Claude.ai 101 for Business Teams: The UK Setup Guide (2026)

Claude.ai is Anthropic's chat assistant, and the setup that makes it useful for a UK business team takes about an hour. Put the team on a commercial plan so your conversations stay out of model training, write your custom instructions once, switch memory on, move shared context into Projects, and default to Sonnet 5. Skills, Artifacts and connectors come after that, in that order.

By Toni Dos Santos, Co-Founder, Spicy Advisory. Published 6 August 2026. Written for UK teams rolling Claude out to people who do not write code.

What Claude.ai actually is

Claude.ai is the chat product: web, desktop app, mobile. You type, it answers, and it can read your files, search the web, write documents, build small tools and reach into apps you connect to it. That's the whole surface this guide covers.

Anthropic also ships Claude Code for engineers and Claude Cowork for agentic desktop work. Different products, different buying decisions, and we've covered where each one fits separately. If your team is in sales, marketing, finance, ops, HR or legal, claude.ai is the surface you'll live in.

1. Pick the plan (and understand what you're buying in the UK)

Anthropic runs seven tiers. Here's what matters when you're the one signing.

Plan	Price (USD)	Who it's for	Trains on your chats?
Free	$0	Trying it out. Tight usage caps.	Yes, by default
Pro	$20/month	One person, daily use. Adds Projects.	Yes, by default
Max 5x	$100/month	Individuals who keep hitting the Pro ceiling	Yes, by default
Max 20x	$200/month	All-day power use	Yes, by default
Team Standard	$25/seat/month ($20 annual)	2+ people, shared Projects, admin console	No
Team Premium	$125/seat/month ($100 annual)	Teams that also want heavy Claude Code usage	No
Enterprise	Custom	SSO, audit logs, custom retention, procurement paperwork	No

That last column is the one people skip. On 28 August 2025 Anthropic changed the consumer terms so Free, Pro and Max users choose whether their chats train Claude, with the setting defaulting to on and retention stretching to five years for anyone who leaves it there. Commercial plans work the opposite way: Team, Enterprise and API traffic is excluded from training as a contractual term in the commercial terms and the DPA, with 30-day default retention.

"Most of the shadow AI I find in UK companies isn't people using banned tools. It's twelve people expensing a personal Pro subscription and pasting client work into an account with the training toggle still on. The fix costs less than the risk."

Work example. A Bristol agency has six people on personal Pro accounts, £96 a month on expenses, six separate privacy settings nobody has audited. Moving to Team Standard on annual billing costs about the same per head, kills the training question outright, and gives the ops lead an admin console showing who's actually using it. That was the entire business case.

Two UK-specific things to budget for. Anthropic bills in US dollars with no sterling price list, so your card issuer's FX spread applies. And UK VAT at 20% sits on top for business purchases. A $25 seat is closer to £24 all in, not the £19 your spreadsheet says. Compared against ChatGPT Business seats, that lands in the same bracket, so the decision rarely comes down to price.

2. Write your custom instructions once

Custom instructions (Anthropic calls them profile preferences) are a text box in Settings that loads into every conversation you have. Free accounts get them too. Most people never open it, then spend six months telling Claude the same five things every morning.

Four things belong in there: who you are and who you serve, how you want output formatted, the house rules on tone, and what Claude should ask before guessing.

Work example. Here's the block a finance director at a UK manufacturer uses. Copy the shape, change the facts:

I'm the Finance Director at a 140-person UK manufacturing business. We report in GBP, our financial year ends 31 March, and our board pack goes out five working days before each meeting.

When you write for me: British English, no American spellings. Dates as 6 August 2026. Numbers in £ with thousands separators. Lead with the number or the recommendation, then the reasoning.

Never invent a figure. If a number isn't in what I've given you, say "not in source" and leave the gap.

If my request is ambiguous about scope, period or audience, ask me one question before you start rather than producing something I have to redo.

That last line does more work than the other three put together. It turns Claude from something that guesses into something that checks.

Styles are the related feature people confuse with this. Open the plus menu in any chat, hover Use style, and you can switch between Normal, Concise, Explanatory, Formal, or a custom Style built from a writing sample. Custom instructions are the standing brief. Styles are what you change for one conversation because you're writing a LinkedIn post instead of a board paper.

3. Turn memory on, then check what it's holding

Memory is off until you switch it on at Settings, Memory. It's available on Free, Pro and Max. Anthropic rebuilt it in July 2026 into individual entries, sorted into categories, that Claude reads and updates as you chat. You can open the list, read every entry, edit them and delete them.

The division of labour is simple. Custom instructions are what you declare. Memory is what Claude notices. You tell it you're a finance director. It works out on its own that you hate three-paragraph preambles and that "the Hastings site" means the plant, not the customer.

