Agentic AI for Database Management Systems: Write SQL using natural language and Deep research

30+ million
in ARR in the 1st half of 2026.
70+
logos purchased before Q4 2026 (GA Q1 2026)
600+
Active users in a day*
12k~
Peak user events in a day*

*Reach out for more details

IBM Db2 Genius Hub is an AI-powered database management system for IBM Db2 autonomous databases. Enterprises can now manage (Db2) databases with AI-driven automation, faster problem resolution and proactive performance optimization.

What is Db2 Genius hub?

Absent of a precedent in IBM Software’s enterprise portfolio, I defined how Deep research and Text2SQL should behave — and what would earn a Database Administrator’s trust: output they can verify, recognize, and control.

Design overview

From the definition of Design principles from UX Research, working directly with backed and front engineering leads, logo design, prototyping, mid & high fidelity design, I am the product designer of Deep Research and Text2SQL.

Role, Scope
The Improve panel in the SQL workbench, showing a side-by-side diff of a diagnostic query — 89 additions and 88 removals across 835 lines — with Reject and Accept controls above it.

Client value

[1]

When a high severity production anomaly needs diagnosing, proactive intelligence can turn a time-sensitive fix into a weeks-long investigation.

Weeks to hours anomaly resolution

[2]

The skill required to troubleshoot is becoming rare. An Intermediate DBA is scarce. In practice, complex scenarios escalate this way: Customers send a support ticket → Customers wait for days → IBM’s internal Db2 Technical Staff provide an advisory and/or remediation plan to the customer. When the problem persists, clients cycle through this process again.

Scalable resolution path

[3]

Db2 is a niche, specialized IT product with a gigantic learning curve. Deep research bridges the knowledge and expertise gap of a Junior Database Analyst from a Senior.

Sustainable talent pipeline and skill enhancement

Design Principles

Database Administrators adopt AI they can verify, recognize, and control.

[1]

Agentic AI is goal-directed, not task-directed. Operational transparency is crucial to AI trustworthiness.

How do we mitigate misalignment in procedural logic?

How do we surface traceability early and empower customers to press the eject button when necessary?

[2]

AI-native, autonomous databases are the future. How do we encourage and normalize the use of AI in their day-to-day workflow?

Given day-to-day use and ingrained mental models of traditional console UI, how do we meet existing users where they are?

[3]

Progressive disclosure is essential to AI usability. UXR insights stressed an appropriate pacing of information while also providing users with the ability to progressively control the flow of information.

How do we prevent an LLM chat dump?

UX Research was performed at IDUG EMEA Düsseldorf, Germany late 2025. We worked with 10+ Database Administrators to test a concept end-to-end design prototype. Phase 1 design considers what is technically feasible at the time.

Solution

[1]

Available in the AI chat full page workspace and the side panel experience, Text2SQL and Deep research is invoked from the Tools menu in the Prompt bar. Users are endowed with monthly Deep research limits to manage token usage traffic and consumption.

The Tools menu open above the prompt bar, offering Deep research with its remaining request count and Generate query for writing SQL in natural language, with the connected database named beside the Tools button.
Previously, Agentic skills could only be invoked through Agent mode. Soon, Engineering enhanced the baseline Database Assistant architecture to inherit it, with the exception of Text2SQL.
[2]

Deep research affords many opportunities: ease of client migration to Db2, the bridge to close the immense expertise between Junior and Senior Database Administrator, and the improved digestibility of extensive product documentation.

The finished HADR setup report open on its Answer tab, beside the chat panel where the request was made — the Deep research card reporting it completed in 4 minutes 16 seconds.
At the time of the technology, Deep research could take from a couple of seconds to potentially a couple minutes to generate a report. Engineering managed to shrink it to a reasonable maximum of a few minutes right before the release — removing the need for notifications.
[2.1]

Text2SQL reimagines how DBAs and analysts interact with complex Db2 databases, replacing manual SQL authorship with a natural language interface that autonomously explores schemas, generates SQL, and surfaces its reasoning at every step.

The Genius Hub SQL workbench with a generated World Cup scores query in the editor, and the AI side panel alongside explaining the query it wrote and offering to recalculate it differently.
[2.2]

The to-do list illustrates the breakdown and real-time progression of a complex task. Operational transparency is crucial to AI trustworthiness. Inspired by an existing AI chat plan component, this Plan componentry captures the progression of the tasks granularly, with the generated query ready to be opened.

