We pay for Gemini. Who still needs a Claude seat?

Start everyone on Gemini. Move people who keep asking engineers questions to Claude Tag in Slack. Give a Claude seat to the people who build things or do long, private work, inside engineering or outside it.

That approach keeps you out of two traps: seats nobody opens, and staff stuck on a tool that can’t do their job.

Someone in your company asks for a Claude licence. You already pay for Google Workspace, so the reflex answer is “you’ve got Gemini”. Ask two questions instead. What are you trying to get done? What have you tried with Gemini?

Where I work, we pay for Google Workspace and Claude Team, and since June we’ve had a third option: Claude Tag in Slack. With three options on the table, I think about AI access as three paths, and the work decides which one someone takes.

flowchart LR
    A[New starter or licence request] --> B[Gemini in Workspace]
    B -- Asks questions about shared systems --> C[Claude Tag in a scoped Slack channel]
    C -- Builds, codes or works alone every day --> D[Claude Team seat]
    B -- Private, specialist or daily build work --> D
    A -- Role with a proven need, such as engineering --> D

For many people, Gemini will be enough. Some move on to Claude Tag. Others need a seat, and for a few roles, such as engineers, you know that on their first day.

The decision matters because a seat costs £20 to £25 a month. That looks like nothing for one person. Make it the default for everyone you hire, and you’ve signed up for a bill that grows with headcount whether people open Claude or not.

I checked the prices and features below on 1 November 2026. Several are in beta, so check them again before you write a policy around them.

Stop one: Gemini

If you run Google Workspace, your staff may have Gemini already, depending on your edition and what your admin has switched on. The Gemini app is a core service in most editions. For drafting an email, summarising a doc in Drive, pulling numbers out of a sheet or prepping a meeting from the calendar invite, Gemini works where the work already sits. Paying for a second chatbot for that work can waste money.

Gemini also keeps the work inside an account you govern. Google says content from users with a Workspace Gemini licence isn’t reviewed by humans or used to train generative AI models.

Refuse approved AI tools, and you risk pushing staff towards personal accounts you can’t govern. Picture someone in finance who pastes a confidential board report into their personal ChatGPT account to get a summary. OpenAI ships Improve the model for everyone switched on for Free, Plus and Pro users, so that conversation can train its models unless they’ve found the setting. Even with it off, your company can’t set retention, can’t see which connectors the account uses, and has nothing to revoke when that person leaves.

So start there. Show people the Gemini features they haven’t found, and give them a week with it.

Why people say Gemini isn’t as good

Ask someone who uses both why they prefer Claude, and they’ll describe the product more than the model. Gemini’s answers do the job for most everyday tasks. The friction shows up around the answers.

• Google spreads its AI across products. Gemini sits in the Workspace side panels and the Gemini app. App building happens in Canvas or in Google AI Studio, which sits outside your Workspace licence. Google’s coding agent, Antigravity, and Gemini Enterprise each come with their own subscriptions and access controls. Each is another product to learn, licence and review.
• Work accounts reach the tools Google lists. Since September 2026, Gemini in Docs, Sheets, Slides and Chat connects to seven services through MCP, including Atlassian Rovo, HubSpot and Salesforce. The Customize hub adds more than 40, most of them US- and English-only for now. You can’t add your own. Google’s custom apps feature “isn’t available if you sign in with a work or school Google Account”, and Google sells custom MCP in Gemini Enterprise, a separate product.
• The limits move. Google’s usage limits page gives Business Starter “basic access” to the Pro and Thinking models, where “daily limits may change frequently”. Standard and Plus get 25 Pro prompts every four hours and 300 Thinking prompts a day. Someone who leans on Gemini all day may hit those caps.
• Sharing goes through Drive. On a work account, a Gemini canvas leaves the chat as a snapshot in Drive, and only if your admin turns that on.

Claude puts those pieces in one product. A Team seat covers chat, artifacts your colleagues can open and use, Claude Design, Cowork and Claude Code, and an Owner can add any remote MCP server as a connector. Claude Tag runs from the same organisation. In my experience, that joined-up product is why people want to keep Claude, whatever the benchmarks say this month.

Still, “Gemini feels clunky” doesn’t earn anyone a seat. Find out which of those gaps blocks their work. If none does, Gemini is the answer.

