The analyst isn't dying. The job is splitting.

Hex put a number on the floor this quarter: one customer, Mercor, used the platform to unlock $100M in revenue. Hex’s pitch is no longer “BI tool for analysts.” It’s “every operator does their own analysis.” That repositioning is not marketing fluff. The self-service BI category went from $8B in 2025 to a projected $33B by 2034. The money is moving to the seat, not the dashboard.

The narrative this triggers is wrong. AI isn’t killing the data analyst role. It’s revealing how much of the role was never analysis to begin with.

Pull a week of any working ops analyst’s tickets and you will find their time splits across three buckets.

Clerical work. “Pull last quarter’s churn by segment.” “How many seats did we add in May?” “Re-export the dashboard with the right date filter.” This is 50 to 70 percent of most ops analyst weeks at Series A through C SaaS companies. It is not analysis. It is translation between a business question and a SQL warehouse, repeated thousands of times because the business doesn’t speak SQL.

Translation work. “Marketing is asking why CAC ticked up but it’s actually a sourcing-mix shift, here’s the cleaner cut.” This is the middle 20 to 30 percent. It requires knowing the business, the data model, and what the asker actually meant. AI tools touch this badly today and well in 18 months.

Judgment work. “We can’t trust this number until we untangle three months of broken event tracking.” “The lift looks real but the population isn’t comparable.” “Here’s what I’d actually do about it.” This is the last 10 to 15 percent. It is the part of the job that compounds. It is the part you cannot automate at any stage of the current technology curve.

Text-to-SQL tools eat the first bucket cleanly. They are starting to chew on the second. They do not touch the third.

This is what every ops leader hiring a “data person” right now should be staring at. You are about to underpay for clerical work that’s about to evaporate, while the judgment work that you actually need stays unfilled.

Working: text-to-SQL with a semantic layer underneath it. Hex with dbt, Mode with their own metric definitions, Definite with a controlled schema. The semantic layer is the thing that keeps the answers consistent across users. Without it, you get the “Tableau problem” — every dashboard shows a different number for revenue and finance refuses to trust any of them.

Working: self-service for the question, governance for the answer. Operators ask in English. Tools answer with versioned, traceable, schema-bound SQL.

Demo-ware: raw text-to-SQL with no semantic layer. Looks magical on a five-row demo. Returns hallucinated joins on a 200-table warehouse. Skip until your governance is real.

Demo-ware: conversational dashboards. The interface is fine. The judgment isn’t there. Every “AI agent for analytics” demo right now is a thin wrapper on a model that doesn’t know your data model.

Sherif Mansour, Head of AI at Atlassian, said it cleanly: “AI isn’t magic, it’s plumbing.” (SaaStr AI Annual 2026 speaker materials.)

That framing is the most useful operator lens we’ve seen this year. If AI is plumbing, the question for ops leaders isn’t “can AI do my analyst’s job.” The question is “which of my analyst’s pipes are leaking, and which ones get replaced first.” Plumbing replaces clerical work. Plumbing does not replace the engineer who decides where the pipes go.

The ops leaders moving fastest right now are the ones running this audit on their own team’s outputs first, before their CFO asks them to.

If you’re an ops leader at a 20 to 200 person B2B SaaS company:

If clerical is over 60 percent, you are about to lose the seat you most need. Use the savings from a text-to-SQL deployment to fund the judgment hire you’ve been deferring. Don’t cut the headcount. Reshape it.

Throughput goes out every Friday at 9 AM ET. One signal, one case, one shift, one quote, one move. Forward this to one operator who’d find it useful. Reply with the company you’d want covered next week.