*Databricks’ own published claims about its own organisation; not independently verified. The 85%, 800+ monthly questions, and 5,000+ total answers are floors. Marge is the marketing organisation’s deployment; the anecdote below concerns sales and finance — no connection between the two instances is established.
On the a16z podcast episode Databricks CEO: Stop Scaring People About AI, Databricks co-founder and chief executive Ali Ghodsi recounted a moment that is easy to read as a minor frustration and considerably more interesting once you look at its structure. He needed his company’s Fortune 500 penetration figure for a presentation. He asked someone in sales operations. Her reply was that she could not log in to Genie at that moment because she was on a flight. Ghodsi’s account of his reaction: “if you’re just gonna log into Genie, I can do that myself. Like I asked you because I thought you had like something authoritative that I don’t have access to.”
He then asked the chief financial officer. The CFO, in Ghodsi’s telling, replied by copying and pasting a screenshot of Genie. His response: “And he just copy pasted a screenshot of Genie back… So I said, does anyone do anything novel here?” This is Ghodsi’s account of what happened and is not independently established here. Neither person beyond Ghodsi is named. Nothing in what follows claims either person did anything wrong, failed at anything, or was underperforming.
Separately, and independently of that anecdote, Databricks published figures on September 15, 2026 describing its marketing organisation’s use of an internal assistant called Marge, built on Genie Agents. According to those published figures — Databricks’ own claims about its own organisation, not independently verified — marketers use data three times more often to make decisions; more than 85% of the marketing organisation uses Marge; usage has grown 50% quarter over quarter; Marge handles more than 800 questions each month and has answered more than 5,000 in total; and flagged incorrect responses have decreased by 25% as the system has improved. Marge is the marketing organisation’s deployment and is not established to be the same instance described in the anecdote, which concerns sales and finance.
The key insight: Ghodsi’s own sentence identifies what he expected to receive. He did not ask because he wanted the number — he asked because he thought the person had something authoritative that he did not have access to. Authority and retrieval were bundled together, and self-serve access to the system is what unbundles them. That unbundling is the structural event the anecdote describes.
A figure saying almost everybody now queries the data directly is also a figure saying the intermediating step has been removed.
The Structural Read
When access to a data system was scarce — when not everyone could log in, run the query, and read the output — whoever held that access became the interface between the system and everyone who needed an answer. The authority and the retrieval arrived in the same package. Answering the question was the job. Nobody needed to separate the two components because they had never been separated.
Self-serve retrieval breaks that bundle at the structural level, not at the level of any individual. When the CEO can run the query himself, privileged access stops being a scarce asset. The package comes apart. What remains — the part that does not get replaced by the query — has to be something other than retrieval: reconciliation between sources that disagree with each other, knowing which definition of a metric a particular board presentation actually requires, and the standing to say that a number looks wrong before it travels further. None of that is novel in the way Ghodsi’s question asks about novelty, which is precisely the problem with the question.
The Authority-Retrieval Package
Scarcity of access is what kept authority and retrieval bundled
When system access was the scarce resource, whoever held it was structurally the answer to every question about that system. Democratising access does not make the answer easier to produce — it removes the reason the question was routed through a person at all. What remains in the role after retrieval is automated is not nothing; it is exactly the work that never looked like work because it was always wrapped around the retrieval step.
The adoption figure from Marge is the same event counted from the other end, and this is worth sitting with. A metric reporting that more than 85% of the marketing organisation now queries data directly — that usage is growing 50% quarter over quarter, that more than 800 questions arrive each month — is also, read carefully, a metric reporting that the intermediating step has been removed at scale. From the vendor’s perspective, that reads as adoption. Experienced from inside the organisation, it is the moment when colleagues begin forwarding one another the same output from the same tool, which is the precise scene the anecdote describes. This is a reading of what a high adoption figure describes, not a criticism of the product, the deployment, or any team.
The question at the end of the anecdote — does anyone do anything novel here? — is both correct and limited, and both halves deserve stating. It is correct because it asks what is being added on top of a capability that everyone in the organisation now shares, which is exactly the question any organisation should ask when a tool removes a step. It is limited because a novelty frame cannot see the work that consists of checking, or of knowing which of several defensible versions of a number the moment calls for, or of being the person willing to say an answer looks wrong before it reaches a presentation. None of that work is novel. All of it is load-bearing. This is a general observation about how the question is framed — not a correction of Ghodsi, not a claim that he is mistaken, and no view is attributed to him beyond the two quotations above.