Stop two: Claude Tag

Anthropic launched Claude Tag on 23 June 2026, in beta for Team and Enterprise plans. It puts Claude into the Slack channels you choose. Anyone in the channel can tag @Claude with a question or task, and Claude replies in the thread.

Two details change the licensing maths:

• No seat needed in channels. The docs say Slack users don’t each need a Claude seat. Channel work bills by usage to an organisation balance that an Owner funds.
• Shared access. In a channel, Claude acts through service accounts an admin set up, including the Claude GitHub App for repositories, and leaves the asker’s own credentials alone.

The second point makes Claude Tag the next stop for anyone whose questions land on an engineer.

Let support ask the code

A customer asks your support team what might keep a payment in pending, or what triggers a refund email. Today that question lands on an engineer, who stops what they’re doing, opens the code and writes an answer.

With Claude Tag, an admin adds a few repositories to the support channel through the Claude GitHub App. The support agent asks in Slack:

@Claude when does a payment move from pending to completed, and which events fire when it does? Name the files and functions you used, explain the behaviour in your reply, and link to the source.

Claude reads the code and explains the behaviour in plain language, naming the files and functions it used. The support agent needs no GitHub account or repository permissions, because Claude reads the code through the GitHub App access granted to the channel.

sequenceDiagram
    participant S as Support agent
    participant C as Claude in #support
    participant G as Claude GitHub App
    participant R as payments repo
    S->>C: @Claude when does a payment move to completed?
    C->>G: Read code with the channel's access
    G->>R: Read authorised source files
    R-->>G: Source files
    G-->>C: Code
    C-->>S: Plain-language answer with file links

Claude can still misread the code. A repository won’t hold runtime config, a third-party service’s behaviour or a manual step someone does in a back office. The support agent can’t open the source links without their own GitHub access, but an engineer can use them to check the answer. Tell support which answers need an engineer’s confirmation before they reach a customer. Engineers will still answer questions, but the routine ones stop interrupting them.

Shared access is your responsibility

The design that makes the support example work also creates a risk. Anthropic’s agent identity docs say anyone in the channel gets the same capabilities, and what Claude can do never changes based on who asks. Grant every repository to #support, and everyone in #support can read every repository through Claude. Check the effective access too: admins can grant repositories to a single channel, a whole Slack workspace or all of Slack, and a channel inherits the broader grants.

flowchart LR
    subgraph Channel["In a channel"]
        A1[Anyone in the channel] --> B1[Claude's service accounts]
        B1 --> C1[Repos and tools the admin granted]
        C1 --> D1[Bills to the organisation]
    end
    subgraph DM["In a DM"]
        A2[One person with a connected Claude account] --> B2[That person's Claude account]
        B2 --> C2[Their own connectors]
        C2 --> D2[Bills to their seat]
    end

Personal connectors open a second path. A member can let Claude use the connectors on their own Claude account, including MCP servers they’ve connected, for a request they make in a channel. Claude asks first, runs the work with that member’s credentials, and posts the answer to the thread where everyone can read it.

Say a product manager in #support connects their own Jira account and asks Claude to summarise an escalation. Claude reads Jira as that product manager, and everyone in #support sees the summary, including people without Jira access. Claude screens results for content such as credentials and HR records before it posts, but Anthropic calls that check “a screen, not a guarantee”. On the Team plan, no admin switch turns personal connectors off.

Some defaults I’d set:

• Grant each channel the repositories and tools its questions need, and no more.
• Keep sensitive systems in private channels with a short member list.
• By default, anyone in the paired Slack workspace can tag Claude. An Owner can restrict that to members of your Claude organisation.
• Tell staff to choose Allow with review when Claude asks to use their connectors on anything sensitive, or to ask in a DM.

Decide, per channel, whether Claude acts with shared access or with one person’s access. That choice matters more than which model you prefer.

Stop three: a Claude seat

A seat makes sense when the work is individual, frequent or private. That covers people who work alone on long tasks, handle material a channel shouldn’t see, use their own connectors all day, or build things.

Engineers make the obvious case. Ours each have a Standard Claude Team seat, which Anthropic lists at $20 a month on annual billing or $25 monthly before tax, and which has included Claude Code since January 2026. In the UK, we pay £20 per seat per month on annual billing, or £25 on monthly billing. We run Claude Code in our repositories and share approved skills and plugins across the organisation, including our security checks. For engineers, the seat pays for itself, so they skip the first two stops.

