Gorgias Cortex Explained: How Gorgias Built Its Internal AI Agent (2026)

Cortex is Gorgias's flagship internal AI agent. 72% of Gorgias employees use it each week, and it handles more than 10,000 prompts a week. It answers questions in Slack, a browser extension, an internal web app and over MCP, and it does so against the company's governed data, metric definitions and business context. Gorgias built it in about two months. It sits on a shared system the company calls the Decision Engine, which also powers two other agents, GAIA and Oracle.

This guide explains what each part does, the numbers Gorgias has published, how the context layer works, and how to build a Cortex-style agent at your own company. Every claim about Cortex links to Gorgias's own writing: the Decision Intelligence post by VP Howard (Greg) Gregory and the Gorgias Engineering blog. If you have read our breakdowns of Ramp's AI product factory, Vercel's agents or Notion's agents, this is the same format.

Gorgias's internal AI agents at a glance

What is Gorgias Cortex?

Cortex started as a Slack agent, "mostly for data analysis," according to Gorgias engineer Adam Treen. It grew into a company-wide platform. Gregory describes it this way on the Gorgias blog: anyone at Gorgias can ask a question in plain language and get a contextualized answer in seconds.

Where it runs, per Treen's engineering post:

• Slack, which Gorgias engineer Antoine Balliet calls "the best platform for an internal AI agent"
• Linear, Notion and GitHub
• A web app and a Chrome extension
• "anywhere over MCP"

What it can do, per the Gorgias blog:

• Read and write Notion pages
• File Linear tasks
• Create HTML apps
• Automate signal delivery to the people who need it

Gregory gives two examples. A sales lead receives a customized PDF with coverage gaps, pre-sale intent signals and product-fit options. A senior product leader receives an ARR breakdown, projected end-of-month attainment, and recommendations tied to an OKR.

Treen lists what sits behind it: a context layer, live warehouse access, a filesystem, a sandbox, memory, skills, custom agents, automations and monitors, and artifacts and apps. Connected tools include the company's BI tool, CRM, Google Drive, Figma, Granola, Datadog and Sentry.

How fast did Gorgias build Cortex?

"In two months, Cortex went from an idea to fielding thousands of questions every week," Gregory writes. The team behind it had been repositioned in late 2025 from a data analytics function into a Decision Intelligence function. Elliot Trabac leads the Data, Context and AI Engineering teams, and Noemie Happi Nono leads Decision Strategy and Operations.

How many people use Cortex?

Gorgias has published three figures, and they measure different things:

• 72% of employees use Cortex each week (Gorgias blog, March 2026)
• 98% of the company has used Cortex at least once (Treen, July 2026)
• 10,000+ prompts a week (Treen; repeated by Arthur Edmond in a July 2026 benchmarking post)

Weekly usage and lifetime usage are not the same metric, so quote them separately.

The Decision Engine: one context system, three agents

Gregory's main point is that Cortex is not a chatbot bolted onto a model. "We didn't achieve this by simply plugging a large language model into our stack," he writes. The Decision Engine is what lets Cortex "reason against governed data, metric definitions, and business context." The same engine powers GAIA and Oracle, so each new agent starts with the context the last one already had.

That is the reusable lesson. Ramp, Vercel and Notion each built several agents. Gorgias built one context system and pointed several agents at it.

How the Gorgias context layer works

Gorgias's Yochan Khoi documented the foundation in "Building a context layer from the ground up". His summary: "The agent is only as good as what it knows."

1. One warehouse as the universal layer. BigQuery holds the data and, with native vector search, the embeddings.
2. Ingestion and modeling. Data arrives via Airbyte or Gorgias's own tool, Bizon, and is modeled in dbt.
3. Metadata the agent can read. Tables such as ctx__model_metadata and ctx__instructions tell the agent what each model means and how to use it.
4. Progressive disclosure. The agent loads detail only when it needs it, an approach Khoi says was borrowed from Claude Code skills.
5. Customer conversations as context. The layer includes embedded Gong transcripts from CSM and sales calls, so answers can draw on what customers actually said.

Treen describes the result plainly: "This is not some sophisticated RAG system or semantic layer, it's simply a knowledge graph of Markdown docs," backed by "hundreds of certified dbt models." The large majority of retrieval, he writes, "comes down to just Markdown + SQL."

