Public transcript

Sam Altman and Ali Ghodsi: OpenAI + Databricks, AI Agents in the Enterprise, The future of GPT-OSS

By: Databricks
Length: 23 min
Language: American English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this webinar conversation, Databricks CEO Ali Ghodsi and OpenAI CEO Sam Altman discuss their newly announced partnership to make OpenAI models natively available in Databricks and Agent Bricks. They cover enterprise AI adoption, the growing task horizons of AI agents, governance and responsible AI, the future of open-source models like GPT-OSS, and what enterprise leaders should do now to prepare. Both agree that data foundations, context, and guardrails—not model intelligence—will determine success.

Key points

OpenAI has seen 7x enterprise growth this year and views Databricks as a key partner for bringing models to enterprise data.
Sam Altman measures progress by the 50% success task horizon, which has grown from 5-second tasks at GPT-3.5 launch to 5-hour tasks with GPT-5.
Ali Ghodsi says governance—privacy, audit logging, access control, and guardrails—will be the fundamental limiter of enterprise AI adoption, not intelligence or price.
OpenAI wants to eventually release an open-source model at GPT-5 quality running on a single device, though they don't yet know how to do it.
Ali notes coding is only 20% of an engineer's time at Databricks, and agents will increasingly help with design docs, PRDs, meetings, and other work.
Both CEOs advise leaders to build a good, secure, accessible data foundation so enterprise context can be fed to AI agents.
Sam argues models are smart enough to figure out agentic workflows without hand-engineering, echoing the 'bitter lesson.'
Real use cases include AstraZeneca agents sifting 400,000 documents, finance firms gleaning alpha from SEC filings, and healthcare agents processing hospital paperwork.

Questions it answers

00:00The partnership announcement

Why did Databricks and OpenAI partner?

Overwhelming enterprise customer demand drove the partnership: customers want OpenAI models working natively on their sensitive enterprise data with privacy, auditing, and GDPR rights. OpenAI has seen 7x enterprise growth this year and also uses Databricks internally for data analytics.

OpenAI models are now natively available in Databricks and Agent Bricks.
OpenAI is itself a demanding Databricks customer.
The companies are geographically close, making collaboration easy.

02:00Enterprise transformation timing

Why is now the moment for enterprise AI?

Models have matured enough that enterprises can bring them into their ecosystems with their data and constraints. Sam Altman points to how code writing changed in 2025 and expects a huge transformation of enterprise thinking in 2026–2027, with AI doing an increasing share of intellectual work.

Proprietary enterprise data as context for agents is the big unlock.
2025's coding transformation is a preview of what happens to other functions.

04:00Research frontiers and task horizons

What research advances will drive enterprise adoption?

Beyond smarter models, two axes matter: tight integration with enterprise knowledge and processes, and lengthening the task horizon a model can succeed at. Sam Altman notes the 50% success horizon went from 5-second tasks (GPT-3.5) to 5-hour tasks (GPT-5). Ali adds that Databricks built a genetic-algorithm-inspired technique to automatically optimize enterprise context.

Task horizons have grown from seconds to hours.
Automatic context optimization can lengthen horizons further.
Enterprise tasks often span months or years, so horizons must keep growing.

08:20Governance and responsible AI

How do you ensure responsible AI in the enterprise?

The integration builds in privacy and security from the ground up: audit logging of everything, access control, and guardrails like staying on brand and not recommending competitors. Sam believes governance, not intelligence or price, will be the fundamental limiter of enterprise AI adoption.

Using OpenAI models inside Databricks gives these guardrails out of the box.
Governance is the 'limiting reagent' for adoption.

09:40Open source and GPT-OSS

What is the future of open-source models from OpenAI?

GPT-OSS is not the last. OpenAI hopes to someday release a GPT-5-quality open-weights model that runs on a single device, though they don't know how yet. Sam argues privacy and freedom will make local models important, and envisions a new AI-era computer that runs a great model locally.

Demand for open models is smaller than for frontier cloud models but intense.
A 120-billion-parameter GPT-OSS already exceeded expectations.
Local models matter for privacy and when connectivity fails.

12:00Agents as co-workers

What are we underestimating about agents long-term?

Agents will become co-workers involved end to end in productive work. Ali notes coding is only 20% of an engineer's time at Databricks—the rest is design docs, PRDs, and meetings where agents can also help, and their involvement would improve the coding itself through added context.

Agents will participate in the whole work process, not just code generation.
Adoption will be gradual as task horizons lengthen.

15:50Use cases today

What enterprise AI use cases are already working?

AstraZeneca used agents to sift 400,000 documents; financial firms extract alpha from SEC filings; insurance underwriting and healthcare document processing are being transformed. Sam adds the 'wouldn't do' category—cheap agents make previously abandoned ideas worth trying, like kicking off one extra agent for a feature no one would ever build.

Superhuman literature and document search is already taken for granted.
Low activation energy changes what work people attempt.
Latency, not just cost per token, is an underinvested industry area.

19:50Advice for leaders

What should a CEO do right now to prepare?

Start with the boring stuff: build a good data foundation, secure it, add guardrails, and break down data silos so all enterprise data is accessible as context. Definitions like churn, revenue, and fiscal quarters exist somewhere in the organization and must be connected to the AI.

Data accessibility determines whether the AI gets tasks right immediately.
Models are smart enough—they need the right context fed to them.

Notes

Partnership Background

Databricks and OpenAI announced a first-of-a-kind partnership to make OpenAI models natively available within Databricks and Agent Bricks.
Ali Ghodsi says customer demand was overwhelming: enterprises want OpenAI models working on their sensitive data with privacy, auditing, and GDPR rights.
OpenAI has had 7x enterprise growth this year and is also a Databricks customer for data analytics. Ali calls OpenAI one of Databricks' most demanding customers, in a good way.

Enterprise AI and Task Horizons

Sam Altman frames progress via the 50% success task horizon: from 5-second tasks at GPT-3.5 launch, to 5-minute tasks with GPT-4 iterations, to 5-hour tasks with GPT-5.
Two key axes: tight integration into enterprise knowledge, data sources, and business processes; and lengthening the horizon to months- or years-long tasks across more verticals.
Ali describes a Databricks technique inspired by genetic algorithms that automatically optimizes enterprise context for the model.

Governance

Ali emphasizes privacy, security, audit logging, access control, and guardrails (e.g., staying on brand, not recommending competitors) built in from the ground up.
Sam believes governance will be the fundamental limiter of enterprise AI adoption—not intelligence or price.

Open Source

GPT-OSS has huge demand, especially as an American frontier open-source model. Sam says GPT-OSS is not the last and OpenAI hopes to someday ship a GPT-5-quality open model on one device.
He cites privacy and freedom as key principles for local models.

Use Cases and Advice

Examples: AstraZeneca agents sifting 400,000 documents, finance firms extracting alpha from SEC filings, insurance underwriting, and healthcare document processing.
Sam highlights the 'wouldn't do' category—cheap, low-latency agents make previously abandoned ideas viable.
Advice for leaders: build a solid, secure, accessible data foundation and connect enterprise definitions (churn, revenue, fiscal year) as context.
Both endorse letting models figure out agentic workflows rather than hand-engineering—the 'bitter lesson.'

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.