OpenAI just launched a series of videos and blog posts on how OpenAI uses AI at their work. https://openai.com/index/building-openai-with-openai/
This led to some stock market impact for companies like Hubspot, Salesforce and Klaviyo. Investors and the rest of us are asking the question.
Read OpenAI SaaS attack has begun
Will Agentic AI disrupt SaaS
Building OpenAI with OpenAI
“Does AI Native SaaS look fundamentally different from traditional SaaS with their AI bolt ons?”
The blog series could be an important take on this thesis. I read all of the articles so you don’t have to - let’s dive in.
They talk about 5 types of Agents in their blog
The Internal Agent crew at OpenAI
Category - Customer Support
Problem - Deals with answering questions from millions of users.
Impact - Support reps evolved from ticket processors to system builders who flag patterns, ship classifiers, and prototype automations
Tech Stack - Agents SDK, Responses API, Realtime API, Evals dashboard
Category - Business Intelligence
Problem - Millions of support tickets buried with valuable information
Solution- Combined dashboards with conversational GPT-5 interface for plain-language queries and instant reports
Tailor - Inbound Sales Agent
Category - Sales Enablement
Problem - Small companies lost in queues for contract negotiations. Lead forms are automatically worked on by the Agent and emails are sent to pitch different products.
Impact - Multi million ARR unlocked
DocuGPT - Contract Processing Agent
Category - Document Intelligence
Problem - Huge volume of contracts requiring data henry
Impact - RAG based agent automated data extraction of different contract terms supported 100X scale with same headcount
GTM Assistant - GTM Meeting Assistant
Category - Meeting Prep Q&A
Problem - Use expert knowledge to do research for meetings and provide answers
Impact - Slack assistant supports Q&A and reviews
Agents deployed in different form factors
I think the most interesting thing about these posts are that each team has created a very different way of interacting with the Agent. The GTM Agent is on slack and the Contract Processing Agent looks like it is actually run in databricks.
When messy workflows meet real world data and get applied to different team settings you can see very different implementations of solutions.
Revenue - Account Management
-Internal Dashboard
GTM Assistant - Inbound Sales Agent
-Slack
Support - Support Agent
-Customer Facing Chatbot
Research - Research Agent
-Internal Dashboard
Legal - Document Extraction Agent
-Databricks Query
So the method of AI delivery seems to be changing with the workflow, the impact and level of sophistication of the user.
SaaS has spent 20 years locking users into the same interface and siloed data layer. Whereas the “One Interface for One workflow” doesn’t seem to apply to OpenAI - the interfaces are fragmented and driven by workflow context.
Each team is building an Agent and acting as a system architect.
What this means for the “AI Native vs Bolt-On Debate”
There is an Agent builder shift - Traditional SaaS sells fixed workflows. AI Native SaaS provides primitives for users to build their own agents. OpenAI is mentioning their own Agents SDK and evals so that each team has its own construction kits.
AI tools will disappear into existing workflows. Examples from OpenAI.
Key Changes
-Legal doesn’t want AI legal tool - they want DocuGPT inside databricjs
-Sales doesn’t want another dashboard - they want slack where deals happen
-Customers doen’t want a new interface for support tickets 0 they want a simple chatbot
UI Habituation will give way for Workflow Invisibility
What This Means for the "AI Native vs. Bolt-On" Debate
The uncomfortable truth: You can't achieve workflow invisibility with a bolt-on. Incumbents like Salesforce and HubSpot face structural disadvantages. 20 years of interface optimization locks them into single UI paradigms Siloed data architectures can't support multi-agent workflows. "Copilot" features still trap users inside their existing interface
AI Native advantage: Meet users where they are, not where you want them to be.
Legal
Don't want: New AI legal tool
Actually want: DocuGPT inside Databricks (where they already work)
Sales
Don't want: Another dashboard
Actually want: Agent in Slack (where deals already happen)
Customers
Don't want: New support portal
Actually want: Simple chatbot (on the existing website)
The New Moat
From: UI habituation and switching costs
To: Agent construction kits and evaluation infrastructure
It is more important for teams to build, measure, and improve their own agents - not figure out how to setup the most feature-complete dashboard. This is what SaaS has to reckon with.
