Key Takeaways

• MCP (Model Context Protocol) is an open standard that lets an AI assistant take real actions inside business software, so work that used to take a dozen clicks can be done by asking in a sentence.
• Gartner predicts AI agents will outnumber sellers 10 to 1 by 2028, yet fewer than 40% of sellers will say those agents improved their productivity, and the main reason is that most agents can only suggest, not act.
• The first revenue work MCP changes is operational: onboarding and offboarding, library cleanup, territory changes, and reporting, which is the work that eats RevOps time without moving pipeline.
• Before connecting any vendor’s MCP server, require five things: real actions, a preview before anything changes, permissions inherited from the user, hard limits on bulk changes, and a read-only mode.
• Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, and the projects that survive start with bounded, measurable tasks rather than broad autonomy.
• MCP removes the clicking, not the judgment. Deciding what to change, and why, is still the job of the person giving the instruction.

Gartner predicts that by 2028, AI agents will outnumber sellers 10 to 1, yet fewer than 40% of sellers will say those agents improved their productivity (Gartner, AI Agents Will Outnumber Sellers 10 to 1 by 2028, 2026). Read that twice. Ten times more agents than people, and most of the people say it did not help.

Gartner’s own warning names the problem: without the right data foundation and workflow integration, sales organizations risk “agent sprawl,” with more digital activity and little improvement in seller impact. That is what happens when an agent sits next to your software instead of inside it. It can draft, summarize, and recommend. Then a human still has to open six tabs and do the actual work.

MCP, the Model Context Protocol, is the piece that closes that gap. This post covers what MCP is in plain language, why the interface to sales software is shifting from screens to sentences, which revenue work changes first, what to demand from any vendor before you connect an agent to your stack, and how to run a pilot that does not end up in Gartner’s cancellation statistics.

What Is MCP, and Why Should a Revenue Leader Care?

MCP is an open standard, introduced by Anthropic in November 2024, for connecting AI assistants to the systems where work and data actually live. A software vendor publishes an MCP server that describes what its product can do. An AI assistant, the MCP client, connects to it and can then read data and take actions in that product on your behalf.

Think of it as a universal adapter. Before MCP, every AI integration was a custom build: one connector for the CRM, another for the demo platform, another for the data warehouse, each maintained separately. With a shared protocol, any compliant assistant can work with any compliant product.

For a revenue leader, the technical detail matters less than the consequence. Your sales software is getting a second kind of user. The first kind is a person clicking through screens. The second is an agent, acting for a person, working through the protocol. Every tool in your stack will eventually be judged on how well it serves both.

Why More AI Agents Have Not Made Sellers More Productive

Most AI in sales tools today is advisory. It writes the follow-up email, scores the lead, summarizes the call, and suggests the next step. All useful. None of it finishes the job.

The seller still copies the email into the sequence tool. The ops manager still opens the admin console to apply the change the assistant recommended. The SE still rebuilds the folder structure by hand after the assistant explained how it should look. The AI removed the thinking time on a task and left the doing time exactly where it was.

That is the gap behind Gartner’s productivity number. An agent that can only suggest adds a step: read the suggestion, evaluate it, then execute it yourself. Multiply that across ten agents per seller and you get the sprawl Gartner describes, more output to review and no less clicking.

The fix is not a smarter model. It is access. An agent that can act inside the system, within rules you set, turns a recommendation into a completed task. That is the difference MCP makes.

The Interface to Sales Software Is Moving From Screens to Sentences

Every generation of business software has had a primary interface. First it was the screen: menus, forms, and dashboards built for a human to navigate. Then came the API, built for developers to connect systems, powerful but out of reach for anyone who does not write code.

MCP is the third interface, and it is the first one built for an agent that takes instructions from a non-technical person. The ops manager does not need to learn the admin console or file a ticket for an engineer to script the API. They describe the outcome, and the agent works out which actions produce it.

This is why “the next user of your sales software is an agent” is not a slogan. When a revenue operations lead says “move everyone in the EMEA team to the new regional folder structure and archive anything nobody has opened since March,” the user executing that request is the agent. The person is the one with the intent and the authority. The agent is the one clicking, at machine speed, across hundreds of items.

