Francois Ajenstat has helped build or shape nearly every major analytics platform of the last three decades. By his own account, what finally pushed him to start a company was going looking for “a Cursor for data,” and finding nothing, and coming away "demoralized."¹ As founder and CEO of Golden Analytics, Ajenstat is now betting that large language models can finally deliver on something business intelligence never quite delivered: giving every person in an organization, not just trained analysts, the ability to ask a real question and get a real answer.
His résumé reads like a history of the BI industry itself. He started his career at an SGML document-parsing startup in Ottawa, then Cognos, (where he first heard the term "business intelligence" used in earnest), then a decade at Microsoft across the SQL Server and Office teams.¹ Then thirteen years at Tableau, more than seven of them as chief product officer, through the company's IPO and its $15.7 billion sale to Salesforce, where he helped popularize the self-service analytics movement.² A stint as chief product officer at Amplitude followed, and today he leads Golden Analytics as its founder and CEO.
On a recent episode of AI Radicals, "Is Business Intelligence Truly Dead?", host Satyen Sangani sat down with Ajenstat to talk through what he thinks the industry, and his own new company, need to get right this time.¹
"The chart was never the result"
Sangani opened by putting his own spicy take on the table. Seven or eight months earlier he had posted on LinkedIn that BI was dead, on two grounds: that models would commoditize visualization, and that they let a person ask a question in their own language rather than in the semantics some report developer had chosen for them.
Ajenstat didn't disagree with Sangani so much as reframe the claim. "Creating a chart, creating a visualization is not the hard part," he says. "It was allowing you to ask questions easily, because that's actually the part that I think is really, really hard."¹ Tools like Tableau, in his telling, were never really about the charts.
His sharper point was about what a dashboard actually is. Most dashboards today, he argues, have quietly become single-purpose reports again, a "new dashboard factory" that answers one question well but can't handle the next. "The chart was never the result," Ajenstat argues.¹ The real product, in his view, was always the journey from raw data to insight, something a static semantic layer can only partially support without a way to keep asking follow-up questions.
A semantic layer being cheap to generate is not the same thing as it being right. Anthropic ran a similar experiment, having a model auto-generate metric definitions from raw tables and query logs. It found that the output looked plausible while encoding the very ambiguities the layer existed to eliminate, and tested worse against its evals than a smaller, human-curated layer. Its conclusion was to let the model draft the documentation and keep a human owning the definition.³ That distinction turns out to be the subject of the rest of the conversation.
Building the "Cursor for data"
Golden Analytics exists, Ajenstat says, because of a gap he felt personally. After spending time with AI coding tools and feeling, as he puts it, unleashed, he went looking for an equivalent for data work and came up empty. "Why do data people have to have these tools?" he recalls thinking. "Why can't they have a Cursor or a Claude for data?"¹
That question led to Golden's core design principle, which Ajenstat calls a "slider of autonomy." It isn't a literal UI control, but a stance: the AI is available at every step, and the user decides how much of the work to hand over. Do the whole analysis manually. Let the AI take it 90% of the way and finish the last 10% yourself. Or hand over the entire task. "I'm in control," Ajenstat says, "or just do all of the work."¹ The result, he says, is a tool that can build a dashboard in two clicks where the old workflow took forty or fifty, and the click count is a proxy for something more important: if it only takes two clicks, you don't need to know as much.¹ (A GeekWire reporter watched him do it with a raw e-commerce dataset on a demo call in April.²)
Why data isn't software
Sangani pushed Ajenstat on a distinction Anthropic had drawn a month or so earlier, in a post with a section headed "Data is not software": in analytics, there is often a single correct answer from a single correct source, and no deterministic way to prove the answer is right.³ A wrong number arrives looking exactly like a right one.
Ajenstat agreed, and walked through the specific engineering problems that distinction creates. Golden can't send a customer's entire dataset, potentially billions or trillions of records, into a language model and still return an answer at interactive speed, so the system reasons over metadata to infer where things are and falls back on SQL, not the LLM, for anything deterministic. "We're actually grounded on SQL," Ajenstat says, "but everything around it, of how you operate with it, is the LLMs."¹
On top of that, no single model handles every step of an analysis well, which turns model selection into an ongoing cost, latency, and accuracy problem.¹ And the system has to add value on data it has never seen before, without six months of upfront curation to lean on.¹
That last constraint sounds like the sharpest version of an objection facing every catalog and governance vendor right now: if the AI can figure your data out on the fly, why invest in describing it at all? Except that isn't where Ajenstat lands.
The context graph problem
The conversation's most technical turn came when Sangani asked Ajenstat about context, a term that increasingly blurs the old boundaries between metadata management, semantics, and governance. Ajenstat's answer was unambiguous, and he offered it unprompted: companies like Alation should be 10x larger, because everybody needs it. He added that he makes this point because he's a believer that more context produces more value.¹ His caveat was about which context a system reaches for, not whether an organization should have any. Knowing which context to call out, and not hallucinating, he says, is still a challenge.
He had the receipts: Golden tried letting its AI freely pull from every connected app to enrich an answer. "It couldn't conclusively connect the dots," Ajenstat says. "It tried. It invented possible correlations."¹ Correlation does not mean causation: the fact that a marketing campaign happened at a given moment doesn't explain a performance degradation elsewhere.
