Most businesses using AI are using it as a better search box. The ones getting real time back have done something different: they worked out which parts of the week were never worth a human brain in the first place, and handed those over deliberately.
Before you add another subscription
Every AI tool below is another monthly bill. The software underneath it doesn’t have to be. A perpetual Windows or Office licence is bought once and owned permanently — no renewal, no price rise, no access lost if you stop paying.
If you’re rebuilding your working setup around AI this year, it’s worth getting the foundation off subscription first.
Office Professional Plus 2024 Windows 11 ProThere is a particular kind of disappointment that comes with AI tools. You sign up, you ask it a few clever questions, it answers well, and then a fortnight later you notice your working week looks exactly the same as it did before. The tool was fine. The problem was that nothing was ever actually handed over.
Delegation to a person works because you decide in advance what someone else owns. Delegation to software works the same way. The businesses we speak to who have genuinely clawed back hours — sole traders, small offices, development teams — almost all did the boring part first: they looked honestly at where their week goes, sorted those tasks into what can be handed over and what absolutely cannot, and only then picked a tool.
This guide covers that process. What to delegate, what to review, what to keep, how the main assistants differ in practice, and what it means now that some of them can operate your computer directly rather than just talk to you about it.
Start with a delegation audit, not a tool
Spend twenty minutes writing down everything you did last week. Not a polished list — the actual list, including the fifteen minutes reformatting a document and the half hour spent finding an old email. Then sort each item into one of three tiers.
The sorting matters more than the tool you eventually choose. Get it wrong in one direction and you’ll waste money on software you never really use. Get it wrong in the other and you’ll hand something over that needed your judgement.
Tier 1 — Hand over fully
Work where the output is checked by reality, not by you
These are tasks where a mistake is obvious the moment you look at the result, and where the work is mechanical rather than considered. You skim the output, you don’t audit it.
- Summarising long documents, email threads, or meeting transcripts into the three things that actually matter
- First drafts of anything that will be rewritten anyway — internal notes, briefs, descriptions, outlines
- Reformatting and restructuring: turning notes into a table, a table into prose, a mess into a document with headings
- Background research where you’ll verify anything important before acting on it
- Repetitive text work — renaming, tidying, standardising, converting between formats
Tier 2 — Hand over, then review properly
Work that leaves your business with your name on it
AI drafts it, you read every line before it goes anywhere. This is where most of the real time saving lives, because starting from a draft is dramatically faster than starting from nothing — but the review step is not optional.
- Client-facing emails, quotes, and follow-ups
- Reports and analysis where the numbers came from somewhere you can check
- Marketing copy, listings, and anything customers will read
- Technical explanations you’ll give to someone who trusts you to be right
- Meeting notes turned into actions and owners — excellent output, frequently confident about things nobody said
Tier 3 — Keep
Work where being wrong is expensive and quiet
Not because AI can’t attempt these, but because the cost of a plausible-sounding error is high and you may not notice it for weeks.
- Final commercial judgement — what to charge, who to hire, what to commit to
- Anything involving credentials, payment details, or bank access
- Legal, tax, and compliance positions — use AI to understand the question, take advice on the answer
- Sensitive personal or HR matters, including anything covered by UK GDPR obligations
- Decisions that depend on context only you hold, about people, history, or relationships
A useful test for Tier 3: if the output were subtly wrong, how long before anyone found out? Anything where the honest answer is “possibly never” stays with a human.
The main assistants, and what they’re actually good at
The differences between the major assistants have narrowed considerably. All of them write competently, summarise well, and handle documents. Choosing between them now comes down to where your work already lives and what you need them to touch.
