Introducing the Querri Library and Librarian: The First AI-Native Semantic Layer Organized Around Business Questions and KPIs

Today we're shipping the biggest change to Querri since we became an end-to-end platform last fall: the Querri Library and the Librarian.

I'll get to what they are. But I want to start with the problem, because if you recognize the problem, you'll understand the product in about two minutes.

There's way more value in your data than what you can use. That's the real problem.

Nobody I talk to is short on data. Every company has a CRM, a billing system, a support tool, an ad platform, a warehouse or two, and more spreadsheets than they can count. The data is there. It just isn't living up to its potential… yet.

The research backs that up. In a global IDC study commissioned by Seagate, 1,500 enterprise leaders reported that only 32% of the data available to their organizations was being put to work. Splunk surveyed more than 1,300 business and IT leaders across seven economies and found they estimated 55% of their data was "dark", meaning untapped and often unknown. Different studies, different definitions, same shape: more than half of what companies collect never gets put to work.

What struck me in that research is why. The leaders in the Seagate study didn't say they lacked interest in their data. They named friction: making collected data usable, getting at data trapped in different silos, knowing whether the right data was even being collected. The problem starts after collection. A company can own a dataset and still fail to get value from it because nobody can find it, or somebody finds it but nobody understands it, or people understand it but don't trust it, or they trust it but only one specialist knows how to query it, and that specialist has a two-week (or two-month) backlog.

Companies don't need another place to store data. They need a better way to turn the data they already have into answers.

AI made asking easy. It didn't make the answer yours.

Two years ago it looked like AI would fix this. You type a question in plain English, the model writes the SQL, an answer appears. And for a demo database with five tables, it does.

Real companies don't have a demo database. When researchers built the Spider 2.0 benchmark out of 632 real enterprise analytics tasks, on databases that often have more than a thousand columns, the same AI agent that solved 91% of the older, tidier benchmark solved 17% of this one. Same models. The difference was the environment. Giving AI access to all of your data is not the same as giving it the right context.

Think about what happens when someone asks "Why did retention fall last quarter?" and the AI is looking at 400 tables, eight different customer identifiers, and five competing definitions of "active customer." It picks one. Ask again tomorrow and it may pick another. Two teams ask for last month's revenue and get two different numbers, and nobody can say where either one came from. The AI didn't make anything up. It just didn't know which of your truths was the one you meant.

There's a subtler version of the same problem, and we measured it. Most questions a business asks its data have one obvious reading and at least one other defensible one. Our best channel. Active customers. Average daily deliveries. Each can be computed two ways that are both correct and that produce different numbers. A system that gives you one of them and moves on isn't "wrong." It has made a decision on your behalf, in silence, and written it into a dashboard, a report, or a company strategy doc. We put this to the test on a public benchmark we didn't write: on questions the benchmark itself ships with more than one accepted answer, Querri identified that another defensible answer existed on 73% of its responses. A leading general-purpose AI agent, given the same data and asked the way a user asks, did so on 12%. On raw accuracy the two were near-tied. The difference wasn't who got the number. It was who told you there was another one.

With general-purpose AI, every answer starts from zero. Or maybe worse, from everyone's separate saved context in their personal agent. We wrote about this a few weeks ago: as an industry we got very good at storing the data and the answer, and much less good at storing the question and discussion that gave the answer meaning.

Start from the question, not the table

The Library is a new foundation for how Querri organizes your data. Instead of a catalog of files and tables, it's a map of your business. At the top are the questions you actually care about. We call them Anchor Questions. Each one connects to the KPI that tells you whether the outcome is improving, the Refining Questions that break it down, the Facts you need to read the number correctly, and the Views that compute it from your data. Everything related to one concern lives together in a Collection. Explore the Library.

Here's one in practice.

Your Sources, the joins, the permissions are all still there. You can't compute a real answer without them. What changed is the front door. You start from the question, and Querri already knows where to go.

That's what we mean by question-first. Every other tool starts from the data and hopes you can find your way to the question. We start from the question. Switch the Library to Visualize and you can see it on one map: your connectors and files, your Collections, your KPIs and Facts, with lines showing how they connect.

