Most businesses already pay for ChatGPT, Claude, or Copilot. And most of them hit the same wall within a month: the assistant is smart, but it can’t see anything that matters. It doesn’t know your customers, your orders, your stock levels, or what’s in your CRM. So your team copies data in, pastes answers out, and the “AI productivity” gain turns into a new kind of busywork.
Model Context Protocol (MCP) is the fix the whole industry has settled on. OpenAI, Anthropic, Google, and Microsoft all support it, and it’s now the default way to plug an AI assistant into real business systems. Here’s what it is, what it’s good for, where it goes wrong, and what it costs to set up properly.
What Is MCP, in Plain English?
MCP is an open standard for connecting AI assistants to outside tools and data. The common comparison is USB-C for AI: one standard plug instead of a different cable for every device.
Before MCP, if you wanted ChatGPT and Claude and Copilot to read your CRM, that meant three separate integrations. With MCP, you build or install one MCP server for your CRM, and any compatible assistant can use it.
An MCP server exposes a short, named list of actions — for example:
find_customer— look up a customer by name, email, or order numberget_open_invoices— list unpaid invoices for an accountdraft_quote— create a draft quote for a sales rep to reviewcheck_inventory— return live stock for a product
The assistant decides when to use each one based on what the user asked. Your team types “What does Acme Ltd still owe us, and when did we last talk to them?” and gets an answer from your actual systems, not a guess.
Why This Matters Now
Two things changed in 2026.
Every major AI assistant supports it. MCP started at Anthropic in late 2024, and it’s now backed by OpenAI, Google, Microsoft, and AWS, with governance under the Linux Foundation. The protocol is still being updated (the latest spec landed in July 2026), but it’s past the point of being a risky bet. Building on it means you aren’t tied to one AI vendor.
Your software vendors are shipping MCP servers. HubSpot, Slack, Notion, Google Drive, GitHub, Jira, and most major databases already have one. That covers your off-the-shelf tools. What’s left is the part that actually runs your business: your custom app, your internal database, your ERP, your spreadsheets-turned-systems. That’s where custom work comes in.
Real Business Use Cases
These are the patterns I see delivering value quickly — all narrow, specific, and tied to a system the team already uses every day.
| Team | What the AI can do via MCP | What it replaces |
|---|---|---|
| Sales | Pull account history, open deals, and last contact before a call; draft follow-ups | 10–15 minutes of CRM clicking per call |
| Support | Look up order status, shipping, and refund eligibility mid-conversation | Switching between helpdesk, store admin, and courier sites |
| Operations | Answer “what’s low on stock?” or “which orders are late?” from live data | Morning spreadsheet exports |
| Finance | List overdue invoices, summarise cash position, draft reminder emails | Manual reconciliation and chasing |
| Management | Ask plain-English questions of the database without writing SQL | Waiting on someone to build a report |
| E-commerce | Query products, orders, and customers across Shopify and a custom backend | Logging in to three admin panels |
If you’ve read the AI agents guide, this is the missing piece: an agent is only as useful as the systems it’s allowed to touch. MCP is how you give it access safely.
Do You Need a Custom MCP Server?
You probably don’t if your tools are mainstream SaaS products with an official server. Connect it, set permissions, test it with your team, done. That’s often a same-day setup.
You probably do if:
- Your core data lives in a custom web app, internal database, or legacy system with no MCP server
- You need write actions (creating records, issuing refunds, updating orders), not just lookups
- Different staff need different permissions — a sales rep shouldn’t see payroll
- You need an audit trail of everything the AI read or changed, for compliance or peace of mind
- You want to combine several systems into one set of business-level actions rather than dozens of raw API calls
That last point is underrated. A good custom server doesn’t just mirror your database; it exposes actions that match how your business talks — “get customer health summary” instead of six separate table queries. That’s what makes the assistant reliable instead of confused.
The Security Part Nobody Should Skip
MCP makes it easy to give an AI access to things. That’s the benefit and the risk. The problems I see most often:
Over-broad access. A server that exposes “run any SQL query” is fast to build and a bad idea. Each tool should do one specific thing, with the minimum permissions it needs.
Prompt injection. If the AI reads untrusted content — customer emails, support tickets, web pages — that content can contain instructions trying to hijack it. Tools that can change data need guardrails that don’t depend on the AI behaving.
Unvetted third-party servers. Installing a random MCP server from the internet is like installing a random browser extension with access to your CRM. Stick to official servers or ones you’ve reviewed.
No human in the loop. Reads can be automatic. Anything irreversible — payments, deletions, customer-facing emails — should require a person to approve it.
A production-ready setup includes proper sign-in (OAuth, tied to each user’s existing account), per-user permissions, approval steps for risky actions, rate limits, and a log of every tool call. This is the same thinking covered in the enterprise AI deployment guide, applied to one connection at a time.
What Does a Custom MCP Server Cost?
On top of the build, budget $20–$150/month for hosting the server, plus whatever you already pay for your AI assistant seats. Because MCP servers usually wrap APIs you already have, the cost is mostly design, permissions, and testing — not rebuilding your systems. If your system has no API at all, that part comes first; the API integration guide covers what that involves.
Timelines are short compared with most AI projects: a read-only server is often 1–2 weeks, and a full read/write server with approvals is typically 3–6 weeks.
How to Start Without Overbuilding
- Pick one question your team asks every day that requires opening another system to answer. “Where’s this order?” and “What’s the history with this client?” are classics.
- Start read-only. Let people use it for two weeks. You’ll learn which actions they actually want before you build anything that writes data.
- Add write actions one at a time, each with a clear approval rule.
- Measure the time saved. If the lookups aren’t getting used, the next tool won’t be either.
This is the same “one workflow first” approach from the internal AI assistant guide. It works because it’s boring: small scope, fast feedback, no big-bang rollout.
The Bottom Line
The AI assistant your team already pays for is far more useful once it can see your business. MCP is now the standard way to make that happen, and it’s supported by every major AI platform. For off-the-shelf tools, it’s often a simple setup. For your own systems, a well-scoped custom MCP server is one of the cheapest, fastest AI projects with a clear payoff — as long as permissions and approvals are designed in from day one.
Want ChatGPT, Claude, or Copilot connected to your own CRM, database, or custom app — safely? Let’s talk. I’ll tell you whether an existing server covers it or what a custom one would cost, before you commit to anything.