| Assistant | Practical strength | Fits best when | Watch for |
|---|---|---|---|
| ChatGPT (OpenAI) |
Broadest general capability and the largest ecosystem of integrations and community knowledge | You want one general-purpose assistant and don’t need it tied to a particular office suite | Feature tiers change often — check what’s included in the plan you’re on rather than what an article says |
| Claude (Anthropic) |
Long documents and sustained, careful writing; strong at working through a large body of text without losing the thread | Your week is heavy on drafting, reviewing contracts and reports, or working across long files | Smaller third-party integration ecosystem than ChatGPT |
| Gemini (Google) |
Tight integration with Google Workspace — Gmail, Docs, Drive, Meet | Your business already runs on Google Workspace | Much of the value depends on being in that ecosystem; less compelling outside it |
| Copilot (Microsoft) |
Sits inside Windows and the Microsoft 365 apps, with access to your own files and mail where enabled | You’re a Microsoft house and want AI where you already work rather than in another browser tab | The genuinely useful version is a paid add-on, and the free and paid tiers differ significantly — see our Copilot free vs paid comparison |
We’ve deliberately left pricing out. All four vendors change plan structures and per-seat costs several times a year, and any figure printed here would be wrong within months. Check current pricing on the vendor’s own site before committing, and pay particular attention to minimum seat counts on business tiers — they catch small teams out regularly.
One honest observation from watching businesses make this choice: most of the benefit comes from using one assistant consistently rather than picking the theoretically optimal one. The gap between a good tool used properly and the best tool used occasionally is not close.
What this looks like in three real setups
Sole trader
The admin tail, not the actual work
If you’re running a business alone, the skilled work isn’t what’s eating your evenings — the tail around it is. Chasing unpaid invoices, replying to enquiries that go nowhere, writing the same explanation for the tenth time, keeping records tidy enough for your accountant.
Practical starting point: build a small set of reusable prompts for your five most repeated messages — the enquiry reply, the polite payment chase, the availability response, the quote follow-up, the referral thank-you. Draft with AI, edit in your own voice, send. That’s Tier 2 work, and it’s usually where a sole trader recovers the first few hours a week.
Small office
Meetings into actions, without the note-taker
The highest-value AI task in a small office is almost always turning a meeting into a list of decisions, actions, and owners. It’s pure Tier 1 mechanical work, it’s universally disliked, and the output is checkable in thirty seconds by anyone who was in the room.
Two cautions worth building into the habit. Check attribution — summaries are confident about who said what and are not always right. And check your data position before recording anything: if a meeting involves personal data or client confidentiality, know where the recording and transcript are stored and who can reach them.
Development team
Triage and explanation, with review intact
Developers were first to real AI adoption and have the clearest view of the limits. The reliable wins are explanation and triage: understanding an unfamiliar codebase, drafting tests, summarising what a change actually does, first-pass review that flags the obvious before a human looks.
The failure mode is well documented — generated code that works, looks right, and carries a security weakness nobody read carefully enough to catch. We covered this in detail in the security risks of vibe coding for businesses. AI-assisted code is Tier 2 work permanently. It does not graduate.
When AI can use your computer
Until recently, delegating to AI meant a conversation: you described a task, it produced text, you did something with that text. That boundary has moved. All the major vendors now offer some form of agent that can operate software directly — opening applications, clicking through interfaces, filling forms, moving between a browser and a document and back.
This is a genuine change rather than a marketing one. A task like “go through this folder, pull the totals out of each invoice, and put them in a spreadsheet” stops being something you describe and becomes something that happens. For repetitive multi-step work across applications, it removes the part where you were still the one doing the clicking.
It also raises the stakes considerably, because a mistake is no longer a paragraph you didn’t like. It’s an action taken on your machine.
Sensible guardrails
- Start with reversible tasks. Reading, sorting, extracting, summarising, drafting. Let it prove itself on work where the worst outcome is that you redo it.
- Never hand over payment or credential entry. No card details, no bank access, no password entry, no confirming purchases. This is a hard line, not a caution — keep those actions on your own hands and eyes.
- Watch the first few runs end to end. You’re looking for where it misreads an interface or makes an assumption you wouldn’t have. That’s cheap to learn early and expensive to learn late.
- Know what it can reach. An agent with access to your desktop can see what’s on it — client files, financial records, anything else open. Decide deliberately what’s in scope before you start, particularly where UK GDPR obligations apply.
