Custom AI Assistant for Team: MCP-Connected Setup Guide

A 12-person recruitment agency in Manchester pays for ChatGPT Team seats every month, but the actual usage is scattered: one recruiter copy-pastes candidate summaries into it, another uses it to draft outreach, most don't use it at all because nobody wrote down what it's good for. Meanwhile the same recruiters spend 20-30 minutes a day manually cross-referencing HubSpot, a shared Google Sheet of open roles, and Slack threads to answer "who's already been pitched this candidate." That's roughly 2 hours per person per week of work a properly connected assistant would do in seconds.

The direct cost is the wasted seat licences - $25-30 per user per month for a tool that's 80% idle. The indirect cost is worse: inconsistent client communication, duplicated outreach, and senior staff answering the same "where do I find X" questions instead of the AI answering them. Multiply that across a team of 10-20 knowledge workers and it's a few thousand dollars a month in dead time, not software spend.

This guide covers what a custom AI assistant for team use actually involves once it's connected to your real tools rather than left as a generic chatbot, how the build and rollout works day by day, what it costs, and the compliance points worth checking before you commit. The system referenced throughout is Bharat AI Saathi's Team AI Assistant (MCP-Connected).

What Team AI Assistant (MCP-Connected) actually does

It's a single branded assistant - deployed inside Claude or ChatGPT Team, or as a lightweight web app if you'd rather not expose the underlying platform to staff - built around your team's actual daily tasks rather than generic chat. Three concrete things distinguish it from just giving everyone a ChatGPT seat:

MCP or API connectors to three of your tools. Commonly HubSpot, Gmail, Google Sheets, Slack or Notion, capped at three by design so the scope stays fixed and the assistant stays fast and auditable rather than becoming an ungoverned integration sprawl.

A 30-prompt library mapped to your team's top recurring tasks. Not generic "write me an email" prompts - specific ones like "summarise this week's pipeline changes from the deal sheet and flag stalled deals," built from what your team actually does daily.

Guardrails and role-based access. Who can trigger which connector, what data the assistant can and can't touch, and a data-handling policy your compliance or ops lead can actually read (a DPA is available on request).

For solo operators - a single consultant or broker rather than a team - there's a lighter $500 Custom GPT/Claude Project tier that skips the MCP connectors and training sessions but keeps the persona and prompt library.

How it works, step by step

1. Kickoff call - 30-45 minutes to confirm which three tools to connect, who the assistant serves, and what "done" looks like for your top five tasks.
2. Access provisioning - your team grants scoped API/MCP access to the chosen tools (read-only where possible, write access only where the workflow needs it).
3. Persona and system prompt build - the assistant is given a name, tone, and boundaries matching how your team actually communicates with clients.
4. Prompt library and connector wiring - the 30 prompts are built and tested against your live data, not sample data.
5. 20-prompt acceptance test - you and your team run the assistant against real scenarios before go-live, not after.
6. Training - two recorded sessions plus a written SOP so new hires can onboard without a live walkthrough.
7. Daily operation and monthly care - the assistant runs inside your existing Claude/ChatGPT Team workspace or web app; the $199/month plan covers prompt and connector updates, a usage review, and migration when underlying models change.

Trigger: Recruiter asks in Slack, "Has this candidate already been pitched to Acme Corp?" Action: Assistant queries HubSpot for the candidate's deal history, cross-checks the open roles sheet in Google Sheets for Acme's current openings, and replies: "Candidate was pitched to Acme for the Ops Manager role on 3 March, marked 'no response.' Acme now has a Senior Ops Manager opening, unposted. Want me to draft an outreach note referencing the earlier pitch?"

Before and after

What it costs and what is included

There's no lock-in - the setup fee buys a working system your team owns and operates; the monthly fee is optional upkeep, cancellable when you no longer need it. Payback logic is straightforward rather than promised: if the assistant saves each of 10 team members even 30 minutes a week on the specific tasks it's built for, that recovers the $199 monthly fee in staff time well within the first month, though actual savings depend on how much your team's work maps to the three connected tools.

