Key Takeaways
- A seat without connectors gives Claude only what people paste in. Your team becomes the integration layer, one copy and paste at a time.
- The gap shows up in the numbers. In McKinsey's 2026 survey, 80% of respondents say AI made them more productive, but only 37% see any EBIT impact.
- Connect the systems where decisions get made first. CRM, ERP, the data room and the shared file store carry most of the context that turns a good answer into a useful one.
- Permissions come before plumbing. Claude connectors inherit each user's own access, so a messy permission model becomes a messy Claude rollout.
- The standard is settled; your wiring is the work. MCP is now widely adopted. Picking which systems to connect, in what order, with what access, is where the rollout succeeds or stalls.
Count the copy-and-paste steps in a normal Claude conversation at a company that just rolled it out. Someone exports a pipeline report from the CRM. Someone else pulls three PDFs out of the data room. A third person pastes last quarter's numbers from the ERP into a message and asks for a summary. Claude does a fine job with all of it. Then the person copies the answer back into an email, a deck, or a ticket.
Every one of those steps is a human doing the job of an integration.
That pattern shows up in the data, too. McKinsey's 2026 State of AI survey found that 80% of respondents say AI has improved their individual productivity. Only 37% attribute any EBIT impact to AI, a share that has barely moved since last year. The same report says the small group of high performers "fundamentally redesign workflows" around AI instead of inserting it into existing ones.
I read that gap as a wiring problem. People get faster at their own tasks. The company doesn't change, because the AI never touches the systems where the company's work actually lives.
Anthropic said this plainly when it introduced the Model Context Protocol in 2024: even the best models are limited by their isolation from data, stuck behind information silos and legacy systems. MCP was built to fix that. Connectors are how that fix reaches your team.
So here's my position. If your Claude rollout ends at SSO and seat provisioning, you bought a very good chat window. The value starts when Claude can see the CRM, the ERP, the data room and the file shares your people already work in.
Context Beats Model Choice
Most rollout debates start with the model. Which one is smartest, which one is cheapest, which one should be the default. Those are fair questions. They matter a lot less than what the model can see.
A strong model with no context gives you a well-written guess. A good model with the right deal history, the right contract, and the right customer record gives you an answer you can act on. The gap between those two outputs comes from the plumbing.
The way I think about it: the moat sits in the context layer. Models get better and cheaper every few months, and every company gets the same upgrade on the same day. Your CRM history, your deal notes, your pricing logic and your contracts belong to you alone. A Claude deployment that can't reach them is using the one input every competitor also has.
Ask what Claude can see before asking which model
Before the next model debate, list the five questions your team asks Claude most often. For each one, write down where the answer actually lives. If most of those answers live in systems Claude can't reach, a better model won't fix it.
Copy and Paste Is a Tax Nobody Budgets For
Every manual hand-off between a system and Claude costs something. Time is the obvious cost. The less obvious ones are worse.
The pasted data is stale the moment it leaves the source. The person pasting picks what to include, so Claude only sees what they thought was relevant. Nothing they paste carries its permissions with it, so sensitive data ends up in conversations it was never meant to reach. And the work never compounds, because every conversation starts from zero.
Fortune's coverage of MIT NANDA's GenAI Divide report put a sharp number on where this leads: about 95% of enterprise generative AI pilots delivered little to no measurable P&L impact. (That number comes from a preliminary research program, so I treat it as a direction, not a precise failure rate.) The researchers didn't blame model quality. They pointed to a "learning gap," where generic tools stall in enterprise use because they don't learn from or adapt to workflows.
A chat window that only knows what you pasted into it is the definition of a tool that doesn't adapt to the workflow.
Count the hand-offs
Pick one recurring task, like a weekly pipeline review or a monthly close summary. Count every time someone moves data between a system and Claude by hand. Each hand-off is a connector you haven't built yet.
Connect the Systems Where Decisions Get Made
Not every connector is worth building first. The ones that pay off fastest sit where your team makes decisions every day. In most mid-market and PE-backed companies, that's a short list.
| System | What Claude can do once connected | Where to start | What to watch |
|---|---|---|---|
| CRM | Pull account history, deal notes and pipeline status into meeting prep, follow-ups and forecasts | Read-only access to accounts and opportunities | Field-level permissions and stale records |
| ERP | Answer questions on orders, inventory, invoices and margin without a report request | Read-only access to a few well-defined tables or views | Financial data access by role, and on-prem systems that need a private route |
| Data room | Summarize, compare and cite documents during diligence or LP reporting | Read-only access scoped to one deal or fund at a time | Deal confidentiality and who can see which room |
| File stores (SharePoint, Drive) | Find and use the latest version of contracts, policies and decks | Read access to a few curated, well-organized libraries | Old files, duplicate versions and overshared folders |
| Email and calendar | Prepare for meetings with full context, draft replies, track commitments | Read access for the individual user | Personal data and any send or invite actions |
The pattern in that table is deliberate. Read first. Narrow scope first. One system done well before five done halfway.
