How to Use This Guide
"Plenty of businesses have opened the app. Far fewer have wired it into how work actually gets done."
Most writing about artificial intelligence is aimed either at engineers or at nobody in particular. This guide is aimed at the people who carry the business: the owner who answers the phone and still has to get the newsletter out on Friday, the general manager or department head who has eleven people asking whether they are allowed to use any of this yet, and the CFO or COO at a company where approving the spend, writing the policy and answering for both land on the same desk.
The argument it makes is narrow and easy to test. Most of the disappointment people report with these tools traces back to two fixable mistakes. The first is using the wrong tool for the job. The second is typing a one-line request where a six-line brief was needed. Fix both and the tools stop feeling like a novelty and start behaving like a capable assistant who happens to work at three in the morning. The rest is a question of wiring it into how work actually gets done.
of small businesses report using generative AI, up from 23% in 2023
average reported return for every $1 invested in generative AI
average weekly time saved by people who use AI at work
Those figures come from the US Chamber of Commerce, an IDC study commissioned by Microsoft, and a Protiviti and London School of Economics study respectively. Set against them, the US Census Bureau, which asks a stricter question about production use rather than any use at all, put adoption among businesses under 250 employees at 8.8% in August 2025. Both pictures are true. Plenty of businesses have opened the app; far fewer have wired it into how work actually gets done. This guide is about the gap between those two states.
1. The Five-Minute Orientation
If you have used one of these tools and come away unimpressed, it is worth knowing what changed. The versions that reached the public in 2022 and 2023 were text predictors with a chat window bolted on. What you are buying in 2026 is closer to a workspace: it can read your files, search the live web, run calculations, hold a project's context across weeks, and produce a working document rather than a wall of text.
1.1 What You Are Actually Buying
Strip away the branding and all five products are the same three things in different proportions. There is a language model that generates text. There is a set of tools it can reach for, such as web search, a code sandbox for crunching a spreadsheet, or a connection to your Google Drive or SharePoint. And there is a memory layer that decides how much of your prior context it carries into the next request.
The practical consequence is that the assistant knows nothing about your business until you tell it or connect it. It does not know your margins, your tone of voice, your best customer, or why last October was a disaster. Almost every complaint about generic output comes back to this. You are asking a capable stranger to write about a company they have never heard of.
The Core Idea
Treat the assistant as a bright new hire on their first morning. Talented, fast, entirely without context. Anthropic uses this exact framing in its own guidance, and it is the single most useful mental model in this guide.
1.2 Context Windows, in Plain Terms
A context window is how much material the assistant can hold in view at once: your instructions, the documents you upload, and the conversation so far. This is the number that decides whether you can drop in a sixty-page supplier contract and ask sensible questions about clause 14. The larger models now hold up to a million tokens, which is roughly 750,000 words, or a stack of documents you would need a weekend to read.
Two things follow. First, uploading the real document nearly always beats describing it. Second, very long conversations degrade. When a thread starts drifting or contradicting itself, start a fresh one and paste in a short summary of where you had got to.
1.3 What It Is Genuinely Good At, and Where It Still Fails
| Reliably good | Still unreliable |
|---|---|
| First drafts of almost anything, from a job posting to a client email | Arithmetic and data accuracy unless it is given a calculator or a file to work from |
| Summarising long documents, transcripts and email threads | Anything requiring guaranteed correctness with no human check |
| Reformatting and restructuring material you already have | Current facts, unless web search is switched on |
| Explaining an unfamiliar concept at whatever level you ask for | Knowing what it does not know, which is why it invents confidently |
| Extracting structure from messy notes, forms and free text | Genuine originality; it produces the average of what has been written before |
| Adapting tone for a different audience | Judgement calls that depend on context it has not been given |
Read that table as a division of labour rather than a scorecard. The left column gets you to a competent eighty per cent in a fraction of the time. The last twenty per cent, the part where judgement and accountability live, stays with you.
2. Choosing Your Stack
You do not need five subscriptions. Most businesses under a hundred people are well served by one paid workhorse for daily writing and analysis, plus one research tool that cites its sources. Everything else is a preference.
2.1 The Short Answer
If your business already runs on Microsoft 365, start with Microsoft Copilot; if it runs on Google Workspace, start with Gemini. Each has the same advantage: it works inside the documents and mailboxes you already have, on your own company data. If you are on neither, pick ChatGPT or Claude on merit: ChatGPT if you want the widest range of capabilities in one place, Claude if the majority of your use is writing and reading long documents. Add Perplexity for research regardless of which you choose, because the ability to click through to the source changes how much you can trust an answer.
