AI Data Readiness: 5 Signs Your Data is Ready for an MCP Server

Data streaming from a keyboard into checked-off documents and reports, representing AI data readiness

“Is our data even ready for this, or do we need to clean everything up first?”

That question came up in our recent Fresche Talks session, Beyond the Hype: What MCP Really Delivers for Your Data. Nearly every leadership team asks it before connecting AI to their business data.

The short answer is that you don’t need perfect data to start. You do need to know which of your data you can trust. That difference decides whether your first AI rollout builds confidence or burns it.

Watch the full session below for a candid, no-pitch look at governed AI analytics, including a live demo on our own data. Or read on for the key points.

Jump to a topic:

  • 2:00 – Why self-service analytics never delivered
  • 5:41 – What an MCP server actually does
  • 6:46 – Three must-haves in any MCP server
  • 9:32 – The data gap we fixed overnight
  • 12:25 – Live demo on our own data
  • 17:40 – “Can’t we just point Claude at our warehouse?”
  • 22:22 – Is your data ready for AI?
  • 26:41 – Isn’t this just BI with a chatbot?
  • 28:38 – MCP servers vs. REST APIs

What is AI Data Readiness?

AI data readiness means a set of your data is accurate, defined the same way everywhere, connected across systems, and governed well enough for an AI assistant to answer questions from it reliably. You measure readiness one domain at a time, such as finance, sales, or delivery, not across your whole data estate at once.

An MCP (Model Context Protocol) server is what connects AI assistants such as Claude, Copilot, or ChatGPT to that data. “An MCP server is simply the layer that allows the large language model to communicate with any system, any application,” Tobias says. That can be your data warehouse, your semantic models, or systems like your ERP and CRM.

A governed MCP server adds three controls:

  1. Deterministic answers: Every revenue question runs the same approved calculation.
  2. User-level security: The server runs as the person asking, so their existing permissions apply.
  3. Audit logging: You can see who asked what and whether they got a good answer.

Why Does Data Quality Matter More With AI?

“It’s the same concept we’ve spoken about for decades,” Tobias says. “Garbage in, garbage out.”

What AI changes is the speed and confidence of the output. A flawed report used to reach a handful of people after days of work. A flawed AI answer reaches anyone who asks, in seconds, written in clear and convincing language. Bad data no longer looks bad; it looks finished.

Without a governed layer, Steve says, the AI is “basically guessing what your question actually means.” It doesn’t know how your finance team treats discounts, rebates, or intercompany transfers, so different people get different answers. “If you want a really rapid way of decreasing confidence in AI and getting people to stop using those tools, that’s it.”

A governed MCP server fixes that. In the live demo, a question about revenue over the last 12 months went straight to the approved finance revenue measure. “There’s only one way to get revenue for our company,” Tobias says. But consistency cuts both ways. If the underlying number is wrong, everyone now gets the same wrong number.

Complexity raises the stakes. Most business questions cross systems. Days sales outstanding (DSO) needs invoices from the ERP and customer records from the CRM. Project margin needs bookings, delivery hours, and billing. If those systems don’t connect cleanly, the AI either guesses at the join or can’t answer at all.

That’s why a data warehouse matters so much. “That’s where we see organizations have the most success,” Tobias says. “Organizations that already have a mature and validated data warehouse can get started with an MCP server in weeks, not months.”

How Do You Know Your Data Is Ready for AI? Five Signs

Use these five checks on any data domain. A domain that passes all five is ready for AI.

  1. Your key metrics have one agreed definition.
    Finance, sales, and operations calculate revenue, margin, and bookings the same way. Those definitions live in a semantic model or metrics layer, not in individual spreadsheets.

  2. Your systems connect.
    Customer, product, and project IDs match across your CRM, ERP, and delivery systems, including IBM i. You can follow a deal from quote to invoice to cash without manual matching. Our own quote-to-cash gap, described below, is what failing this check looks like.

  3. People already trust the reports.
    Leaders use existing dashboards in meetings without asking where the numbers came from. Trusted reporting is the strongest sign that a domain is ready.
  4. Access rules are defined.
    You know who can see salary data, customer contracts, and margin by client, and those roles exist in your data platform today. Because the MCP server runs as the user, good data governance carries straight through to AI.

  5. Someone owns the data.
    Each domain has a named owner who fixes errors and approves changes to definitions. Without data ownership, issues that AI surfaces have nowhere to go.

Few companies pass all five checks in every domain. That’s normal, and it doesn’t stop you from starting. If you run on IBM i, the IBM i Analytics and AI Readiness eBook walks through readiness for that platform.

What If Your Data Isn’t Ready? Start With One Trusted Domain

You don’t need to fix your entire data estate before giving people AI access. You need one domain you can stand behind.

Steve sees this often in real deployments. Organizations “might not have a very mature data warehouse, but they have some reporting that they feel a lot of confidence [in]. They’ve gone through a lot of validation. That’s a great starting point.”

That domain is often finance reporting that has been through month-end close, audit, and repeated validation. It might be the sales bookings your leadership team reviews every week. Whatever it is, people already use it to make decisions.

A governed MCP server lets you choose exactly which domains to expose. Start with the trusted one and hold back the rest. Users get reliable answers from day one, and early wins build the confidence that drives adoption.

It also protects trust. A user who gets a wrong answer in week one may never come back. A user who gets three right answers in a row starts to rely on the tool. As each new domain passes the five checks, you add it, and your AI coverage grows as fast as your data quality.