Start read-only, then earn write access
Read access delivers most of the early value with very little risk. Write access, where Claude updates the CRM or files a record in the ERP, is where things get interesting and where mistakes cost real money. Earn it one workflow at a time, with a person approving each write until the workflow has a track record.
Permissions Come Before Plumbing
Connectors expose your permission model. Whatever shape it's in, Claude will reflect it.
Anthropic's guide to custom connectors explains the default: on Team and Enterprise plans, an admin adds a connector for the organization, then each user connects individually, so Claude can only reach the tools and data that user already has access to. The exception is a connector that uses a shared credential, like a single API key. Then everyone who uses it gets that credential's access.
That's a good design. It also means a SharePoint site that half the company can quietly open becomes a SharePoint site Claude will quietly open for half the company. Rollouts that skip this step tend to discover their oversharing problem through a Claude answer instead of an audit.
Map access to roles before launch
Before turning a connector on, write down which roles should reach which data through it. Fix the obvious oversharing in the source system first. Prefer per-user authentication over shared keys. And log what connectors do, so security can review it later. This is the same work PressW's AI Security, Compliance & Governance practice does alongside a deployment.
Connectors Turn Prompts Into Workflows
A connector on its own makes Claude a better research assistant. The bigger shift happens when you build on top of it.
Once Claude can read the CRM and the calendar, a meeting-prep skill can pull the account history, recent emails and open deals before every call without anyone asking. Once it can read the data room, a diligence workflow can check every new document against a standard list of questions. The connector supplies the context. The skill or agent supplies the process. Neither is worth much alone.
This is where workflow logic gets cheap. If you can describe a process in plain language and Claude can reach the systems involved, a lot of the custom software that used to sit between your team and its data starts to look optional. I think that's one of the biggest changes in how mid-market companies will run over the next few years, and it starts with boring connector work.
Pick one workflow per system
For each connector you turn on, name one workflow it should improve and one person who owns that workflow. A connector with no workflow attached tends to get used for a week and forgotten. That's the same discipline PressW brings to Custom Claude Agents & Workflow builds.
The Standard Is Settled. Your Wiring Isn't.
A year ago, picking an integration approach for AI felt like a bet. It doesn't anymore. When Anthropic donated MCP to the Linux Foundation's Agentic AI Foundation in December 2025, it reported more than 10,000 active public MCP servers, adoption by ChatGPT, Gemini, Microsoft Copilot and others, and a Claude directory of over 75 connectors.
So the protocol is the easy part now. The hard part is your environment.
Your ERP may run on-prem with no public endpoint, which means a private route instead of a directory connector. Your data room vendor may not offer an MCP server at all. Your CRM may hold ten years of duplicate accounts. Every one of those is solvable, and none of them is solved by buying more seats.
I'll also be honest about what's still open. How much autonomy to give an agent that can write to your systems, how to review what it did at scale, and how to set trust levels per workflow are questions the whole industry is still working through. Anyone who tells you these are fully solved is selling something.
Treat connectors as infrastructure
Give every connector an owner, a documented scope and a review date. Test it when Anthropic ships a new model or feature. Retire the ones nobody uses. Connectors drift the same way any integration does, and a broken one fails quietly: Claude just answers with less context.
A Seat Is Where the Rollout Starts
I'm bullish on Claude as a work tool. But the license is the cheapest and easiest part of the rollout. The value sits in the wiring between Claude and the systems your company runs on.
I don't have a universal answer for which connector to build first. It depends on where your team makes decisions and where your data is cleanest. What I'm confident about is the order of operations: permissions, then read access to the systems that matter, then workflows on top.
If your team has Claude today, try one exercise this week. List the three systems your people paste from most often. That list is your connector roadmap. We help companies plan and build that wiring through PressW's Claude Deployment, Configuration & Maintenance service. If you want to compare notes on where to start, Book a Discovery Call.
FAQ
Claude connectors link Claude to outside tools and data, like a CRM, file store or data room, so it can read and act on live information instead of only what users paste in. Most are built on the Model Context Protocol (MCP). On Team and Enterprise plans, admins add connectors for the organization and users connect individually.
The Model Context Protocol is an open standard, created by Anthropic, for connecting AI applications to tools and data sources. It matters because it gives you one consistent way to connect Claude to many systems. It is now widely adopted across AI products and is governed by the Linux Foundation's Agentic AI Foundation.
Start with the systems where your team makes daily decisions, usually the CRM, the shared file store, and for PE and finance teams the data room. Begin with read-only access and a narrow scope. Add the ERP and write access once permissions and one or two workflows are proven.
By default, a connector added to a Team or Enterprise plan only reaches data the individual user can already access, because each user authenticates separately. The risk comes from oversharing in your source systems and from shared credentials like a single API key. Clean up permissions and log connector activity before launch.
Yes, but not always through the connector directory. Remote connectors are reached from Anthropic's cloud, so a system inside your private network needs a private route or a custom MCP server built for it. Plan that work early, because ERPs often hold the data finance and operations teams care about most.
Watch how people use it. If most conversations start with someone pasting in data from another system, your rollout needs connectors. Count those hand-offs on one recurring task. Each one points to a system Claude should be able to read directly.