2.2 Claude, from Anthropic
Claude has a reputation among people who write for a living as the best drafter of the group, and it holds up. It is less prone to the breezy listicle voice, it follows detailed style instructions closely, and it handles very long documents without losing the thread halfway through.
| Claude | |
|---|---|
| Model tiers | Haiku for fast, high-volume work. Sonnet as the everyday default. Opus and above for hard reasoning. Version numbers move; the tier names have been stable. |
| Price | Free tier, which now runs the same mid-tier model as Pro. Pro US$20 a month, or US$200 a year. Max at US$100 or US$200 a month for heavy users. Team from US$25 per seat a month. Enterprise is quoted. |
| Standout features | Projects, which hold your files and standing instructions in one workspace. Artifacts, which render a working document, calculator or chart beside the chat. Very large context. Connectors to Drive, Gmail, Slack and Notion. |
| Best for | Proposals, client correspondence, contract and transcript review, policy documents, anything where the writing itself is the deliverable. |
| Watch out for | Consumer plans now train on your conversations unless you opt out. |
2.3 ChatGPT, from OpenAI
The generalist, and the safest single choice if you want one tool for a small team with mixed needs. It writes well, analyses spreadsheets without you writing a formula, generates images, holds a voice conversation, and now routes your question automatically to a fast or a slow model depending on how hard it is.
| ChatGPT | |
|---|---|
| Model tiers | A fast default for everyday chat, a Thinking mode for harder problems, and a research-grade tier on paid plans. In most cases the router picks for you. |
| Price | Free tier, ad-supported in the United States. Go US$8 a month. Plus US$20, and the sensible default for an individual. Pro at US$100 or US$200 a month. Business US$25 per seat a month. Enterprise is quoted. |
| Standout features | Advanced data analysis, where you upload a spreadsheet and ask questions in English. Deep Research, which reads dozens of sources and returns a cited report. Projects, Canvas, image generation, and a wide set of connectors. |
| Best for | Mixed teams, spreadsheet work, campaign and content production, and any business that wants one subscription to cover most requests. |
| Watch out for | Personal plans train on your data by default. The Business tier does not. Put the team on Business. |
2.4 Perplexity
Perplexity is a different kind of product. Rather than owning a model, it runs your question through live web search and then has a frontier model write the answer, with numbered citations beside every claim. That one design decision makes it the only tool of the group you can audit quickly, which is exactly what you want for market sizing, competitor checks, supplier due diligence and regulatory questions.
| Perplexity | |
|---|---|
| Modes | Search for quick cited answers. Pro Search for messy questions. Research for a multi-page cited report. Labs, on higher tiers, to produce spreadsheets, dashboards and decks. |
| Price | Free tier with roughly five Pro searches a day. Pro US$20 a month or US$200 a year. Max US$200 a month. Enterprise Pro US$40 per seat a month. The Comet browser is free on every tier. |
| Best for | Any question where you need to show your working, or where being six months out of date would be expensive. |
| Watch out for | It researches, it does not operate. It will not draft in your voice as well as Claude, and it cannot act inside your CRM or helpdesk. |
Unlocked. Your download has started.
The rest of section 2 is now readable below. Sections 3 to 10, including the eight rewritten prompts and the thirty-day plan, are in the PDF. Download it again.
Unlock the Full Field Guide
Get the Microsoft Copilot and Gemini breakdowns, the five-platform comparison table, three costed starting stacks, eight prompts rewritten before-and-after, the trust boundary, and a thirty-day rollout plan. 32 pages, free.
Free for owners and operators.
What the Full Guide Covers
The remaining twenty-four pages are the operational half. Section 3 explains the escalation ladder every vendor now sells under different names, and the overthinking tax you pay for reaching past it. Section 4 breaks a working prompt into six components you can reuse. Section 5 is the sharp end: eight real requests small businesses actually type, each shown weak, then rewritten, with a note on why the rewrite works.
Sections 6 and 7 cover the habits that compound and the failure modes worth recognising, from confident invention to the quieter problem of bias in hiring and marketing copy. Section 9 is the one most guides skip: what trains on your data, what never goes into a personal account, and an eight-line policy you can adopt today. Section 10 is a thirty-day rollout order that does not leave you retrofitting governance onto an established habit.
Frequently Asked Questions
Which AI assistant is best for a small business?
It depends on where your business already works. If you run on Microsoft 365, start with Microsoft Copilot. If you run on Google Workspace, start with Gemini. Both work inside the documents and mailboxes you already have. If you are on neither, pick ChatGPT for the widest range of capabilities or Claude if most of your work is writing and reading long documents.
Do I need more than one AI subscription?
Most businesses under a hundred people are well served by two: one paid workhorse for daily writing and analysis, plus one research tool that cites its sources. The second subscription costs about twenty dollars a month and removes a whole category of risk, because you can click through to the source rather than trusting an uncited answer.
Does AI train on my business data?
On free and personal plans, your conversations are generally used to improve the models unless you turn that off. On business and enterprise plans, they are contractually excluded. That single distinction is why a paid business tier is not a luxury for any company handling client information.
Why does AI produce generic output for my business?
Because the assistant knows nothing about your business until you tell it or connect it. It does not know your margins, your tone of voice, your best customer, or why last October was a disaster. Almost every complaint about generic output traces back to this: you are asking a capable stranger to write about a company they have never heard of.
How much do AI assistants cost for a small team?
Individual plans cluster around US$20 a month across Claude Pro, ChatGPT Plus, Perplexity Pro and Gemini AI Pro. Business tiers run roughly US$21 to US$30 per seat per month and add admin controls plus the contractual guarantee that your data is not used for training. Microsoft and Google add-ons sit on top of a qualifying base plan, so the real cost is the base plan plus the add-on.