Delivery timeline

Day 0: Kickoff

Scoping call, tool selection, access requests sent.

Days 1-4: Build

Persona and system prompt drafted, connectors wired, prompt library built and tested against your live data.

Days 5-6: Acceptance testing

20-prompt acceptance test run with your team; adjustments made based on real feedback.

Day 7: Go-live and training

Assistant deployed in your Claude/ChatGPT Team workspace or web app; two recorded training sessions delivered along with the SOP.

What you need to have ready

• Active ChatGPT Team or Claude Team subscription (or agreement to set one up)
• Admin access or a designated admin contact for the three tools you want connected
• A shortlist of your team's top 5-10 recurring tasks the assistant should handle
• Internal sign-off on data-handling policy and role-based access scope
• A named point of contact for the kickoff call and acceptance testing
• Any existing compliance requirements (DPA needs, data residency preferences) flagged before kickoff

Doing it yourself vs done-for-you

DIY is genuinely fine if your team is small, technically comfortable, and has someone willing to own prompt maintenance indefinitely. It tends to break down once more than a handful of people rely on it and nobody owns the upkeep.

Compliance and risk notes

• Confirm what data will flow through the connected tools and whether any of it is regulated (client financial data, health information, candidate PII) before granting write access.
• If your team operates in the UK, EU or a state with data residency rules, check where the underlying AI platform processes and stores data, not just where your own servers sit.
• A DPA is available on request - loop in whoever owns compliance at your company before kickoff rather than after.
• Role-based access limits who can trigger which connector, but it doesn't replace your own internal policy on what staff should and shouldn't paste into any AI tool.
• Platform terms (OpenAI, Anthropic, HubSpot, Google) can change; the monthly care plan covers adapting to those changes, but you remain the account holder and are bound by those terms directly.
• This guide is not legal advice - if your sector has specific regulatory requirements (financial services, insurance, healthcare), have your compliance lead review the data-handling policy before go-live.

Mistakes buyers make

• Connecting five tools instead of three. The cap exists for a reason - more connectors means more failure points and slower responses. Fix: rank your tools by actual daily use and commit to the top three.
• Skipping the acceptance test. Teams sometimes want to go live immediately. Fix: the 20-prompt test catches wording and access issues cheaply, before staff hit them mid-task.
• Not naming a real task list before kickoff. "Just make it useful" produces a generic build. Fix: bring 5-10 specific recurring tasks to the kickoff call.
• Granting write access everywhere by default. It's tempting to give the assistant full access "to be safe." Fix: start read-only where possible and add write access only where the workflow requires it.
• Treating training as optional. Skipping the recorded sessions means adoption relies on one enthusiastic staff member. Fix: use the SOP for onboarding new hires from day one.

FAQ

Can the assistant work with tools other than HubSpot, Gmail, Sheets, Slack or Notion?

Those are the most common, but the three connector slots can generally be pointed at other tools with MCP or API support - this gets confirmed on the kickoff call based on what you actually use.

What happens if we outgrow three connectors?

You can add capacity as a scoped add-on later; the base system stays fixed-price and fixed-scope so costs don't creep silently.

Do we need to already be paying for ChatGPT Team or Claude Team?

Yes, or you need to be setting one up alongside this - the assistant deploys inside that subscription (or as a standalone web app if you'd rather not expose the underlying platform).

Is our data used to train any AI models?

That depends on the underlying platform's own policies (OpenAI or Anthropic), not on us - this is worth checking directly against their current enterprise/team terms before kickoff.

What if we want to cancel the monthly care plan later?

There's no lock-in - you can stop the monthly plan and keep using the assistant as-is; you'll just lose ongoing updates and migration support.

How is this different from just training staff to use ChatGPT better?

Training helps with prompting skill, but it doesn't connect the assistant to your actual data in HubSpot, Sheets or Slack - this system does both, plus the guardrails and documentation that training alone doesn't produce.

Next step

Email support@bharataisaathi.com or message via WhatsApp with your team size, current AI seats, and the three tools you'd want connected. Review the full scope on the offer page or see how we work before the kickoff call.