Work example. A recruitment consultant switches memory on in January. By March, Claude knows which four clients she's actively placing for, that she screens on notice period before salary, and that her candidate summaries go out as a table with five fixed columns. She stopped writing that brief in every chat somewhere around week three.

One governance point for UK teams. Memory is a store of information about how you work, and if your people habitually paste candidate names, patient details or client financials into chat, some of that can end up in an entry. Tell people to open Settings, Memory once a month and clear anything holding personal data. It takes two minutes and it's the kind of control an ICO DPIA will ask you to describe.

4. Context engineering, which is most of what people mean by prompting

The prompt-engineering framing has aged badly. What separates a useful answer from a mediocre one is almost never clever phrasing. It's whether Claude had the right material in front of it.

A working prompt carries four things:

1. The role and the goal. Who Claude is being, and what the output is for.
2. The raw material. The actual document, transcript, export or numbers. Paste it, upload it, or connect it.
3. The constraints. Length, audience, tone, what to exclude, what you already know.
4. The shape of the output. Table, memo, email, slide notes, with the sections named.

Weak prompt	Working prompt
"Summarise this customer call."	"You're helping me prep for a renewal conversation. Attached is the transcript of Tuesday's call with Halifax Logistics. Pull out: every objection raised and who raised it, anything they said about budget or timing, and the two things they seemed most positive about. Table, max 12 rows, their words not paraphrase. Skip the pleasantries at the start."
"Write our Q3 board update."	"Draft section 3 (Commercial) of our Q3 board pack. Attached: the CRM export and last quarter's pack. Two pages max, written for four non-executive directors who don't know our product. Lead with the number, then variance to plan, then the one thing I want them to worry about. Match the voice of last quarter's section 3."

Three habits improve results more than any prompt template. Give Claude an example of what good looks like, because a past board paper teaches it more than a paragraph of description ever will. Ask it to ask you questions before it starts, on anything with real stakes. And when the answer misses, say what was wrong with it and let it try again in the same conversation rather than starting a fresh chat.

The other side of this is cost. Long conversations carry every previous turn forward, which is why people hit their limits by mid-afternoon. Starting a new chat when the topic changes is the single biggest saving available to a heavy user, and we've written up the other seventeen.

5. Projects: where the context lives

A Project is a workspace with its own knowledge base and its own instructions. Available on Pro and above, shared across the team on Team and Enterprise. Every chat you start inside a Project sees the documents you've put there.

Two parts to set up. Project knowledge is the files: tone of voice guide, pricing sheet, last year's contracts, the template everyone works from. Project instructions is a second custom-instructions box that applies only in that room.

Work example. A London insurance broker runs one Project per major client. Inside "Client: Camden Group" sits the current policy schedule, the last three renewal reports, the claims history and a note on who at Camden reads what. Project instructions say: always check the policy schedule before answering a cover question, never quote a premium that isn't in the attached schedule, output in the firm's renewal report format. Any account handler can now open that Project and get a briefing that would have taken forty minutes of reading.

The pattern that works for teams: one Project per client, per product line, or per recurring deliverable. The pattern that fails: one giant Project called "Marketing" with ninety files in it, where Claude spends its attention on things nobody asked about.

6. Skills: your process, written down once

A Skill is a set of instructions Claude loads when it recognises the task, or when you call it by name with a slash command. Anthropic ships built-in Skills for Excel, PowerPoint, Word and PDF work, and you can add your own at Settings, Capabilities.

The distinction that helps non-technical teams: a Project holds what you know, a Skill holds how you do it. Your brand guidelines are Project knowledge. The eleven steps you follow to turn a customer interview into a case study is a Skill.

Work example. A SaaS marketing team wrote a Skill called "case study". It carries the interview question list, the five-section structure they always use, the rule that every claim needs a number the customer said out loud, the legal sign-off checklist, and two finished case studies as reference. Someone types /case-study, pastes the interview transcript, and gets a draft in their format. Before the Skill, that brief lived in one person's head and a Google Doc nobody could find.

Start with one. Pick the deliverable your team produces most often and argues about most, write the method down, and see whether the output holds up. Teams that try to write fifteen Skills in a week end up with fifteen mediocre ones. We went deep on what building 50 of them actually taught us, including where it goes wrong.

7. Artifacts: output you can actually use

When Claude produces something substantial, it opens in a panel beside the chat instead of scrolling past in the conversation. That's an Artifact. Documents, tables, code, and small working tools you can click on.

The part business teams underuse: Artifacts can be interactive. Claude will build a working calculator, a decision tree, a sortable dashboard, a form, without you touching code. You can then share it with a link.