The 25% reduction in flagged incorrect responses is the number in the published set that is easiest to move past and the most structurally revealing. A flagged incorrect response is a person reading an output and deciding it was wrong. The improvement in the system is being measured through human judgement exercised on that output. The same capability that, when it is merely forwarded, generates the anecdote’s frustration is the same capability that, when someone reads it critically, generates the signal by which the tool gets better. No claim is made here about any error rate, accuracy level, or reliability for any system; the piece does not state how often anything is wrong; and nothing here asserts that the flagging is complete, representative, or correct.
Three Implications
AUTHORITY VERSUS RETRIEVAL — WHERE VALUE SITS IN A ROLE
When a role was built around privileged access to a system, authority and retrieval were structurally inseparable. Removing the access barrier does not remove the value — it relocates it. What was previously wrapped around retrieval (reconciliation, definition-choice, the credibility to flag a wrong number) now has to stand on its own. Organisations that recognise this distinction will ask different questions about how their data functions are structured than organisations that only read the adoption metric as a success figure.
ADOPTION AS DISINTERMEDIATION — COUNTED FROM THE OTHER END
A high adoption figure for a self-serve internal tool is simultaneously a description of how many intermediating steps have been removed. Databricks’ own published figures — not independently verified — report more than 85% of its marketing organisation querying data directly, with usage growing at 50% quarter over quarter. The internal experience of that curve is colleagues forwarding one another identical outputs, which is the scene the anecdote captures. Vendors and organisations are reading the same metric from opposite directions, and both readings are accurate.
WHAT A NOVELTY FRAME CANNOT SEE
The question does anyone do anything novel here? is structurally blind to the work of checking, definition-choosing, and error-flagging — because none of that work is novel. Yet the 25% reduction in Marge’s flagged incorrect responses (Databricks’ own published claim; not independently verified; no error rate or accuracy level claimed here) is a published measurement of exactly that checking capacity turned into a system-improvement signal. The least visible work in the organisation is currently the instrumentation by which the tool gets better.
Business Engineer Framework
The Map of AI — Where Authority Sits in the Stack
The Databricks anecdote is a precise illustration of a pattern the Map of AI traces across the full stack: when a capability layer becomes broadly accessible, the value does not disappear — it migrates up to whoever can act on the output rather than merely produce it. Understanding which layer you are operating in, and what the access economics of that layer are, is the foundational question for any organisation deploying internal AI tools at scale.
Ghodsi’s sentence — I asked you because I thought you had something authoritative that I don’t have access to — is the cleanest description available of what self-serve retrieval actually removes, and it comes from the person whose tool removed it. The authority was never in the number; it was in the scarcity of the access. Once that scarcity is gone, the remaining value in a data function is reconciliation, definitional judgment, and the willingness to say a number looks wrong — none of it novel, all of it real, and none of it visible to a question framed around novelty. That is the structural shift the anecdote describes, and the adoption figures confirm it from the other direction.
This is businessanalysis. It is not investment advice, no view is expressed on any security, and no recommendation is made.
The anecdote above is Ali Ghodsi’s own account, quoted from the episode, and is not independently established here. No individual other than Ghodsi is named, neither colleague is described beyond their role, and nothing above claims that either did anything wrong, failed at anything or was underperforming. The argument concerns where value sits within a role. Nothing above claims that any role is redundant, makes any claim about headcount, employment or staffing, or suggests anyone’s job is at risk. Databricks’s figures are its own published claims about its own marketing organisation’s deployment, an internal assistant called Marge built on Genie Agents, and are not independently verified. The stated share of the marketing organisation using it, the monthly question volume and the cumulative total are all floors rather than exact values. That marketing deployment is not established to be the same instance described in the anecdote, which concerns sales and finance; no figure from one is applied to the other, and no connection is claimed beyond both being Databricks’s own AI analytics. Nothing above claims any error rate, accuracy level or reliability for any system, says how often anything is wrong, or claims that response flagging is complete, representative or correct. No Fortune 500 penetration figure, revenue, customer count, headcount, price, competitorproduct or figure, or time or cost saving appears above, and nothing is predicted. This is businessanalysis. It is not investment advice, no view is expressed on any security, and no recommendation is made.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