The agent to-do card “Query World Cup scores — 2 of 3 steps completed” above the Plan component, with the generated World Cup scores query beside it and an “Open in SQL workbench” tooltip.
[4]

I lent my expertise in Text2SQL and Human-in-the-loop UI/UX to an emerging design system workgroup called Ambient Intelligence and the watsonx.data product portfolio.

A SQL script in the editor with an AI-proposed index maintenance change shown as a red and green diff, and inline Approve, Reject, Explain and Refine controls waiting for confirmation.
Patterns carry cross-portfolio and cross-organization.

Conversational design

Keep conversations natural: Handling out of scope requests

Educate users on out of scope queries. Inform next steps and suggest concrete examples to prevent dead-end experiences.

Three Database Assistant conversations side by side: the what's new and best practices greeting, a refusal that lists the queries Deep research can answer instead, and an out-of-scope weather question answered with the assistant's actual skills and documentation links.

Handover escalation and error states

Sometimes, not all situations can be handled by Agents. An effective handover escalation method is to escalate it to the right person or role.

Two Database Assistant conversations where a troubleshooting task is cancelled: one ends in a connection error telling the user to contact an administrator, the other in a system error after the agent could not resolve the issue.

Early explorations

Cursor, Claude Code, and other IDEs were largely studied for Agentic developer use cases. First to release Agentic experiences for database management use cases and beyond, Palantir AIP, Supabase and Databricks were studied.

Text2SQL could be leveraged at any point in Database Administrators’ workflow - whether creating code from scratch or editing existing code. Considering accessibility, large visual transformations, and strained engineering resources at the time, Text2SQL Phase 1 is primarily an AI side panel chat experience.

Two SQL editor explorations side by side: on the left a Database Assistant prompt opened in line over a selected subquery, on the right an inline AI edit reading “Change this query to find purchases from the last 90 days instead of 30 days” with Reject, Edit in-line and Accept controls above a red and green diff.
I enhanced existing patterns provided by the AI Design system design team. The original patterns served less complex AI chatbot use cases relative to Agentic IDEs.
An empty Genius Hub worksheet in dark mode with a natural-language request for a diagnostic query above the editor, generating its response.

What didn’t ship

Design shifted through many iterations: constrained by rigid architectural limitations, expanded by eureka-moment architectural possibilities, and guided by ever-changing highly experimental comparative Agentic UI products in the market.

Inversely, a user could pass context from the SQL Worksheet to AI chat side panel. Users could work from the file natively independent from the AI chat side panel. Agentic IDEs like Cursor are known for this ease of use. However, it was not architecturally possible at the time to pass context from front-end to the Agent.

A dark SQL worksheet with a block of the query selected and a small Improve / Explain menu floating beside it.

As emphasized by UX research findings, minimal, necessary information and visual componentry is highly preferred by Database Administrators. While these UI patterns were present in Agentic IDEs at the time, these patterns occupied more space that I had envisioned. For example, while Agentic IDEs could set font sizes as small as 12, body font sizes cannot be less than 14 for Accessibility reasons.

The workbench beside a Database Assistant panel, which reads the open query, explains how it is modifying it, and offers the edited version for approval.

In a scenario where in-line editing directly from the SQL Worksheet Editor was not possible, the selection of the ‘Generate with AI’ button on the SQL Worksheet toolbar could take users to a modal workspace. This added layer was discarded due to redundancy.

A full-screen Generate dialog for writing or editing a SQL query with Database Assistant, with Reject, Accept and edit, and Accept and apply along the bottom.

This effort is not short of pivots and experimental concepts. Reach out to learn more.

Impact

[1]

Set a higher bar for Agentic design as a design org AI SME

Design teams across IBM came to me for competitive analysis and direction on in-flight work. Through 1:1 advisory and cross-portfolio shareouts, I shaped AI product decisions beyond my own organization (Core software) and a sister design org (Platform software).

[2]

I owned a 0→1 project and became Product and Engineering’s peer on AI product decisions.

There was no playbook for Agentic AI UI/UX at IBM — product teams had only ever built authoring experiences. I navigate ambiguity and lead design without waiting for requirements, co-creating direction with PM and customer feedback. With my growing domain knowledge in Agentic AI and AI for design, I confidently argued trade-offs with Engineering, and shaped what shipped.