The fix, he says, is being directive rather than exhaustive: pointing the system at a specific source of truth, whether that's a catalog, a semantic model, or a wiki, rather than letting it search everywhere.¹ Golden's own contribution is what Ajenstat calls a context graph built from usage itself, the "exhaust" of every step a user takes inside the product, recorded so the system can recognize a repeated workflow like territory planning and offer to run it again rather than starting from zero.¹
Inferring what matters from what people actually do, rather than from what they get around to documenting, is an approach we've believed in for a long time. Deriving popularity, top users, and lineage from ingested query logs is what Alation's Behavioral Analysis Engine has done from the beginning.⁷
It’s worth being precise about what a context graph is, though, because the vocabulary here is starting to blur. Golden is capturing the path an answer took: which steps a user strung together to resolve a question inside one tool. That's genuinely useful, and it is not lineage. Lineage traces how data moved and transformed between systems, where a number originated, what changed it on the way, and what else depends on it now. Conflating the two is how an organization ends up believing it has provenance when what it actually has is a well-organized record of clicks.
There's evidence for that distinction beyond the definitional one. When Anthropic gave its analytics agent direct access to thousands of prior SQL queries from dashboards, notebooks, and past analyses, accuracy moved by less than a point, even though the correct answer was present in that corpus about 80% of the time. Access to the record of past work wasn't the bottleneck; structure was.⁵ Usage exhaust is raw material for context, which is a real and useful thing to have. But it is not a designated answer, and it is not provenance. Both are worth having. Only one of them tells you why the number is wrong.
Three questions to take back to your own environment
If supervision is the real constraint, the useful exercise isn't evaluating a tool. It's checking whether your organization is currently in a position to supervise one. Three questions worth asking your team this week:
If an AI tool asked your data estate what revenue was last quarter, which source should it read — and does it know that?"
Who is accountable for keeping that source correct six months from now, after the next reorg?
If it returned a confidently wrong answer next Tuesday, how would anyone find out?
Each of those points at a specific capability rather than a posture. The first is certification: a trusted source has to be marked as trusted somewhere a machine can read it, which is the difference between a certified, governed data product published in a marketplace and a table someone shared in Slack.⁸
The second is stewardship, ownership assigned to a named person and visible to everyone who queries the asset rather than held in someone's head. The third is lineage and impact analysis: tracing a wrong number back to where it broke, and forward to every report, model, and agent that consumed it.
It's worth noting how hard the third one is even for the people furthest ahead. Anthropic's mitigation is a provenance footer on every answer, showing which source tier it came from, how fresh the data is, and who owns the model, and it describes the silent wrong answer as the failure mode it hasn't fully solved.⁶ If the company running 95% of its business analytics through agents still names that as the open problem, "we'll notice" is not a plan.
If question two doesn't have a name attached to it, that's the gap to close before the tooling question matters much. Curious to learn more? Book a demo today.
Every quote in this piece is checkable. The full episode transcript, with timestamps, is published alongside the episode.¹
Sources & notes
Every external claim on this page is independently verifiable. The public sources are listed here, including the complete transcript of the episode itself.
All quotations from Francois Ajenstat and Satyen Sangani, and all descriptions of the conversation, come from AI Radicals, Season 4 Episode 3, "Is Business Intelligence Truly Dead?" Full transcript with timestamps. — Alation ↗ https://www.alation.com/podcast/episodes/business-intelligence-dead-francois-ajenstat/
Ajenstat spent 13 years at Tableau, more than seven as chief product officer, through the IPO and the $15.7 billion Salesforce acquisition; later CPO at Amplitude and a venture advisor at NEA. Golden Analytics launched April 2026 with $7 million in seed funding co-led by NEA and Madrona, runs roughly 120 model calls through an orchestration layer, and later raised a $14 million seed extension alongside a public beta. — GeekWire, 7 April 2026 ↗ https://www.geekwire.com/2026/former-tableau-product-chief-launches-golden-analytics-using-ai-to-challenge-the-bi-old-guard/
"Data is not software": in analytics there is often a single correct answer from a single correct source with no deterministic way to prove correctness. On auto-generating a semantic layer, LLM-generated metric definitions encoded existing ambiguities and tested net-negative against a smaller human-curated layer; the recommendation is model-drafted documentation with human-owned definitions. — Anthropic, 3 June 2026 ↗ https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude
Anthropic reports its offline analytics accuracy drifting from roughly 95% at launch to roughly 65% over one month before documentation maintenance was treated as an engineering discipline. Self-reported internal figures. — Anthropic, 3 June 2026 ↗ https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude
Query-corpus ablation: giving the agent direct access to thousands of prior SQL queries moved accuracy by less than one point, with the correct answer present roughly 80% of the time. — Anthropic, 3 June 2026 ↗ https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude
Every response carries a provenance footer naming source tier, data freshness, and model owner; the silent wrong answer is described as the failure mode not yet fully solved. Anthropic also reports that 95% of its business analytics queries are automated via Claude. — Anthropic, 3 June 2026 ↗ https://claude.com/blog/how-anthropic-enables-self-service-data-analytics-with-claude
Alation's Behavioral Analysis Engine derives popularity, top users, and lineage from ingested query logs. — Alation product documentation ↗ https://docs.alation.com/en/latest/datasources/AddDataSources/QueryLogIngestion.html
Certified, governed data products published to a marketplace. — Alation product documentation ↗ https://docs.alation.com/en/latest/DataProductsandMarketplace/ManageMarketplace/ManageMarketplace.html
- AI
- Data Catalog
- Data Intelligence
- Digital Transformation