Work example. An operations manager at a facilities firm asked for a staffing calculator: enter site size, shift pattern and cover ratio, get headcount and monthly cost. Ten minutes of back and forth, one Artifact, shared with four regional managers who now stop emailing her the same question. It replaced a spreadsheet that broke every time somebody sorted a column.

Treat the first version as a draft you argue with. "The cost line should include employer NI at 15%" is a perfectly good next message, and Claude rebuilds it in place.

8. Connectors: reaching into where the work already is

A connector is a permissioned bridge between Claude and another app. You authenticate once in Settings, and Claude can then read from it, and in some cases write to it. The directory has grown past 400 integrations, and the ones UK business teams use most are Google Drive, Gmail, Google Calendar, Microsoft 365, Notion, Canva and Atlassian.

Connectors are free on every paid plan. The Microsoft 365 one covers SharePoint, OneDrive, Outlook and Teams in a single connection, and it stays read-only until an admin enables write tools, which became possible on 7 July 2026.

Work example. A professional services firm connected Google Drive and Calendar. Their Monday routine used to be a partner opening six folders to prep for client meetings. Now it's one message: "Look at my calendar for this week, find the folder for each client I'm meeting, and give me a one-paragraph brief per meeting with anything that changed since we last spoke." Fifteen minutes back, every Monday, per partner.

On Team and Enterprise, admins control which connectors people can add. Do that before rollout rather than after. A finance team connecting Claude to a live accounting system is a conversation worth having in advance, not a discovery you make in an audit.

9. Which model, and when

The model picker sits under the chat box. Most people leave it alone forever, which is fine, and occasionally expensive.

Model	Use it for	Work example
Haiku 4.5	High volume, low stakes, repetitive	Sorting 200 inbound CVs into three buckets against fixed criteria
Sonnet 5	The default for real work	Client emails, contract review, first-draft reports, spreadsheet analysis
Opus 5	Judgement calls and long documents	Reading a 90-page tender and telling you whether to bid
Fable 5	Long-running research and the genuinely hard stuff	A market entry analysis pulling from twenty sources over several sessions

Sonnet 5 is where most business users should sit. It handles multi-step work without hand-holding and it's fast enough that you don't lose your train of thought waiting. Opus 5 arrived on 24 July 2026 and is the default on Max. Fable 5 sits above it, built for long-running agentic work.

The rule of thumb: if you'd give the task to a graduate, use Sonnet. If you'd give it to your best analyst and expect to read the answer twice, use Opus. If the task has no obvious method and you'd expect it to take someone a full day, that's where Fable earns its cost.

Your first two weeks

Week one, one person. Set up billing on Team. Check the privacy toggle on any existing accounts. Write your custom instructions. Switch memory on. Create one Project for the thing you do most and put five real documents in it. Use it daily for a week without telling anyone.

Week two, the team. Run a 90-minute session where everyone sets up their own custom instructions and builds one Project, using their real work rather than a demo. Agree what never gets pasted into Claude. Connect Drive or Microsoft 365 for the group. Pick the one deliverable worth writing a Skill for and write it together.

What kills rollouts is the gap between licences bought and licences used. The teams that get there run the second session on real work, not on a slide deck. That's the entire difference, and it's why we build Claude training around the workflows a team already has rather than a feature tour.

What to settle before you roll out (UK specifics)

Four things worth having written down before the seats go live.

The plan is your data boundary. Commercial plans are excluded from model training by contract. Consumer plans are not, unless each person opts out. If any client, candidate or patient data touches Claude, that decides the plan for you.

Retention. 30 days by default on professional plans. Enterprise can go to zero-data retention or set a custom window. Note that even under ZDR, safety classifier results persist, and Fable 5 carries a 30-day retention requirement introduced on 9 June 2026.

The paperwork exists. Anthropic offers a DPA and holds SOC 2 Type II, ISO 27001:2022 and ISO/IEC 42001:2023, all listed at the Anthropic Trust Center. Your DPO will ask. Have it ready.

A DPIA is likely. The ICO expects one where processing is likely to result in high risk, and a firm-wide assistant rollout in a regulated sector usually qualifies. Our UK governance framework walks through what that document needs to cover.

None of this needs to take a quarter. A one-page usage guideline, the right plan, and a named owner covers most of it. The companies that stall are the ones waiting for a policy document that never gets written.

"The hour you spend on custom instructions, one Project and the privacy settings pays for itself in the first fortnight. Almost nobody spends it, which is why most Claude rollouts look like an expensive search engine."

Pick one recurring deliverable this week. Next Wednesday, whoever owns it builds the Project, uses it for the real thing, and tells the rest of the team what broke.