One number, one definition, for everyone

Remember the two teams with two revenue numbers? In the Library a KPI isn't a label on a dashboard tile. It's an actual calculation, written down once: measured from one table, with a value for each period. Open one and you can see exactly how it's calculated, which table it comes from, and its latest value. When that table rebuilds, the KPI recalculates on its own. And when anyone asks a question that touches it, Querri pulls in that KPI and uses its calculation. Whoever asks, it's worked out the same way. The headline number on the dashboard and the answer in chat come from the same definition, so there's nothing to reconcile.

When everyone works from the same definitions, and anyone can open an answer and see how it was computed, people stop arguing about whose number is right and start acting on it. And once a KPI or a View is saved, re-running it replays the exact code that ran the first time. No model is involved, so the report you scheduled can't quietly show a different number next month because an AI made a different judgment call. That's what being data-driven actually requires: not more dashboards, but numbers the whole team trusts enough to act on.

Every KPI sits in at least one of the Five Dimensions of Business Value: Revenue, Cost, Efficiency, Quality, and Risk. Churn can sit under both Risk and Revenue, because that's how businesses actually think about it. The dimensions give you a simple way to see what your business is measuring and where the blind spots are. Many companies discover they're tracking Revenue closely and very little else. More on KPIs and the five dimensions.

An answer you can check before you share it

The Librarian is the AI agent you're talking to when you ask Querri anything. It lives in a chat panel that follows you around the app: beside the Library, in your Inbox, on your lists of projects and dashboards.

When you ask a question, it doesn't start from raw tables. It looks first for a View or a KPI your team has already built that answers it, and uses that only when it fits the question you asked. If nothing fits, it works straight from your source data, and if the result is worth keeping it offers to save it as a new View. Either way, the answer shows the rows it returned, the exact calculation behind them, and which Views and Sources it read. On the Library map, the parts it looked at light up, so you can follow the route from the answer back to your data.

And if the data is too messy or simply isn't there, you're told what the gaps are and what steps you can take to move forward, instead of getting a confident number that happens to be wrong. I think that's one of the most important things we've ever shipped.

Behind that trail is a lot of checking you never see. Before a question becomes a query, Querri profiles what's actually in the columns it's about to touch: gaps, duplicates that disagree with each other, values that won't parse, text that doesn't match the typical format, numerical outliers. That's how Querri can flag that a churn rate that's arithmetically perfect has 2,777 of 2,796 records with no value behind it, and that you probably shouldn't use this number. The arithmetic isn't the problem. It's the tendency of AI to tell you what you want to hear and be confidently wrong, and that is a big part of what we've been solving for. Then the answer is worked out twice, by two routes that never see each other's work, and compared: same columns, same joins, numbers that reconcile. Where they disagree, you hear about it. And your own definitions, the rules your business runs on, reach both routes before either one starts. So the Librarian isn't deducing what "churn" means from a column name. How Querri checks an answer before you see it.

The Library pays off here in a second way. In our testing, a question the organization had answered before, with a View already saved in the Library, came back 4.6 times faster and used about a third of the tokens of an AI agent exploring the same database from scratch. Not because the model is faster. Because it isn't doing the work twice.

Know what to ask before you know what to connect

Most data tools are a blank slate until you connect data. Then once you do, you're staring at a list of tables and no clear steps to get to insights. You know your business fine. What you don't know is which table (or many tables, joined together somehow) holds the answer you need.

Querri flips data management around and starts with the real question:

You give it a few words on a topic, the Librarian suggests business questions, you pick the ones that fit, and you do that for three topics. Before you've connected a single file, you have a Collection with an Anchor Question, Refining Questions, and up to three KPIs, marked aspirational until there's data to measure them from. Then you bring in your data, and the questions are already waiting for it.

As your Library evolves, you keep working on all of this conversationally. Connect a file, ask a question, and the Librarian works from your source tables. Ask it to build a View, define a KPI, or group questions into a Collection and it does. Settle a business rule in conversation ("sales means paid orders only") and it records a Fact without a form. The structure builds up as you use it. For a team without a data engineer, that's the difference between a data project that starts and one that doesn't.

Find a problem once. Clean it once. Every answer after starts clean.

Remember the top obstacle the leaders in the Seagate study named: making collected data usable. Nobody's data arrives clean. Blanks hiding as "N/A." Amounts stored as text with a dollar sign in front. Two date formats in one column. An invisible character splitting one customer into two. Each one becomes a confidently wrong number downstream, and finding them is usually a Monday morning spent staring at an export.

