Your AI Agent Is Guessing. It Just Sounds Confident About It

Quick Answer

• AI agents don't fail because the model is weak, they fail because they lack business context: your definitions, KPIs, and rules.

• MCP (Model Context Protocol) is the connector that lets an agent securely reach your systems (CRM, calendar, ticketing) and take action, not just talk.

• Mosaic is the semantic layer that makes what the agent finds trustworthy: one definition enforced everywhere, with deterministic SQL instead of a model's best guess.

• Together, they cut token and infrastructure costs (37-70%+ depending on scale) while requiring no migration of your existing data, cloud, or LLM.

Every enterprise is now running the same experiment: point an AI agent to solve internal system problems, and see what it can get done. The results usually fall under a similar bucket, the agent is articulate, fast, and confident at first (and your instinct tells you to believe it) but then it often ends up wrong in ways that are hard to catch until someone points it out.

"AI is powerful, but it doesn't understand your business,"

says François Dupont, who leads Sales Engineering for Strategy across Central Europe.

"AI doesn't know your customer definition. It doesn't know your discount rules, your KPI priorities. It doesn't know what's been approved by your finance department."

That gap doesn't stop an LLM from answering… it just means the LLM fills the gap itself silently, and states the result with the same confidence it would use for something it knew. Ask the same question five times, Dupont points out, and "you may get five different answers, five different formulations, and it's not an AI problem. It's a shared definition problem."

Strip away the buzzwords and an AI agent doing real work looks almost unremarkable.

Take this example: A risk officer is driving to work with AirPods in, and he asks his assistant a simple question: any specific fraud alert from overnight? It doesn't just say "yes," it tells him exactly what happened: an anomaly in transaction pattern, hundreds of microtransactions in eleven minutes, all under a 500-euro threshold. That pattern has a name, called structuring, and is a classic way people try to move money without tripping a fraud alert.

He follows up with: who's responsible for this account? The assistant checks the CRM and comes back with a name and a branch. He asks for a meeting with compliance to sort it out. The assistant checks his calendar and the other two people's calendars, finds a slot that works for everyone, drafts a short summary of the situation, and books it. By the time he parks the car, the issue has been flagged, the right person has been identified, and a meeting is already on the calendar with the context attached. He walks into the office already caught up, instead of spending the first hour of his day getting there.

This might sound far-fetched, but nothing about that sequence is hypothetical. "It's powered by MCP together with Mosaic, and it's available right now," Dupont says. "We are not talking about the future. We are talking about reality."

The same pattern shows up in quieter, everyday work, too: an agent scanning customer health data, flagging an account with a steep drop in order volume, opening a retention opportunity in the CRM, and spinning up a follow-up task, all from one prompt.

"Everything that you can interact with today; you can connect to your AI. And MCP makes that possible."

— François Dupont, Sales Engineering, Central Europe, Strategy

MPC is like the USB-C of AI

MCP is the connective layer. "You can imagine MCP as a universal connector to connect your applications, your data, to AI," says Dupont. "It's like the USB-C connecting multiple worlds" one standard that reaches Salesforce, a calendar, a ticketing system, whatever the agent needs to touch, "bringing the data where it lives to AI, and also respecting governance and permission." An agent only sees, and only acts on, what the person behind it is cleared to see.

Mosaic is the layer that makes what comes through that connection trustworthy. Business logic — what a customer means, how a KPI is calculated, which rules apply — is defined once in Mosaic and enforced everywhere it's consumed, agents included. Rather than letting each model guess at how to translate a question into a query, Mosaic's engine generates the query deterministically. Ask the same question through five different agents and you get the same answer, because they're all working off the same definition instead of five separate guesses.

"Mosaic makes your data trustworthy. MCP makes it actionable."

— François Dupont, Sales Engineering, Central Europe, Strategy

1. Trust prevents AI from guessing at business definitions, ensuring consistent answers.

2. Cost claimed ~56% reduction in data/cloud compute costs and 50%+ reduction in AI token costs due to better context.

3. Flexibility decouples business logic from underlying tech (cloud, data platform, LLM), enabling "easy swaps" as technology changes.

Trust is the headline issue, but a quieter one is that every bit of missing context an agent has to compensate for costs money.

"Every unnecessary request, every clarification on a wrong assumption, burns tokens."

Multiply that by an organization running the same kind of query thousands, or tens of thousands, of times, and the waste stops being trivial. Feeding agents pre-defined business logic instead of a raw schema cuts token consumption by roughly 37% per query on average, climbing to 50–70% at production scale, and higher still for agents built to take direct advantage of it. On the infrastructure side, consumption costs drop by 56%, largely from caching that keeps repeat queries from hitting the warehouse at all.

The other objection Dupont hears constantly is really a question about lock-in: does adopting this mean re-platforming?

"Technology changes faster than business definitions," he says.

"Cloud evolves, data platforms evolve, AI models evolve. Your revenue, your customer base, your margin definition — that shouldn't have to change every time."

Because Mosaic sits as an independent layer above the LLM, the cloud, and the database rather than inside any one of them, switching any single piece (a new model, a different warehouse) doesn't touch the business logic underneath it. Dupont calls it "the easy swap": change what you consume with, or change where the data lives, and the definitions travel with you either way.

An AI agent without business context doesn't know it's guessing, so it sounds just as confident whether it's right or wrong. Closing that gap takes two pieces: MCP connects the agent to your systems so it can take action, and Mosaic makes sure what the agent finds is correct, so one definition is used the same way everywhere. Together, they make agents more trustworthy, cheaper to run, and easy to set up without changing your existing tech stack.