4 questions to get started with AI agents

Hi Everyone,

In December 2025, Anthropic, OpenAI, Google, Microsoft, and AWS agreed on a single open standard for how AI talks to your business tools. The Model Context Protocol (MCP) now has more than 10,000 servers running publicly. A year ago, it didn't exist.

AI agents are starting to do what tools like Zapier have done at the team level for years. They move information between your systems and decide what to do with it.

Today, we're breaking down what's shipping, where it works, and four questions to ask before you commit.

What's shipping now

By late 2025, Salesforce had closed 9,500 paid Agentforce deals, up 50% in a single quarter. Combined with Data Cloud, the product line is now at $1.4B in annual recurring revenue. 1-800Accountant resolved 70% of chat engagements during peak tax season. Pandora deflected 60% of customer cases.

Microsoft 365 Copilot rolled out to 31,000 UK government seats after a 20,000-person trial. A Forrester study commissioned by Microsoft found users recaptured roughly half their time on tasks Copilot touched. Amplitude built five AI agents on Workato's platform that handle 60+ use cases and save over 4,500 hours per week.

Why is this different from traditional integration

Old-style integration moves data on rules. A trigger fires, and a workflow runs with a predictable result. Zapier has done this at the team level for years. Larger platforms like Workato and Boomi do the same at a company scale.

Agents read across your CRM, your email, your support tickets, and your product data, then decide what to do. The same input can produce different outputs, which is both the value and the risk.

The Model Context Protocol gives agents one way to connect to any system that supports it. Before MCP, every vendor built proprietary connectors. Now, an agent built on Claude or ChatGPT can read from your Notion docs and update your Linear tickets through the same protocol.

Where it isn't working

McKinsey's 2025 State of AI report (1,993 organizations) found that fewer than 10% are scaling agents in any function. In product development, 73% aren't using agents at all. A Zapier survey of 546 enterprise executives in September 2025 found only 9% have implemented AI agents at scale, and 91% face friction from AI tool sprawl.

Across the McKinsey and Zapier data, the same two reasons keep coming up. Data feeding the agents is messy, and no named person has authority to pull the agent back when its outputs drift. The technology rarely causes the failure.

Four questions to ask before you commit

1. Does it speak MCP? If an AI agent can't connect to your other tools through an open standard, you're locking into their ecosystem. As of December 2025, the major model providers and enterprise platforms all support MCP.
2. What's the data behind it? An agent that reads your CRM produces good answers only if your CRM is clean. A 2026 Salesforce study found 69% of marketers can't respond promptly because they can't access the context they need.
3. Who owns the agent when it makes a bad call? The output of an AI agent varies more than a fixed workflow does, so the owner can't be 'the AI team' as a group. Name one person who reviews what the agent does and decides when to pull it back.
4. What's the kill criterion at 90 days? Pick a measurable result up front, such as cases resolved or hours saved, and stop the project if you're not hitting the success metrics.

Try this today

Open your current AI tool list. For each tool, write down which other systems it reads from and which it writes to. A tool that doesn't read from or write to anything is simply an AI feature inside a single product. And the ones that connect through workarounds and one-off scripts are building integration debt every week.

Go deeper

👉 Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation – the official announcement of MCP becoming an open standard in December 2025.

👉 Forrester: The Total Economic Impact of Microsoft 365 Copilot – an independent measurement of where Copilot saves time and where it doesn't.

👉 Workato: How Amplitude Saved Thousands of Hours with AI Agents – a detailed customer story on building five AI agents that handle 60+ use cases.

👉 Anthropic: Introducing the Model Context Protocol – the original launch announcement from November 2024, useful for understanding what MCP solves.

Coming up tomorrow

Tomorrow, we're covering how to add a forward-looking metric to your board dashboard so your directors can act before the next miss.

That's it for today!

P.S. Is anyone in your org specifically owning AI agents, or is it spread across teams? We're curious how this is shaping up.