In Querri that's the Data Audit's job. Every table you upload or connect gets more than 180 checks the moment it lands, before anyone asks a question. Each problem becomes a note with the evidence behind it: fifty-three spellings of customers' last names that fold down to forty. And the notes are ranked by what depends on the column. If a KPI or a saved question reads from it, a problem there would change a number somebody looks at, so it goes to the top. A messy column nobody uses can wait.

Then you press Clean this table, and the Janitor goes to work. It tries each fix on your own rows and stops only for the calls it can't make alone: month-first or day-first dates, which duplicate row to keep. It shows you the plan, and nothing is built until you say so. Then it builds a cleaned View with every fix applied, your raw data stays exactly as it arrived, and every new answer starts from the clean version. See how automatic data cleaning works.

And your Librarian learns while you sleep

Here's the part I'm personally most excited about. Overnight, Querri runs a pass over your Library. It groups what your team asked that week into themes and decides, theme by theme, whether a consistent and fast answer already exists, or should be built. It looks for questions nobody has asked yet that your data could answer. It profiles new data, finds the joins between your sources, and spots the same customer living under two names in two systems. And it reads each KPI's history and flags one that's moved in a way its history says is unusual.

None of that changes your Library on its own. Suggestions arrive as cards in your Inbox: Question worth a view, Metric to track, Worth remembering, Can't answer this yet. Each one says why it matters, who's been asking, and what accepting would add, with the evidence behind it. You accept, edit, or dismiss. The Library learns; you stay in charge of what it acts on.

That last card, Can't answer this yet, turned out to be a report I didn't fully appreciate until we saw it on real data. If eight people have asked which customers are likely to churn and nothing in the Library can compute it, that's not a failed query. That's your next data project, already prioritized for you. It's the dark-data problem from the other direction: instead of guessing which of your data matters, your own team's questions tell you.

Most software gets messier the longer you use it. The Library gets sharper. A dashboard shows you what you asked for. An analyst tells you what you should have asked. How Nightly Insights works.

The right data reaches the right people, without losing control of it

Remember the friction the Seagate study named: data trapped in silos. In most companies the silos exist for a reason. Sales owns its CRM. Finance owns invoicing. Marketing owns campaign performance.

Nobody wants to give out admin-level access to their carefully curated systems either, or pay for a hundred more licenses just so someone can pull the occasional report.

The Library fits the shape your company already has. A workspace is a room; a team's work belongs to the people in it. When Finance needs something from Sales, Sales doesn't need to hand over the raw tables. They build a View, cleaned, filtered, joined, and shaped to the question Finance actually asks, and share only that. Nothing is moved or copied, and Sales still owns it. When someone outside the company needs a result, a link reaches them, with a password and an expiration date. And underneath all of it, an access policy filters the rows person by person, so two people can open the same dashboard and each see only their region.

Each team lends its work to the teams that need it, without giving it up. That's how data stops being dark: by giving every department and employee the tools and control they need to share what they have, and access what they need, whenever and wherever they need it. Workspaces and row-level access.

Monitors that never sleep

Let's face it, not everyone wakes up most mornings thinking about data like I do. But that doesn't mean it's safe to refresh a dashboard once a month, or even once a week, and see if anything jumps out at you.

Your sales may look steady this month. But is every store steady? One may be crushing it with a new promotion or a viral product that half the country is missing out on, while an entire region dips because of inventory stock-outs. Averages can hold flat while one region slides and another climbs, and checking every breakdown by hand, every week, isn't a job anyone keeps up for long.

That's why we're also announcing Monitors. A Monitor watches a KPI per region, per store, or per rep, and tells you which one moved. Each slice is compared with its own past, never with the other slices. West being a third the size of East isn't news. West falling to a third of its own usual level is. Start one from any KPI or any View, break it down by up to two columns, and it checks every time the underlying data refreshes. Anything that falls outside its typical range lands in your Inbox, and in your email if you want it. So you can focus on running and growing the business, not watching it. Monitors will do that for you. Monitors are part of Nightly Insights.

Why we had to change the foundation instead of adding a chatbot

A chatbot on top of your data is easy to build. I can say that with some authority, because that's roughly where Querri started. What's hard is making the answers yours: grounded in your definitions, consistent with your dashboards, and getting better over time instead of resetting every conversation.

