# Closing the AI speed divide with Genie Agents on Databricks

> When a decision takes three weeks and a customer's agent takes minutes, the money leaves first. What one Databricks platform and Genie Agents change.

Published: 2026-10-31
Author: Preetham Reddy
Tags: Databricks, Applied AI
Canonical: https://www.techfabric.com/blog/closing-the-ai-speed-divide-with-genie-agents-on-databricks

---

I recently went to the [Databricks Data + AI World Tour](/events/databricks-data-ai-world-tour) in Chicago, where TechFabric was a sponsor, and attended Jennifer Miller's keynote on the strategic shift that AI enables in business. Her point was that companies should stop measuring AI by adoption and start measuring it by speed, and these are my notes from the session, alongside what our engineers have learned building on Databricks with more than fifty agents in [Fabric](/case-studies/fabric-pro).

![Two cards: Adoption struck through, with an arrow pointing to Speed.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/01-adoption-to-speed.png "Miller's slide set deployment and usage metrics against velocity that reaches competitive impact.")

Miller runs banking and payments go-to-market for Databricks and shared the welcome keynote with Spencer Cook, but her argument fits any business that loses money while a report sits in a queue. She measures AI by the time a company takes to get from the first sign of trouble to a decision. In her demo, a bank took three weeks to answer a deposit outflow, and $94M in deposits left while it waited. On Databricks, she got to the bottom of the same problem in five minutes. We think that comes from putting all of a company's data on one platform, with the lakehouse for its history, Lakebase for its live records and Genie Agents to do the work.

## Where most firms sit on the five stages of AI

Miller's slide laid out five stages of AI and showed most firms sitting in the first two: Experiment, where teams run isolated proofs of concept, and Sprawl, where uncoordinated tools pile up in silos. The next three stages move from targeted workflow gains in Optimize to changing a core process in Redesign and building the business model on AI from the start in Reinvent.

![Five stages from Experiment to Reinvent, with most firms marked at the first two.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/02-five-stages.png "Most firms sit at Experiment and Sprawl, and the speed divide opens on the way to Redesign.")

A bank can hand every branch an AI assistant and still sit at stage two, because the weekly report, the tickets and the executive deck all still run. To us, adopting AI completely means reaching Redesign and Reinvent, where an agent does in minutes what took a department weeks and work that nobody could justify before becomes ordinary. The gap between the companies that get there and the ones still at stage two is the speed divide.

## A deposit outflow answered in five minutes

Miller's demo used a sample lender, Crestline National Bank, whose head of consumer deposits says, "Our deposits are leaving faster than our forecast, and I don't know why." The bank learned about the drop from a weekly report, weeks after the money had started to leave. Analysts then spent days on tickets and manual extracts to confirm it was real, and more days reconciling spreadsheets by hand against competitor rates to find the cause. An executive review took another 5 to 10 days to settle on a response. The whole process took three weeks, and $94M in deposits left the bank while it waited.

![Four cards for detect, confirm, diagnose and decide, each with its delay, above a band reading $94M in deposits gone.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/03-four-steps-three-weeks-v2.png "In Miller's demo, detect took weeks, confirm and diagnose took days each, and decide took 5 to 10 days.")

On Databricks, the same four steps took five minutes, because nobody had to wait for a report, a ticket or a spreadsheet. A Genie task caught the drop and sent an email: balances were $47.0M below forecast over three days, driven by a competitor's rate promotion in Ohio, with affluent customers in Columbus, Cleveland and Cincinnati leading the outflow.

![Demo alert showing $47.0M below forecast, with bars for Columbus, Cleveland and Cincinnati Affluent.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/05-genie-task-email.png "The demo's alert named Columbus Affluent at $11.1M, Cleveland Affluent at $7.6M and Cincinnati Affluent at $5.5M.")

In chat, Genie took the question of which Ohio deposits were a flight risk and which retention moves to deploy, and it searched the Genie Ontology for the bank's own definition of flight risk.

![A question about Ohio flight-risk deposits, with the three steps Genie took to gather context.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/06-genie-chat.png "Genie read the past conversation and searched the Genie Ontology before it answered the flight-risk question.")

Her slide showed the signal at 7:00 and the decision by 7:05.

![Four steps on a line from 7:00 to 7:05, above a long three-week bar and a thin five-minute bar.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/04-five-minutes-v2.png "The same four steps, five minutes from signal to decision.")

