Last week the data industry spent four days in San Francisco on familiar 2026 themes: agents, governance at scale, and natural-language query over the warehouse. Then the conference ended, the analysts flew home, and on Monday morning they opened Excel.

That sequence repeats after every major conference. It is not that the audience missed the point, or that their firms are behind. A large share of the work those analysts do has never had a home anywhere else.

Two kinds of analysis

---

Business users do two broad kinds of data work, and the two behave differently.

The first is scripted. It is the recurring revenue figure, the weekly exposure report, the metric someone monitors on a cadence. This work has an owner, a refresh schedule, and a known shape. It is the work that dashboards, BI platforms, and now natural-language assistants were built to serve, and they serve it well. If your question is one a hundred people will ask again next quarter, it deserves to be modeled, and modern tooling models it cleanly.

The second is unscripted. Sometimes it is the question that will not hold still — the one an analyst chases across a sprawling set of workbooks, returning day after day to tinker, because each answer reshapes the next question in a direction no one mapped in advance. Sometimes it is the genuine one-off — the what-if a deal team needs by end of day, the hunt for the single transaction that looks wrong, the extract an auditor has requested in a particular layout by Thursday, done once and not in that shape again. Neither has an owner or a fixed cadence, and both live in Excel.

Every wave made the same promise

---

Every wave of analytics tooling has promised to absorb that second category, and every wave has optimized the first one instead.

Business intelligence promised self-service and delivered governed dashboards. Self-service analytics promised ad-hoc exploration and delivered a better catalog of the questions someone had already anticipated. The current wave of natural-language assistants promises that you can finally just ask, and for anticipated questions, you mostly can.

This year gave an unusually exact measure of where that “mostly” ends. Databricks’ marquee announcement at the summit was Genie Ontology, a context layer that encodes a firm’s own business semantics — what its fiscal year is, what counts as a churned customer — so an assistant can resolve a familiar question against governed definitions. By the figures Databricks showed, it lifted the best agent from roughly 50 percent accuracy to about 84. The improvement is real, yet bounded: the ontology sharpens only the questions a firm already knew to encode. The unscripted question is the one it has never seen.

An analyst who answered business questions correctly 84 percent of the time would not last. For work that ends in a regulatory filing, an audit response, or a figure a board will treat as fact, an answer wrong on roughly one question in six is not a savings but a liability — the cost of the one wrong answer swamps the rest. And the verdict does not soften at 90 or 95 percent: close only counts in horseshoes.

In each case the scripted category got faster and cleaner, and the unscripted work flowed back to Excel. That recurrence — not any single tool’s limits — is the evidence.

Why the work returns

---

Two things send the work back to the spreadsheet, and neither is habit.

The first is the surface it needs. A question that keeps changing shape cannot be met by a schema built in advance, because the next variation is the part no one modeled. Excel is a blank computational grid: every cell is addressable, traceable, and editable by hand. That is what makes “find the one row that looks wrong” and “now change this assumption and watch the rest move” native operations rather than features bolted onto a fixed query. Governed platforms make the anticipated easy and the unanticipated hard, which is the deliberate cost of governance, not a defect — but the unscripted question is, by definition, the unanticipated one.

That same surface is the answer to the accuracy problem. The fix is not that a model must never help form the question, but that whatever it returns has to land where a person can see it, trace it, and correct it before it counts. A confident answer you cannot inspect is unusable for work you will have to defend, at 84 percent or at 98.

The second is the form the work takes when it leaves. Much unscripted work ends not in a decision but in a file handed to someone outside the firm. You cannot hand a regulator a chat session. You cannot email an auditor a dashboard, or a link to a notebook they hold no license to open. The deliverable is a file, in a shape they specified, that they can open, check, and keep. The moment data has to leave your hands and land cleanly in someone else’s, the spreadsheet is not a fallback. It is the format of record.

The category that never gets a dashboard

---

The spreadsheet is not surviving in spite of modern data tooling. It is the permanent home of a category of work that, by definition, never gets a dashboard — because the moment a question is worth a dashboard, it has left that category. The two will not converge, because they answer different questions. One is built for the question you saw coming. The other, for the question you did not, and will still have to defend.

That leaves one unsolved problem. If that work lives in Excel permanently, getting current, governed data into the spreadsheet — without analysts falling back on stale exports and uncontrolled copies — is not a workflow detail. It is the problem itself. That layer is what we build.