Your buyers are making the same shift. Gartner’s Top Strategic Predictions for 2026 and Beyond forecasts that by 2028, 90% of B2B buying will be intermediated by AI agents (Gartner, 2025). Agents are arriving on both sides of the deal at once. We covered the buyer side in why your demo is your first AI agent. This post is about the seller side: the agents running your own operation.

Which Revenue Operations Work Changes First

The first work MCP changes is not selling. It is the operational layer underneath selling, the work that is high volume, rule based, and tedious enough that it slips.

Onboarding and offboarding

A new regional team starts. Someone has to invite every person, assign each one the right role, check that seats are available, and add them to the right groups. Done one screen at a time, that is an afternoon, and seat shortages surface halfway through. Through an agent, it is one instruction with the checks run up front.

Offboarding is worse, because it is where things get lost. When a rep leaves, their demos, links, and assets have owners that need to change before access is removed. Miss that step and a live buyer link breaks, or a library of work becomes orphaned. An agent can list everything a person owns, propose where it goes, and complete the transfer as one operation.

Library and asset hygiene

Every sales content library decays. Names drift from conventions, folders multiply, and outdated assets stay live because nobody has time to audit hundreds of items. Bulk renaming, tagging, moving, and archiving are exactly the kind of tasks an agent does well: tedious for a person, trivial when the rules are clear.

Reporting and adoption questions

“Which teams are actually using the licenses we bought?” “Which demos drove the most engagement last quarter?” These questions usually become a request to an analyst, a CSV export, and a spreadsheet. With an agent connected to the data, the person asking gets an answer in the conversation, and can follow up immediately.

Change management after a reorg

When territories, segments, or messaging change, the operational follow-through is a long list of small edits across tools. This is where strategy quietly fails: the plan changes and the systems do not. An agent that can apply a well-defined change across a whole library shortens the gap between the decision and the execution.

What MCP Does Not Change

It is easy to oversell this, so be clear about the limit. MCP removes clicking. It does not remove judgment.

An agent can archive every asset nobody opened since March. It cannot tell you whether the one asset a strategic account still relies on should be an exception. It can restructure a library in minutes. It cannot decide what the right structure is for how your team sells. It can pull adoption data instantly. It cannot tell you whether low adoption means the tool is wrong or the enablement was.

That is the right division of labor. AI is very good at shortening the distance between a decision and its execution. The decision itself, the story behind it, and the accountability for it stay with the person. Teams that get value from agents are the ones that get sharper about their instructions, not the ones that stop thinking about them.

Five Things to Require From Any Vendor’s MCP Server

Not every MCP server is built with the same care. An agent with broad write access to a system your buyers touch is a real risk if the guardrails are thin. Before you connect one, require these five things.

• Real actions, not just reads. An MCP server that only exposes search and summaries is a chatbot over your data. The value is in completing work: inviting, transferring, organizing, archiving.
• A preview before anything changes. The agent should show exactly which items will be affected, and wait for confirmation, before it writes. No preview means no undo moment.
• Permissions inherited from the user. The agent should be able to do exactly what the person instructing it can do, and nothing more. An agent with admin rights acting for a junior user is a privilege escalation, not a feature.
• Hard limits on bulk changes. A sensible ceiling on how many items one instruction can touch turns a bad instruction from a disaster into a fixable mistake.
• A read-only mode. Many people only need to ask questions. Giving them an agent that cannot change anything is the right default for them.

These are the same questions we argue for in how to evaluate AI sales tools: what happens when the AI is wrong, and who is accountable. Governance belongs in the evaluation, not after the rollout. If you are also working out who owns the demo library now that everyone can build one, start with demo operations.

What a Well-Built MCP Looks Like in Practice

Here is a concrete example of those five requirements in one product. Walnut, the interactive demo platform, built its MCP server around the operational work above.

The Walnut MCP connects to Claude and covers three jobs: member administration (inviting people, managing roles, offboarding members and reassigning their work), library management (searching, renaming, moving, tagging, and archiving templates, demos, playlists, and deal rooms in bulk), and analytics (session counts, completions, engagement, and seat usage). It works with the same permissions the user already has, so a general user can analyze and search while member management stays with account owners and admins. Before anything changes, it shows a preview of exactly which items will be affected and asks for confirmation. Bulk changes are capped at 500 items per operation, and access can be granted as full or read-only.

For teams that want the data in their own systems rather than in a conversation, the Walnut API exposes demo session data, per-user activity, and seat utilization for reporting in the tools RevOps already uses.

