MCP connections vs. a unified data platform for CPG suppliers
The bottom line
MCP lets AI talk to your systems. A unified data platform makes sure what it hears is true. For most suppliers in the CPG world, the winning move is both, with the platform underneath.
- Simple questions: MCP wins. One system, one question, fast answer, no build required.
- The questions that run your business: the platform wins. Margin by customer, stock-out risk, cost variance and OTIF each pull from three to five systems that disagree on names, units and calendars.
- Someone always has to reconcile your data. With direct MCP, AI does it again in every prompt, out of sight. With a platform, it's done once, agreed with Finance and Operations, and reused by every tool. Direct MCP is spreadsheet sprawl for the AI era.
- Over time, the gap widens. A platform gets more valuable with each system you add. Direct connections get more complex with each one.
Two ways to connect AI to your business
MCP (Model Context Protocol) is an open standard that lets AI assistants connect directly to business systems. It's quick to set up and increasingly common. The question isn't whether to use it, but what it should connect to.
- Direct MCP: the AI connects to each system one by one and stitches answers together itself.
- Unified platform: your systems feed one lakehouse and semantic model. AI, Power BI, Excel and Tableau all draw from the same definitions.
The demo: simple questions vs. real questions
Ask the same AI assistant two kinds of questions in each part of the business. The simple question lives in one system, and direct MCP handles it well. The real question crosses systems, and that's where the approaches split.
| Area | Simple question (MCP does well) | Real question (needs a common model) | Systems involved |
|---|---|---|---|
| Sales | What did we invoice last month? | Which customers are growing volume but shrinking margin? | ERP, CRM, pricing and rebates |
| Production | What ran on Line 2 yesterday? | What's our true cost per unit by SKU vs. standard, and which lines drive the variance? | MES, ERP, labor and timekeeping |
| Inventory and planning | How much of SKU X is on hand? | Which SKUs stock out in the next 6 weeks, given forecast, open POs and the production schedule? | WMS, ERP, demand plan, 3PL |
| Procurement | What's open with Supplier Y? | How are raw material price changes flowing through to margin by finished good? | ERP purchasing, BOMs, commodity pricing |
| Quality | How many holds this week? | Do holds trace to specific lots, suppliers or lines, and what did they cost us? | QMS or LIMS, MES, ERP |
| Finance | What's AR over 60 days? | What's contribution margin by customer after trade, freight and deductions? | ERP GL, freight, deductions, customer portals |
| Customer and market | What did we ship to our largest customer? | Are sell-through, distributor depletions and our shipments in sync, or are we loading the channel? | ERP, distributor data, syndicated retail data |
What happens behind the scenes on a real question (Finance example):
- Direct MCP: The AI pulls invoices from the ERP, freight costs from the carrier system (keyed by shipment, not invoice) and deductions from a customer portal (with different customer names). It then invents its own rules to match and allocate them. Ask again next week, or let a colleague ask, and the number can change.
- Unified platform: The matching and allocation rules were agreed with Finance once and built into the model. The AI asks for Contribution Margin and gets the same number Finance closes the books with.
The translation problem
Every system describes the same business in its own language. Here is one customer, one product and one month, as three systems might record them:
| Concept | ERP | Warehouse or plant system | Customer or market data |
|---|---|---|---|
| Customer | Kroger Co. (acct 10442) | KROGER-DC-04 | Kroger Atlanta Division |
| Product | SKU 100245 | VAN-EXT-1GAL | UPC 0-12345-67890-1 |
| Unit | Case of 6 | Each | Pounds |
| Calendar | Fiscal 4-4-5 period | Calendar month | Week ending Saturday |
| Cost | Standard cost | Actual landed cost | Contract price |
Someone has to translate between these. The only choice is whether it happens once, in a governed model, or again in every prompt, where nobody owns the logic.
You've seen this before: spreadsheet sprawl
Direct MCP is the new version of everyone building their own spreadsheet model. Most organizations already know how that story ends.
| Spreadsheet models | Direct MCP prompts | |
|---|---|---|
| Where the logic lives | In the author's formulas and tabs | In the prompt and the AI's reasoning that day |
| Who understands it | The person who built it | The person who asked, if anyone |
| What happens over time | Copies multiply, versions drift | Every user and every prompt reconciles differently |
| The meeting it creates | "Whose number is right?" | "Why did the AI say something different this time?" |
| When the author leaves | Logic leaves with them | Logic was never written down |
There is one key difference: a spreadsheet at least sits in a file someone can open. The logic behind a direct MCP answer is often invisible and gone once the chat ends. The cure is the same one that solved spreadsheet sprawl: a shared, governed model that everyone builds from.