McKinsey's estimate for real banks points the same way. [Bank profit pools could fall by $170 billion, about 9 percent, over a decade or so](https://www.mckinsey.com/industries/financial-services/our-insights/global-banking-annual-review-2025) if banks don't change, largely because customers' own AI agents could move money out of near-zero checking accounts. Investors reacted in the days after Meta launched its Muse agent in September, when [Meta added about $200 billion of market value in one session](https://finance.yahoo.com/markets/stocks/articles/meta-added-200-billion-one-181131033.html) and [Wells Fargo fell 3.9% and JPMorgan 3.4%](https://finance.yahoo.com/technology/ai/articles/metas-muse-just-put-banks-170657037.html) as analysts at BofA Securities described an agent that spots surplus cash, compares deposit yields and might act for the customer. A depositor's agent can compare rates every hour of the day, so a three-week answer arrives after the money has gone.

![Meta up $200B beside falling bars for LPL, Schwab, Wells Fargo, JPMorgan, Morgan Stanley and the KBW Bank Index.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/07-market-reaction.png "Miller's slide showed LPL down 7.5%, Schwab 6.1%, Wells Fargo 3.9%, JPMorgan 3.4%, Morgan Stanley 2.9% and the KBW Bank Index 2.4%.")

## What all your data on one platform makes possible

The demo worked because one agent could reach deposits, forecasts, competitor rates and customer segments in one place. Most banks hold those in separate systems, which is why the tickets and the spreadsheets exist in the first place.

The lakehouse is where that changes, and getting there takes less time than most leaders expect: we [moved a manufacturer off Azure Synapse onto Databricks in six weeks](/blog/synapse-snowflake-databricks-migration-airlift), when a move like that normally takes six months or more.

Lakebase adds a managed Postgres database for operational and transactional workflows. [Synced tables](https://docs.databricks.com/gcp/en/oltp/projects/sync-tables) create a managed copy of Unity Catalog data in Lakebase for serving applications, and it can hold [an agent's state and memory](https://docs.databricks.com/aws/en/oltp/projects/state-management). We moved [TotalLoop](/case-studies/dealeriq) off Cosmos DB and SQL Server onto Lakebase, which removed the separate copy that had to be kept in step with the analytics. We also [branched the database behind a live demo app in 4.5 seconds](/blog/lakebase-database-branching-coding-agents), which gives an agent its own copy to experiment on.

Genie reads what those two hold. [Genie Ontology](https://www.databricks.com/blog/introducing-genie-one-genie-ontology-and-genie-agents) builds a graph of how a company works from its tables, queries, dashboards and pipelines. Genie Agents, the new name for Genie Spaces, are specialists built on that context, and they can read documents as well as tables. A Genie Agent can now [take actions in external tools through MCP connectors](https://docs.databricks.com/aws/en/ai-bi/release-notes/2026), in Beta.

![A table of the lakehouse, Lakebase, Unity Catalog and Genie Agents, with what each holds and our work on it.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/08-what-sits-where-v2.png "Each layer of the platform, beside what our engineers have already done on it.")

Put together, every system lands in the lakehouse, Lakebase holds the live records and Unity Catalog governs the data in both, so Genie Ontology builds the context from governed sources. A Genie Agent then finds and explains the signal, a person decides, and a governed action runs and writes its result back to Lakebase.

```mermaid
%% caption: One governed platform gives a Genie Agent the data and context to take a signal to a person's decision in minutes.
flowchart TD
  A["Every system lands in the lakehouse"] --> B["Unity Catalog governs the data"]
  C["Lakebase holds the live records"] --> B
  B --> D["Genie Ontology builds the context"]
  D --> E["Genie Agent finds and explains the signal"]
  E --> F["A person decides"]
  F --> G["A governed action runs"]
  G --> C
  class E accent
```

## Use cases that were never worth staffing

Most businesses decide in batches because a person's time costs more than one small decision is worth. A bank reviews segments every week since it can't assign an analyst to each account every morning. With the data in one place and a Genie Agent reading it, the unit of decision can shrink to the single account, file, machine or product.

A lender could have an agent read every open refinance file each morning against that day's rate sheets and move the files that now clear a lender's rules to the top of a processor's queue. We've seen what clearing the manual layer does in this business. [Auto Approve's](/case-studies/auto-approve) agents copied each request by hand between the CRM, lender sites and Excel, which took two to three days, and after we built the approval engine on Azure the company doubled its monthly loans without hiring more agents. A Genie Agent would start from that kind of cleared path and keep looking for the next file.

A payments team could have an agent watch disputes merchant by merchant, explain why one merchant's chargebacks jumped overnight, and draft the note to the account manager before the weekly review exists. A manufacturer could join line sensors, maintenance history and Business Central purchase orders, then have an agent flag a line whose load has been climbing, confirm the spare part is in stock and ask a planner to approve the stoppage. A retailer could compare stock in every store with local demand each morning and propose a price or a transfer for each product, a call that today gets made by category.

![A table of what a Genie Agent would read and do for a lender, payments team, manufacturer and retailer.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/09-use-cases.png "Each scenario shrinks the decision to one file, merchant, line or product.")