Can AI Help You Explore Data You Don’t Trust Yet?

Yes, with the right audience. Steve calls this one of the more surprising uses he’s seen: a governed MCP server “can be a great tool for exploration of new data.”

One company we work with is going through an acquisition. The new business brings different verticals, different data architecture, and different platforms. Its analysts are using the MCP server to explore that data early, so they can decide whether to consolidate everything onto one platform or combine the data at the report level.

The difference is who gets access. Exploration belongs with analysts who know the data hasn’t been validated. Business users get access only once a domain passes the five checks.

How to Get Your Data AI-Ready in Six Steps

A small team can run this process, or you can work through it with us in an AI in Action workshop.

  1. List the questions leaders ask most: Collect the 15 to 20 questions behind weekly and monthly decisions, such as revenue trend, pipeline coverage, or overdue receivables.

  2. Trace each question to its source: Note which system holds the data, which definition applies, and who owns it. Gaps show up fast.

  3. Score each domain against the five checks: Pass, partial, or fail is enough. Don’t over-engineer this step.

  4. Launch with your strongest domain: Expose it through a governed MCP server to a small group of users who ask those questions every week.

  5. Watch what users ask. Logs show which questions the server answers well and which fail. Failed questions point straight to the next data gap to close.

  6. Expand one domain at a time: Fix the gaps, rescore the domain, and open it up. Repeat.

Step five is the one most teams skip, and it’s where we’ve seen the biggest payoff.

How We Found a Data Gap Nobody Reported

At Fresche, we run our own MCP server internally. When our new CFO joined in May, he asked IT for a quote-to-cash report to track how long it takes from first quote to cash in the bank, so he could better plan cash flow.

The report stalled. Quotes, invoices, and payments lived in three different systems, and the link between them wasn’t exposed. The analyst building it moved on to other work and didn’t flag the problem.

That evening, an agent that reviews failed user requests spotted the attempt. It worked out what data was missing and sent our analytics team a proposed fix: join the CRM quotes to the ERP invoices. By the next morning, the team had published a patch to the MCP server and told the analyst. He finished the report, and it’s still in use today.

“That’s the first time I’m seeing this sort of behavior in my career,” Tobias says. “Users that run into an issue never report it. We can capture it, and before they just give up on this new system, we’ve solved the issue.”

Usage data tells you where to invest next, so your data roadmap follows real business demand instead of guesswork.

Frequently Asked Questions

Do We Need to Clean Up All Our Data Before Using AI?

No. You need validated data in the domains you expose, not a perfect data estate. Start with one trusted domain, such as finance reporting, and add others as they pass the five readiness checks.

How Long Does It Take to Deploy an MCP Server?

It depends on your data foundation. According to Tobias Eld, organizations with a mature, validated data warehouse can get started in weeks, not months. Organizations without one should first build reliable reporting on their most important data.

We Already Have Copilot, Claude, or ChatGPT. Why Not Point It Straight at Our Data Warehouse?

You can, but without a governed layer the AI infers what your question means each time, instead of using your approved definitions. Different people get different answers, and you lose visibility into who is accessing what.

Is an MCP Server Secure?

A governed one is. It runs as the user asking, so people see only what they already have permission to see. It uses strong authentication, guards against prompt injection, logs every request, and runs in your own tenant. “You shouldn’t let anybody else host your MCP for you,” Tobias says. Blocking AI access is often riskier, because people upload exports to personal AI accounts instead.

Isn’t an MCP Server Just a BI Tool With a Chatbot?

No. It’s a governed layer between your data and any AI tool, not a feature of one chatbot. It works whether you run on Microsoft, IBM, Snowflake, or AWS, and with whichever AI model fits your needs, so you can switch tools as AI changes without rebuilding your data access.

Start Where Your Numbers Hold Up

Data readiness isn’t all or nothing. You don’t need a perfect data estate to get value from AI analytics. You need to know which data you trust, expose that data first, and let real usage show you what to fix next. Trust grows from there.

When you’re ready to go deeper, MCP Servers Explained: Your Guide to Governed AI and Self-Service Analytics is our free guide for IT, finance, and operations leaders who want AI working on their data without ripping anything out. It covers the one prerequisite that decides whether an MCP server will work for you, how governed AI compares with Fabric Data Agents, warehouse-native AI, and ChatGPT builds, and what security and deployment timelines look like.

Download the free Fresche ebook MCP Servers Explained: Your Guide to Governed AI and Self-Service Analytics

About the Authors

Tobias Eld is General Manager of Fresche Analytics at Fresche Solutions, where he leads Analytics Operations. Over more than 20 years, he has worked as a data warehouse developer, CTO, and GM, designing enterprise data architecture for retail, CPG, and manufacturing companies across Microsoft Fabric, Azure, D365, and IBM i environments. He designed and built a governed MCP server that’s now in production, handling hundreds of queries a day with user-level security and audit logging. Connect with him on LinkedIn.

Steve Mika is Commercial Lead for Data & Analytics at Fresche Solutions, where he helps organizations build the data foundation that AI depends on. He has more than five years of data and AI consulting experience across healthcare, senior living, government, and manufacturing, specializing in Microsoft Fabric, Power BI, and Azure AI. He works with directors, VPs, and C-suite leaders to align data strategy with business goals and move AI projects past the pilot stage. Connect with him on LinkedIn.

Request a Demo

See our solutions in action.

Book a Discovery Call

Talk with our team about where you are and what comes next.