It's worse than spreadsheets in one more way: AI is probabilistic, not deterministic. A spreadsheet formula gives the same result every time. An AI model works by predicting the most likely next step, so the same question can produce a different approach, join or assumption from one run to the next. That isn't a bug; it's how these models work. When AI has to reconcile raw systems on its own, that variability lands directly in your numbers.
A semantic model takes that risk out of the math. The definitions and calculations are fixed and deterministic. AI's job shrinks to understanding the question and picking the right governed measure, which is what it does well.
Upsides and downsides
Neither approach is wrong. They're built for different jobs.
| Factor | Direct MCP to each system | Unified data platform |
|---|---|---|
| Time to first answer | Days | Weeks to months |
| Upfront investment | Low | Moderate to high |
| Single-system questions | Strong | Strong |
| Cross-functional questions | Weak: AI reconciles on the fly | Strong: reconciled once |
| Same answer every time | Not guaranteed | Yes, governed definitions |
| History and trends | Limited: many systems only expose current state | Full history, snapshots preserved |
| Data freshness | Live | As of last refresh (often hourly or daily) |
| Security | Access managed per user, per system | Managed once, in one place |
| Audit trail ("where did this number come from?") | Hard to trace | Traceable to source |
| Cost per question at scale | Rises with usage (API calls, AI processing, rate limits) | Low: queries hit an optimized model |
| Taking action (create a PO, update a record) | Strong | Not designed for it |
| Works with Power BI, Excel, Tableau and AI alike | No, each tool connects separately | Yes, one model serves all |
The long-term impact
A platform compounds value. Direct connections compound complexity.
- Each new system adds value to a platform. Add quality data, and cost, margin and supplier reports all gain context without being rebuilt.
- Each new system adds work to direct connections. Every new source has to be matched against every existing one, and that logic lives in prompts and people's heads. Different teams end up with different answers.
- Your business logic should outlast your AI tools. AI tools will change often. Definitions built into a semantic model carry over to the next tool. Logic buried in prompts doesn't.
- History you don't capture is gone. Many systems only show current state. A platform keeps snapshots, so you can compare this year's inventory or costs to last year's.
Which approach fits your organization
| Direct MCP is a good fit when you... | A unified platform is a good fit when you... |
|---|---|
| Run on one or two core systems | Run an ERP plus plant, warehouse, quality and customer data |
| Mostly ask questions inside one function | Need answers that cross sales, operations and finance |
| Are testing what AI can do for you | Need numbers that tie to the financial close |
| Want AI to take actions, not just report | Have several teams relying on the same metrics |
| Have little need for trends over time | Report to owners, lenders, a board or major customers |
| Plan to grow, add plants or acquire |
Signs you've outgrown direct connections: two people get different answers to the same question; Finance rebuilds AI numbers in Excel before trusting them; nobody can explain how a number was calculated.
The best answer: layer them
Mature organizations don't pick one. They use each where it's strongest.
- Foundation: a lakehouse that brings every system's data together, with history.
- Semantic model: one set of definitions for customers, products, calendars and metrics like margin, OTIF and cost variance.
- MCP on top of the model: AI assistants connect to the semantic model, not to raw systems, so they answer with the same numbers as Power BI, Excel and Tableau. Power BI already offers an MCP server for semantic models, so this works today.
- Direct MCP for actions: keep system-level connections for operational tasks, like creating a PO or updating a CRM record.
The result: AI is fast and flexible, and its answers are ones Finance and Operations will stand behind.
Next steps
Fluid Ascent builds the data platforms and semantic models that make AI trustworthy for CPG suppliers. A good place to start:
- Map your systems. We inventory where your sales, production, inventory, quality and finance data lives today.
- Pick three real questions. Choose cross-functional questions your team struggles to answer now.
- See both approaches on your data. We show what direct MCP returns and what a governed model returns, side by side.
You leave with a clear view of where AI can help today and what foundation it needs to scale.