Teams miss the case outside engineering. Product managers, designers and ops leads now build working software themselves, which I wrote about in When does a vibe-coded app become a real product?. On a Team seat, an artifact can call Claude, store data between visits and use your connectors, including custom MCP servers. Claude Design builds prototypes on your design system and hands the chosen one to an engineer through Claude Code. Cowork runs multi-step work on files from the desktop app, and the Dashboards and Motion betas turn warehouse data into live dashboards and short animated explainers.

Picture an ops manager who tracks supplier renewals in a spreadsheet. They ask Claude for a tracker that flags contracts due in the next 60 days and drafts the renewal email. By lunchtime, colleagues are testing it, and nobody waited for an engineer. Someone who builds like that every week has as strong a case for a seat as an engineer. Before the team depends on the tracker, someone has to own it, know what data it stores, and decide when it needs an engineering or security review.

Count the builders, not everyone who looks at their work. Colleagues sign in to your Claude organisation to open what you share. Guests from outside need a Claude account, and the parts of an artifact that use Claude or your connectors don’t work for them. Check the share options on each artifact before you promise anyone access.

Who goes where

A few starting points:

Treat the table as a first guess. People move between paths: a product owner who starts prototyping will outgrow Claude Tag, and a finance analyst might hit something Gemini can’t do. Review each seat after 30 to 90 days, and reclaim it when someone leaves, changes role or stops using it.

Count all three costs

A licence decision carries three costs, and finance sees the monthly fee first:

Subscription creep

Each small approval adds a recurring commitment, and new starters keep adding more. The sums below use £25 a seat a month, our monthly-billed price. On annual billing at £20, each figure drops by a fifth.

Suppose only 30 of those 100 people need a seat. The bill drops from £30,000 to £9,000 a year, leaving £21,000 for Claude Tag usage, training or other tools. Claude Tag costs money too, so treat that figure as a ceiling on the saving.

Per-use pricing can cost more

Twenty occasional users on £25 seats cost £500 a month. On Claude Tag, the same twenty might cost less, or more. Anthropic doesn’t bill short replies that need no working session, or Claude reading the channel. Once Claude starts a working session, for example to clone a repository and run tools, the work draws on your balance. Anthropic publishes no per-task cost guidance, so you’ll find out by trying it.

On Team, you buy usage credits before Claude will respond in any channel. An Owner sets a monthly spend limit, and Claude posts a warning in Slack at 75% and 90%. You can set limits per channel too, so one channel per team gives you a per-team budget.

Saying no costs money too

If a seat saves someone five hours a month, arguing over £25 wastes more of your time than the seat costs. The same goes for Claude Tag: if a support agent saves ten minutes each time they skip an engineer, and the engineer keeps their focus, both savings count. You want seats where they earn their keep, and nobody should have to fight for one.

Test the paths before you write the policy

Take 20 real questions that your support team or product owners sent to engineers last month. Ask Gemini and Claude Tag each one, using only the data each tool may see. Score every answer:

• Was it right?
• Did it link to something you can verify?
• How long until you had an answer you’d trust?
• Did an engineer have to check it?
• Did support resolve it without an engineer?
• What did it cost?

Count an answer as right only when someone checked it against the code or an engineer. Score two failures separately: answers a tool got wrong, and questions it couldn’t answer because it couldn’t reach the data. A wrong answer is a quality problem. A missing one is a product or configuration gap, and you fix it with a connector or a different tool. Then compare with the engineer’s time on the same 20 questions. You’ll end up with a number you can show finance.

Put every new starter through the same decision, fast enough to fit inside onboarding. For roles where you’ve already seen the value, make the seat the default. Nobody should have to prove what you already know.

Stop counting licences

AI doesn’t fit the old procurement loop, where someone asks, a manager approves and finance counts seats. One person needs AI for five minutes a week, another for five hours a day. One needs their own documents, another needs a shared repository.

So give everyone Gemini. Put the questions that interrupt engineers into Claude Tag. Give seats to the people who build, wherever they sit in the org chart. Get it wrong, and you pay one of two ways: seats nobody opens, or staff who can’t reach the tool that would make them faster.

When the next request arrives, skip “why do you need another licence?” Ask what they’re trying to do and which path gets them there.