Crunch Is Open. Hold It to Both.

We spent the last few pieces arguing that a question isn’t a query, and that an answer you can’t check isn’t a cheaper answer. BumbleB Crunch™ is now open to early users, and the arguments become a test anyone can run.

The question usually shows up late in the day, after the work that pays the bills is done. Something in the numbers looks off, or looks too good, and there is nobody down the hall to ask. On a small team the analyst is whoever is still awake. The data is all there, in the store, the site analytics, the ad accounts and a few exported spreadsheets, and the question still goes unanswered, because turning it into something those tools can run is its own evening of work.

That gap has been the subject of this blog for months. Today it stops being only an argument.

What the last two pieces asked for

A Question Isn’t a Query argued that plain-English tools fixed the wording of a question and left the harder part, working out what the question actually is, where it always was. Every real question holds two things. One is what matters: what the company is deciding and what would change its mind. That part stays with the person. The other is the runnable form: which metric, which people, which period, which definition. That part is labor, and every generation of tooling has quietly handed it back.

You Do Not Need to Read SQL to Catch a Wrong Number argued the other half. The errors that change decisions are rarely arithmetic. They are errors of premise, choices about the business made silently by something that does not know the business. The founder who has never written a query is often the best person to catch them, provided the premises are written down where they can see them. And a system can only show work it actually did.

Put together, those two pieces describe a requirement rather than a product: something that takes the half-formed question, does the formulation in the open, and lets the person who knows the business check every reading it made along the way.

What opens today

BumbleB Crunch™ is the AI analyst that investigates. Point it at any of your data — your website, your store, the files you export, your own database — and ask in plain words, half-formed if that is how the question arrived. Crunch works out the real question, plans the analysis, runs it, and shows how it read you and how every number was counted.

Each of those verbs answers one of the arguments above.

Works out the real question is the formulation step, taken on rather than assumed. Vague is a valid place to start. Turning it into something analyzable is the first step of the analysis, not homework the person has to finish before typing.

Shows how it read you is formulation in the open. Plain words get an explicit reading: a phrase like “sign-ups” taken as form submissions, “read all the way down” taken as a stated scroll depth. Each reading sits on the page, in the language of the business, where a wrong one can be corrected in a line.

Shows how every number was counted is the premise half. The date window, what counts as a visit or a user, which denominator a rate uses: spelled out under every answer. What gets checked is the reasoning, not just the result.

Why the reading has to be visible

The case for showing the reading is not politeness. It is where the errors live.

A system that turns a sentence straight into a number has made a string of choices on the way, and if none of them is visible, the only thing the person can evaluate is whether the number feels right. That is exactly the check that fails when it matters, because a confident number for a question slightly to the side of the one asked usually feels fine.

When the readings are visible, the check changes shape. Nobody has to audit a query. They have to look at a sentence like growth, taken here as paying accounts rather than sign-ups and say whether that is what they meant. That is a judgment the person asking is better placed to make than any system, because it depends on knowing what the business is actually trying to decide. The division of labor holds: the system carries the work of making the question runnable, and the person keeps the judgment the whole answer turns on.

There is a short example on the site of what this looks like. A company spent the same amount on two ad channels and asks how many sign-ups each one brought in, what each sign-up cost, and whether the channel that looked cheaper per click was actually the cheaper one. The question spans two datasets, ad spend and website behavior, and it only means something once “sign-up” has a definition. The investigation states that definition, joins the two, and lands somewhere the click-level view would have hidden: the channel with the cheaper clicks was the expensive one per sign-up. The finding is the product. The chart beneath it is the proof.

Any data, deep where behavior lives

The reach is deliberately broad. Crunch takes a website through a single tag, a store, exported files, or a company’s own database through a connector that anonymizes every field inside the customer’s network before anything leaves. Two of those can sit side by side and be asked one question across them instead of stitching reports together by hand.

Breadth is not the claim, though. Plenty of tools accept files and warehouses. The depth is in behavior: what people actually do on a site and in a product. Crunch reads clickstream the way an analyst does — visits, funnels, sources, pages, repeat behavior — so drop-offs and channel quality come back as findings rather than raw counts. Having access to an event table is not the same as understanding what the events mean.

“Every analytics tool says it has AI now”

That is true, and it is the right objection to raise on launch day. Nearly every product in this market takes a sentence now, and most of them will describe what happens next in similar words.

The useful response is not a counter-claim. It is a test. Bring one real question, the kind that has been sitting unanswered, and ask it of anything that promises an answer. Then look at what comes back. Can the path from the words in the question to the number in the answer be followed? Is it clear what each word was taken to mean, which data was used, which period, which denominator? Could someone who knows the business but not the query language spot a wrong turn?

That test applies to Crunch exactly as much as to anything else. It was built to pass it, and the fastest way to find out where it does not is for people with real questions to try.

What early means

Early is meant literally. Crunch is open to anyone who wants to use it, free to start, and it will read some questions wrong. That is part of why the readings are on the page: a wrong reading that is visible is a correction, and a correction is the most useful thing an early user can give.

Alongside early users, BumbleB is looking for design partners: teams with a question their current tools keep getting slightly wrong, who want a hand in how the investigation works and are willing to say plainly where it breaks. The arrangement is simple. Bring the hard questions; expect the work to be shown; push back on any reading that is off.

For a long time the hard part of getting an answer from data was the mechanics, and the mechanics are largely solved. What is left is the part that was always there: working out what was really being asked, and being able to trust what came back. That is the part Crunch is built for, and from today it can be held to it.

Analysis isn’t translation. It’s investigation.

Bring the question your data hasn’t answered: hub.bumbleb.co/signup.

Key Takeaways

• The mechanics of asking data a question are largely solved. What remains is working out what was really asked, and being able to check what came back.
• A system that turns a sentence straight into a number has made choices you cannot see. Visible readings turn “does this feel right?” into “is this what we meant?”
• The person asking keeps the judgment; the system carries the labor of making the question runnable. The reading on the page is where the two meet.
• Breadth of data is not a differentiator. Depth of investigation is, and behavior is where that depth matters most.
• The fair test for any AI analyst, this one included: can the path from the words of the question to the number in the answer be followed?

This piece continues the argument in “A Question Isn’t a Query” and “You Do Not Need to Read SQL to Catch a Wrong Number”.