Databricks [prices Genie by usage, with no seats](https://www.databricks.com/blog/introducing-genie-one-genie-ontology-and-genie-agents), so that retailer would pay for the morning's runs across its products. A license per buyer would never have covered that work, because no team of buyers could have done it.

## Why speed wears down a large company's advantage

One of Miller's slides listed what AI will disrupt, predicting that 90% of the technology doesn't exist yet and that monopolies will end, and we share that view. Genie Agents gained the ability to act in external tools in Beta in October, so a company with its data already in Unity Catalog could try that within weeks of its release, while a company at stage two still has to find its data first.

![Five rows of what AI will disrupt, ending with 90% doesn't exist yet and monopolies will end.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/10-what-ai-disrupts.png "Miller's other rows said no segment is immune, every job gets redesigned and workflows get reimagined for agents.")

Size protected incumbents because only a large company could afford the people to process that many decisions. We watched this shift in our own business. Fabric's agents hold what our strongest engineers used to keep in their heads, and across our clients, delivery work that took ten people now takes three.

McKinsey titled its 2025 banking review [Why precision, not heft, defines the future of banking](https://www.mckinsey.com/industries/financial-services/our-insights/global-banking-annual-review-2025), and it says AI pioneers could raise their return on tangible equity by up to four percentage points. A challenger with forty staff and all its data in one catalog could check every account every morning. We expect the companies that move this way to take customers from much larger ones, in banking first and then in every business where a slow decision costs money.

## What Genie Agents need before they can move this fast

Miller's slide on why firms are slow listed silos, missing context and approval gates, and it paired them with one lakehouse, Genie and Unity Catalog. We'd add what our own builds taught us about the last two.

![Three problems, silos, no context and gates, each paired with its Databricks approach.](/blog-media/closing-the-ai-speed-divide-with-genie-agents-on-databricks/11-why-firms-are-slow.png "Miller's three reasons firms are slow, each paired with the Databricks approach to it.")

Context has to be agreed by the people who own the number. Her slide said Genie encodes business definitions so agents hit 95%, not 50%. Databricks' [own June benchmark](https://www.databricks.com/blog/introducing-genie-one-genie-ontology-and-genie-agents) measured 84.5% on the first attempt across 28 questions. We'd treat either figure as something to test on your own questions. Our [Genie Accuracy](/databricks/genie-accuracy) work starts by sitting with the people who know the right answers, writing down the questions they ask and running those questions as a regression suite, because two teams with different definitions of active customer will each get an answer the other calls wrong.

Controls have to live inside the action. In [AutoFabric](/case-studies/autofabric), our auto lending platform on Databricks, each step carries its permissions and policies in its definition, and [no agent can call a step that has none](/blog/ai-governance-for-lending-and-payments-agents-on-databricks). An approval queue delays every call, and a policy in the action holds only the calls that break it, so a regulated company keeps its record and gets the speed. One customer confirmed a 40% efficiency gain from AutoFabric as a whole. We made a similar call on Fabric, keeping its context store inside Unity Catalog though a separate store would have been quicker to build, since a second store means copied data and two sets of permissions.

We'd plan around two things in Genie today. [Scheduled tasks run at a set time](https://docs.databricks.com/aws/en/genie-one/chat), so we'd use one for a 7:00 email and build a pipeline and an alert for an hourly alarm. Agent actions in external tools are in Beta, and we'd keep a person in front of anything that moves money.

Pick the slowest decision that costs your business money and time it from signal to action, using the dates on real tickets. Then build one Genie Agent against that clock, with your data under Unity Catalog and a benchmark set written by the people who own the number. Our [AI Readiness Assessment](/ai/readiness) is where we'd name that first decision with you.

## Frequently asked questions

### What are Genie Agents in Databricks?

Genie Agents are specialist agents built on your governed data and documents, and the name replaced Genie Spaces this year. Genie One is the general coworker for business users, while a Genie Agent is scoped to one domain and shared with a team. We'd build the first one around a single business question with an owner.

### Does speed matter more than AI adoption?

Yes, once a company has the tools in place. Adoption shows who can use AI, and speed shows whether a decision changed. We'd track time from signal to action on one process and leave seat counts to finance.

### How can a bank spot deposit outflows sooner with Genie?

Land daily net flows by segment under Unity Catalog beside the forecast, then have a scheduled Genie task email a breach. Competitor rates are the piece most banks don't hold, so we'd add a rate feed first. A signal faster than daily needs a streaming pipeline.

### Do we need Lakebase to use Genie Agents?

No. Genie Agents run on data governed in Unity Catalog, and Databricks' announcement doesn't name Lakebase as a source for them. We'd add Lakebase when an application or an agent has to write live records, then sync those records to the lakehouse so Genie can read them.