AI-ready data
Your AI can't answer questions about your business.
You bought the licenses. It drafts email beautifully. But ask it why margin fell in the Midwest and it either refuses, or — worse — makes up a confident number. That isn't a model problem. Nothing in your company has ever assembled the answer.
Tuesday, 9:14am — $8.4M
Tuesday, 4:02pm — $7.9M
No error. No warning. Two numbers, one of which is going in a board deck.
The part everyone skips
The interface was never the hard part.
Connecting an AI tool to a database takes an afternoon. That's why it feels like the whole problem is solved. It isn't — because a database connection hands the model a wall of raw tables and asks it to guess what your business means.
On tidy textbook databases, frontier models write correct SQL about 86% of the time. Point the same model at a real enterprise database — the kind with a thousand columns and fifteen years of naming decisions — and it gets 6% right.
- Your schema isn't self-explanatory. Nothing in it says which of your four revenue columns finance actually reports.
- The joins don't exist yet. "Mike" in your ERP and "Michael" in your CRM are two different customers until someone proves otherwise.
- Guessing looks like answering. A model with no context doesn't stop. It fills the gap and returns a number.
Frontier models on the Spider 2.0 benchmark; bottom row from Snowflake's own internal Cortex Analyst benchmark
The failure mode nobody plans for
It doesn't fail loudly. It fails confidently.
A broken pipeline pages someone at 3am. A broken answer gets forwarded to your board. When an AI tool queries your systems directly, the ways it goes wrong are all silent.
- Results get cut off mid-answer. Assistants cap how much data a tool can return. The model isn't told anything was removed, so it answers as if it saw everything.
- The same question drifts. Practitioners running these setups at temperature zero still get different SQL and different answers between runs.
- Longer context makes it worse, unpredictably. Research across eighteen frontier models found reliability degrades as input grows — and not in a straight line you could plan around.
- Nothing is reproducible. There's no lineage, no trace, no way to ask the answer where it came from.
Consistency is the whole basis of trust in a number. If asking twice gives you two answers, the meeting stops being about the decision and starts being about whose number is wrong.
# requested: all orders, last quarterrows_returned: 1,000rows_in_table: 486,214truncated: silently# the model was not told. it answered anyway.Assistant tool-result caps are measured in tokens, not rows
Token economics
The "cheaper" way is the expensive way.
Pointing an AI tool straight at your systems feels free. What it actually does is push your analytics into the slowest, most expensive layer in your stack — and charge you per question, per person, forever.
- The menu costs more than the meal. Tool definitions alone can consume tens of thousands of tokens before anyone asks anything. Anthropic's own engineers documented a workflow going from 150,000 tokens to 2,000 by keeping that overhead out of the context window.
- Every row travels through the model. Anthropic again: intermediate results pass through the context window, and with large payloads "models may be more likely to make mistakes when copying data between tool calls."
- More connections, worse answers. An academic study across six commercial models and thirty tool suites found that adding tool-protocol context degraded task performance by 9.5% on average while inflating input tokens by up to 236×.
A warehouse does the arithmetic in the warehouse and hands the model twelve numbers. That's what it's for.
Figures from Anthropic engineering and the MCPGauge study
Structurally out of reach
Some questions a connection can never answer.
"How did this trend over 18 months?"
Your systems store what's true right now. When a deal changes stage or an order gets re-dated, the old version is overwritten. Nothing can query history that was never kept.
"Profit per customer, across every system"
That number needs your CRM, your ERP, and your billing system to agree on who a customer is. There's no shared key. Entity resolution is work, not a prompt.
"Give me every order from last year"
Business systems throttle API traffic on purpose — often a hundred requests per ten seconds. You hit the wall long before you finish the question.
"Where did this number come from?"
An improvised query leaves no audit trail. You can't check work that was never written down.
"Tell me when this changes"
Chat is pull-only. You have to remember to ask. Nothing is watching your data between conversations.
"Whose definition of churn is this?"
Metric definitions live in budgets, board decks, and people's heads. A model reading your schema has no way to know which one your company actually reports.
Not our opinion
Everyone who built this concluded the same thing.
This isn't a Synopsis argument. Every serious AI-analytics product on the market requires a modeled layer underneath it — which is the clearest possible evidence that the layer is the product.
- Databricks caps its AI analyst at thirty tables, and its own documentation tells you to pre-join related tables into views first. Pre-defined metrics exist specifically to improve response accuracy.
- Microsoft requires a star schema, explicit measures, and an AI-readiness pass before Copilot works on your data. Their guidance: these are "prerequisites, not optional polish."
- Snowflake built an accurate natural-language analyst — and it doesn't run at all without a semantic model. Same models, 51% to 90%+, on the strength of the layer alone.
"Data and analytics agents are essentially useless without the right context."
Andreessen Horowitz — Your Data Agents Need ContextThe disagreement was never about whether you need this layer. It's about whether you spend six months building it.
Every serious platform requires a modeled layer. The only question is who builds it.
What we actually do
We're what your AI should be pointed at.
Synopsis is the layer that makes the answer possible: every system connected, the data cleaned and joined, duplicate records resolved, your business modeled, and your own metric definitions enforced on every surface. Then we hand that to whatever AI you already use.
- Connected, not exported. Hundreds of pre-built connectors, custom ones built in minutes, including the spreadsheet your CFO swears is temporary.
- Modeled automatically. Raw events become clean, linked entities. No schema design, no transformation project, no engineer to hire.
- Your definitions, reverse-engineered. We learn how your company calculates its numbers from documents you already trust, then reconcile until our figures match the ones you report.
- History that persists. Your data gets a memory, so "was it always this bad?" becomes an answerable question.
- Governed and traceable. Permissions, lineage, and an inspectable query behind every answer. SOC 2 Type II, in your warehouse.
- Open to your AI tools over MCP. We're not a replacement for the assistant you like. We're the substrate that makes it right.
Midwest gross margin fell 4.2 points, almost entirely from freight. Three accounts moved to expedited shipping in March and never moved back — together they're $1.9M of revenue at 11 points below regional average.
Verified sources: ERP · order management · billing · carrier invoices
Bring the question nobody can answer.
Think of the one question about your business where the reply is always "let me pull that together." Give us twenty minutes and we'll answer it on your data, live.
Sources
- Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows — arXiv:2411.07763
- Help or Hurdle? Rethinking Model Context Protocol-Augmented LLMs (MCPGauge) — arXiv:2508.12566
- Anthropic — Code execution with MCP
- Chroma — Context Rot: How Increasing Input Tokens Impacts LLM Performance
- Snowflake — Cortex Analyst: Evaluating Text-to-SQL Accuracy for Real-World BI
- Databricks — Curate an effective Genie Agent
- Microsoft — Semantic model best practices for data agent
- Gartner — Lack of AI-Ready Data Puts AI Projects at Risk
- Gartner — Over 40% of Agentic AI Projects Will Be Canceled by End of 2027
- Andreessen Horowitz — Your Data Agents Need Context
- CloudQuery — MCP Server Gotchas We Learned the Hard Way
