guidance on AI and data protection, and Data minimisation; Harber v HMRC [2023] UKFTT 1007 (TC); R (Ayinde) v Haringey [2025] EWHC 1383 (Admin); UK GDPR Articles 22A to 22D; ICO draft guidance on automated decision-making (31 March 2026); Anthropic and OpenAI help pages on agents, read 1 October 2026; Our one-page AI policy guide; OpenAI and Anthropic privacy pages, read 1 October 2026; ChatGPT and Claude pricing pages, read 1 October 2026; OpenAI and Anthropic help pages on usage and credits, read 1 October 2026; Anthropic and OpenAI billing help pages, read 1 October 2026; HMRC VAT Notice 741A; GOV.UK, Expenses if you're self-employed; R (Ayinde) v Haringey [2025] EWHC 1383 (Admin); our five-step check; Our readiness check, written from the rules in seven of our guides; ICO, Personal data breaches: a guide; How to deal with data protection complaints; ICO DPIA guidance; vendor pages dated 8 January to 24 September 2026; each step names its source, and the linked guide shows the working.
Free tool: AI assistant rollout checklist for a small UK firm: plan, contract, policy, first month
We wrote ten questions from the rules set out in 7 of our guides; for 4 of the ten the answer that is in order is "no", so the check cannot be passed by ticking yes. The first asks: Is every AI account used for work on a business plan that the firm itself pays for? The ICO says there must be a written contract every time a firm uses a processor. OpenAI lists a data processing addendum for ChatGPT Business, Enterprise and the API, and Anthropic builds one into its commercial terms; neither lists one for a personal plan.
Source: ICO, When is a contract needed and why is it important?; OpenAI and Anthropic data pages; OpenAI and Anthropic policy pages on model training; The four vendors' own data pages, as set out in our comparison; ICO, What privacy information should we provide?; ICO guidance; UK GDPR Articles 22A to 22D; R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin); UK GDPR Articles 22A and 22C; ICO draft guidance (31 March 2026); ICO, When do we need to do a DPIA?, and Guidance on AI and data protection; Anthropic, Use Claude Cowork safely; OpenAI help pages; Microsoft Copilot Terms of Use; each rule is set out with its source in the guide the question links to. The test record is in tests/readiness/tools-check.txt.
Free tool: Is your firm's use of AI assistants in order? Ten yes-or-no questions for a small UK firm
In Harber v HMRC [2023] UKFTT 1007 (TC), decided on 4 December 2023, all nine tribunal decisions a taxpayer relied on had been invented by an AI tool; in Ayinde [2025] EWHC 1383 (Admin), decided on 6 June 2025, 18 of 45 citations in one case and five in the other did not exist, and the High Court said the critical safeguard is to check any output against an authoritative source.
Source: Harber v HMRC [2023] UKFTT 1007 (TC) and R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), both read in full on the National Archives' Find Case Law site on 2026-10-01; OpenAI Europe Terms of Use (updated 16 January 2026), Anthropic Consumer Terms of Service (effective 8 October 2025) and Commercial Terms of Service (effective 17 June 2025), Microsoft Copilot Terms of Use (effective 18 August 2026), Google Generative AI Additional Terms of Service (last modified 9 August 2023), each read in a browser on 2026-10-01. Not legal advice.
On 1 October 2026 the documentation of OpenAI, Anthropic, Microsoft and Google each described an AI agent as a model that plans a task and uses tools to act over several steps, where a chatbot answers one message and stops; each also tells the user to check or supervise it, and Anthropic's help page advises manual approval for anything hard to undo, such as sending messages or making purchases.
Source: OpenAI, developer guide "Agents" and help articles "ChatGPT Work and Codex", "ChatGPT agent" and "ChatGPT Workspace Agents for Enterprise and Business"; Anthropic, "Building effective agents" (19 December 2024) and help articles "Use Claude Cowork safely" and "Use Claude in Chrome safely"; Microsoft, Copilot Studio overview (last updated 25 September 2026), Microsoft 365 Copilot agents overview (last updated 30 September 2026) and Copilot Terms of Use (effective 18 August 2026); Google Cloud, "What are AI agents?" and Gemini Apps Help, "Use Gemini Spark". All read in a browser on 2026-10-01.
Across the ten admin scorecards on this site we listed 80 office jobs, eight per trade, and judged that an AI assistant can prepare 60 of them for a person to approve while 20 stay with a person; of those 20, 12 are a decision, advice or a sign-off and 8 need someone on the phone, in the room or doing the work by hand. Counted on 1 October 2026 from the job lists behind the scorecards; it is our judgement, not a survey.
Source: Our own count, made on 2026-10-01, of the job lists behind the ten admin scorecards published on this site (eight jobs per trade, each marked as prepared for approval or kept by a person, with example weekly hours). The method is set out in the guide and can be repeated from the scorecard pages.
AEO Lookup editorial team
We write practical guides for people who run small UK firms, and we check every fact against its source on the date shown. Read how we review, or tell us if something has changed.