- Keep a supervised stop. Prefer setups that ask before consequential actions rather than acting and reporting back. The pause is the whole safety mechanism.
- Keep the machine underneath it current. Agent features assume a supported, properly activated operating system with current security updates. An unsupported machine is a poor place to add software that can click things.
That last point is worth sitting with. Handing a piece of software the ability to operate your computer is precisely the moment to check that the computer itself is on supported, activated, up-to-date footing — not running an operating system past its security support date. If you’re weighing that up, our guide to whether Windows 11 is worth it covers the practical position, and if data residency is your specific concern, running LLMs securely: local vs cloud GPU looks at keeping the model itself in-house.
The part most businesses skip
Delegating to AI has the same failure mode as delegating to people: handing something over and never checking whether it worked. Put a date in the diary a month out and ask three questions. What did I actually hand over? What came back that needed redoing? Where did I say I’d delegate something and quietly carry on doing it myself?
That third question is the revealing one. It usually points either at a task that belonged in Tier 3 all along, or at a habit that hasn’t caught up with the decision. Both are worth knowing about, and neither shows up in any tool’s usage dashboard.
Frequently asked questions
For most small businesses the deciding factor is which office ecosystem you already use, not which model scores highest on benchmarks. If you run Microsoft 365 or Windows, Copilot puts AI where your work already is. If you’re on Google Workspace, Gemini has the same advantage. If you’re neither, or you want one general assistant across everything, ChatGPT and Claude are both strong general choices — Claude tends to suit document-heavy work, ChatGPT has the broader integration ecosystem.
Consistency beats optimisation. One assistant used properly every day will return far more than the theoretically best one used now and then.
It depends entirely on what you grant and to whom. Business and enterprise tiers from the major vendors generally commit to not training on your data by default, while free consumer tiers often make weaker commitments — read the terms for the specific plan you’re on rather than assuming.
If you handle personal data, UK GDPR obligations follow that data wherever it goes, including into an AI tool. Practical minimum: use a business tier rather than a personal account for business data, grant access to specific folders rather than everything, and keep client-confidential and financial material out of scope unless you’ve checked the terms properly. Where data must not leave your premises at all, a locally hosted model is the alternative worth investigating.
Copilot’s deepest integration with the Office applications is tied to Microsoft 365 subscriptions, so the fully embedded experience does require an ongoing subscription.
That doesn’t mean you need one to use AI in your work. A perpetual Office licence — bought once, owned permanently — runs Word, Excel, PowerPoint and the rest with no renewal, and you can use ChatGPT, Claude, Gemini or Copilot’s standalone version alongside it perfectly well. Draft in the assistant, work in Office. For most small businesses that combination costs considerably less over five years than subscribing to both, and you keep the software if you ever stop paying for the AI. Our comparison of Office alternatives without a subscription sets out the numbers.
Honestly, less than the marketing claims and more than the sceptics expect — and it varies enormously by how much of your week is Tier 1 and Tier 2 work. Someone whose week is heavy on drafting, summarising and admin will see a real difference. Someone whose week is mostly meetings, physical work, or judgement calls will see very little, because there’s less to hand over.
The audit at the top of this guide is the only reliable way to find out which you are. If your list comes back mostly Tier 3, that’s a genuinely useful answer — it means the money is better spent elsewhere.
Yes — the major vendors now offer agents that can operate applications directly, opening software, clicking through interfaces and moving between programs. For repetitive multi-step work spanning several applications, it’s genuinely useful.
Whether to allow it is a question of scope rather than yes or no. Start with reversible tasks, never delegate payment or credential entry, watch the first several runs closely, and be deliberate about what the agent can see on your machine. Prefer setups that pause for confirmation before consequential actions. And make sure the underlying operating system is supported, activated and patched before adding software with that level of access.
Gary Walsh is the Head of Tech Support at Software Supplies, with more than 20 years in the IT industry. Fully Microsoft-certified and experienced across the full business software stack — from Windows and Office to cloud infrastructure and device management — Gary delivers practical, no-nonsense advice that helps users and businesses get the most from their technology.