The point is not the product. It is the pattern: action, preview, inherited permissions, limits, and a read-only option. Hold every vendor in your stack to it.

Why 40% of Agentic AI Projects Will Be Canceled, and How to Avoid Being One

Gartner expects that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024. In the same research, Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls (Gartner, Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 2025).

Both numbers are true at once. Agents are arriving in nearly every tool, and a large share of the projects built on them will not survive. The difference between the two groups is rarely the technology. It is scope.

Projects fail when they start with broad autonomy: “let the agent run our pipeline.” Value is hard to measure, risk is hard to bound, and the first visible mistake ends the project. Projects survive when they start with bounded tasks where the before and after are obvious: onboarding took four hours and now takes ten minutes; the quarterly library audit took a week and now takes an afternoon.

Revenue operations is a good place to start precisely because so much of its work is bounded. The tasks are repetitive, the rules are known, and the time saved is easy to count.

How to Run Your First MCP Pilot in 30 Days

A pilot should prove value on one workflow before anyone talks about autonomy. Here is a sequence that works.

Week 1: Pick one workflow and time it. Choose something frequent and painful, such as onboarding a team or a monthly library cleanup. Measure how long it takes today, end to end, including the waiting.

Week 2: Connect read-only first. Give a small group of operations users read-only access and let them ask questions: who owns what, which assets are stale, how seats are being used. This builds trust in the agent’s understanding of your data before it changes anything.

Week 3: Turn on actions for the chosen workflow. Run the workflow through the agent with the preview step on every time. Have a second person review the first few previews. Track every case where the preview was wrong or incomplete.

Week 4: Compare and decide. Put the time, the error rate, and the user feedback next to the baseline from week one. If the workflow is faster and the previews were reliable, expand to the next workflow. If not, you have learned that cheaply.

The teams that get this right treat the agent like a new operations hire: a narrow remit, close supervision at first, and more responsibility as it earns trust. For the presales side of the same capacity problem, the SE team’s guide to scaling demos without burning out covers where human time should go once the busywork is gone.

The Take: Build for the Second User Now

Gartner’s 10 to 1 prediction will come true whether or not your team is ready for it. The question is whether those agents will be the sprawl kind, generating suggestions for humans to execute, or the useful kind, completing bounded work inside systems that were built to let them.

That choice is made now, in two places. In your own stack, by requiring MCP servers that act safely rather than just advise. And in your team, by getting precise about what you want done, because once the clicking is free, the quality of the instruction is the whole job.

Frequently Asked Questions

What is MCP in simple terms?

MCP, the Model Context Protocol, is an open standard that lets an AI assistant connect to business software and take actions in it. The software vendor publishes an MCP server describing what the product can do, and an assistant like Claude connects to it. Instead of clicking through screens, you describe what you want and the assistant carries it out within your permissions.

How is MCP different from an API?

An API is built for developers: someone has to write code to call it and maintain that code. MCP is built for AI assistants taking instructions from people who do not code. The assistant reads what the MCP server offers and works out which actions fulfill a plain-language request. Many products offer both, using the API for system-to-system data flows and MCP for conversational work.

Is it safe to give an AI agent access to our sales tools?

It can be, if the MCP server is built with guardrails. Look for a preview of every change before it happens, permissions that match the user instructing the agent, limits on how many items one instruction can affect, and a read-only option. Start with read-only access and one bounded workflow, then expand as the agent proves reliable.

What sales and RevOps tasks can MCP automate?

The best early candidates are high-volume, rule-based operational tasks: onboarding and offboarding users, reassigning ownership when someone leaves, bulk renaming and archiving content, applying changes after a reorg, and answering adoption and engagement questions. These tasks are tedious for people, easy to define, and easy to measure, which makes the value clear.

Will MCP and AI agents replace sales operations roles?

No. They replace the clicking, not the judgment. An agent can execute a library restructure in minutes, but someone still has to decide what the right structure is, which exceptions matter, and whether the change served the strategy. The role shifts toward defining rules, reviewing previews, and owning outcomes.

Why do so many agentic AI projects fail?

Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value, and weak risk controls. Most failures start too broad, with vague goals like “automate our pipeline.” Projects that survive start with one bounded workflow, measure the before and after, and expand only once the agent has earned trust.

Ready to see what an agent-ready demo platform can do for your pipeline? Start for free with Walnut.