That turned out to require changing the foundation, not adding a feature. Once business questions and KPIs are the organizing structure and the data sits underneath them, a lot of things fall into place at once. You can write your questions down before you connect data, because the questions don't depend on the data. Chat and dashboards agree, because they compute from the same KPI. The Librarian can show you its reasoning, because it walks the same map you can see in Visualize. There isn't a second, hidden representation for the AI. One map, for people and for the agent, with every tool needed to draw out the right insight whenever you need it. That's why we call Querri the question-first data platform.

What this means for your business

Last fall I wrote that Querri had become an end-to-end conversational data platform: connect, clean, analyze, automate, share, all through conversation. I meant it. But there was one layer that was useful in the past and became critical in the age of AI: the semantic layer, the shared, agreed-upon definitions of what your numbers mean and how they're computed. It's the layer that turns "we have the data" into "we can answer the question."

Traditionally it takes a data team months to model, and it's written in a language only they can read. Now the Librarian builds it with you, in plain English, and maintains it from there. Which means the whole stack, from a raw QuickBooks export to a shared dashboard your CEO trusts, runs in the language of your business.

What that looks like depends on who you are.

For small businesses

You are the data team, usually at ten o'clock at night with a spreadsheet. Querri gives you what a company ten times your size pays an analyst for: the questions worth asking, the numbers that answer them, and a heads-up when one of those numbers moves. All from the tools you already run, whether that's QuickBooks, HubSpot, Stripe, Google Drive, or a spreadsheet, connected in a couple of clicks with help from the Connector Agent, and none of it waits for you to learn a new tool.

For consultants

Every client has their own definitions, exceptions, and strange little rules. What counts as an active account. Which months are fiscal. Why Q2 excludes that one region. Usually that context lives in notes, spreadsheets, or the consultant's head, and has to be rebuilt every time the work changes hands. The Library gives it somewhere to live, right next to the questions, KPIs, and data it applies to. Define it once, reuse it across the engagement, and every result gets easier to explain to the client. You get to their answer in the first meeting instead of the third, and you leave them with a Library they can keep asking after you've gone. One firm we spoke with saw it the other way around too: something simple enough to hand to clients directly.

For operations, customer success, and revenue teams

These teams rarely have simple questions. Revenue lives in one system, product usage in another, support history somewhere else, and "which accounts are at risk this quarter?" needs all three. The Library gives those questions a shared place to land. Instead of opening an analyst ticket or building one more spreadsheet, you ask directly, and you can still see which data and which definitions produced the answer. And the answer sticks. Your definitions, your recurring questions, and the number everyone agreed on stay put when people move on. Fewer meetings about whose number is right. More decisions made on the first number.

For data teams

This isn't about replacing the data team. It's about giving them a better front door. Common questions route to prepared, governed Views instead of being rebuilt from scratch in a chat window. Business definitions stay attached to the data they describe, so the rule for "active customer" doesn't live in a Slack thread from last year. And when people keep asking a question the company can't answer yet, the data team gets a clear signal about what to build next, instead of a backlog of one-off requests. Fewer weeks spent keeping pipelines alive and answering the same question five ways. More time building what the business actually needs.

For software companies

Your customers don't want a chatbot in your product. They want answers about their own business, in their own terms, that get better the longer they use you. The Library gives each of them exactly that: an analyst that learns their definitions, their questions, and their data, and gets smarter for them over time. And the whole thing can be completely white-labeled and embedded in days, not quarters, so you can ship your own AI analyst in a sprint.

Try it

The Library and the Librarian are available now for every Querri customer.

Already using Querri? The app looks different, and we've written up what changed and where things went. Your data, projects, and dashboards come with you, and the Library is additive. Log in, choose Revisit onboarding from Home, and the Librarian will build your question map around the data you already have. If you'd like us to walk you through it, request a demo.

New to Querri? Start free and the Librarian will ask you what keeps you up at night before it asks for a single file. Plan on 10 to 20 minutes to your first real answer.

Want to see it first? Book a demo and bring the question that keeps you up at night. We'll start there.

The Library - By Querri

The Foundation Your AI Is Missing

Check it out on Product Hunt →

You already have more data than you know what to do with. The hard part was never collecting it. It's finding the right data, understanding what it means, connecting it to the question, and turning it into an answer. That's the pain we set out to relieve. I'm glad to finally put it in your hands.

— Dave