Models and tooling

• Built on LangGraph with traces in LangSmith; a provide_feedback() tool and a #cortex-help channel support debugging (Balliet, May 2026)
• Default model as of mid-2026: Kimi K2.7 Code served on Fireworks AI, after mostly defaulting to OpenAI, Anthropic and Gemini models until mid-May; Claude Sonnet judges evals (Edmond, July 2026)
• Skill retrieval uses BM25 after skills grew "from about 20 skills to 315 in a month" (Edmond, August 2026)

Where the customer evidence comes from

An agent like Cortex can only surface customer signal if that signal exists somewhere structured. Separately, in BuildBetter's case study with Gorgias, the team describes consolidating feedback from support tickets, sales and customer calls, research interviews, Slack, roadmap feedback and CRM context into one signal layer.

Shayan Memarzadeh at Gorgias put the goal this way: "I want BuildBetter to be the source of truth, and that shoves into all our tools..." Estelle David described the value: "This is powerful because we have the calls plus the written feedback plus the support tickets. It's a good 360 view."

This is the general pattern behind any Cortex-style agent. The agent is the interface. The context it reads is the product. Warehouse metrics tell it what happened; structured customer evidence tells it why.

How to build a Cortex-style agent at your company

You do not need Gorgias's team size to copy the architecture. Start in this order:

1. Pick one surface. Gorgias started in Slack. Ship there first.
2. Certify a small set of metrics. Define the 20 to 50 numbers people actually ask about, in dbt or your BI tool, with plain-language descriptions the agent can read.
3. Write instructions as data. Store "how to answer questions about X" as Markdown the agent loads on demand, not as one giant system prompt.
4. Add customer evidence as a first-class source. Calls, tickets, surveys and Slack feedback, structured by product area and account, so the agent can cite real customer quotes next to the numbers.
5. Build a feedback loop. Copy Gorgias's provide_feedback() tool and a help channel, so wrong answers get fixed in the context, not just forgotten.
6. Point a second agent at the same context. That is the Decision Engine lesson: reuse beats rebuild.

Build vs. buy for each layer

If you only build one layer yourself, make it the metric definitions. They are specific to your business. Customer evidence is the layer most teams underbuild, because it is messy and spread across tools, which is why Gorgias consolidated its feedback sources in one place.

Common mistakes when copying Cortex

• Starting with the model. Gorgias switched default models in 2026. The context layer survived the switch; a prompt-only agent would not have.
• Uncertified data. If two dashboards disagree, the agent will too. Certify first.
• No customer voice. An agent that only reads the warehouse answers "what" and never "why".
• No feedback path. Without a way to flag wrong answers, trust drops after the first bad one.

How Cortex compares with other companies' internal agents

Gorgias is not alone. Several companies have published how their internal agents work, and the designs differ in instructive ways.

Two patterns stand out. First, Gorgias's Cortex and Vercel's d0 both live in Slack, where questions already happen. Second, the agents that scaled across the whole company, Cortex and Vercel's d0, both sit on certified data rather than raw tables. Vercel found its data agent got faster and more accurate after removing most of its tools; Gorgias found its agent improved as the context layer got better, even as it changed models.

The difference is scope. Ramp and Vercel built many purpose-built agents. Gorgias put its effort into one context system and let several agents share it. If your team is small, the Gorgias approach is cheaper: you maintain one layer of truth, not one per agent.

For the full breakdowns, see Ramp's AI tools explained, Vercel's internal agents and Notion's internal agents. If customer feedback is the context you are missing, our guide to building a customer insight agent walks through the pipeline.

FAQ: Gorgias Cortex

What is Gorgias Cortex?

Cortex is Gorgias's flagship internal AI agent. Employees ask it questions in Slack, a browser extension, an internal web app or over MCP, and it answers using the company's governed data, metric definitions and business context. It can also write to Notion, file Linear tasks and build HTML apps.

Is Cortex available to Gorgias customers?

No. Cortex is internal. Gorgias's customer-facing product is its separate AI Agent for ecommerce support.

How many Gorgias employees use Cortex?

72% use it each week, according to the Gorgias blog (March 2026). Gorgias Engineering reported in July 2026 that 98% of the company has used it and that it handles more than 10,000 prompts a week.

What models does Cortex use?

Gorgias has used OpenAI, Anthropic and Gemini models, and in mid-2026 switched its default to Kimi K2.7 Code served on Fireworks AI, with Claude Sonnet as its eval judge.

What are GAIA and Oracle?

Both run on the same Decision Engine as Cortex. GAIA is described as an experimentation agent for AI Agent Guidance design; Oracle supports investor readiness and board preparation.

How did Gorgias build Cortex in two months?

By investing in context before the agent: a BigQuery context layer, certified dbt models, readable metadata and instructions, and progressive disclosure. The agent itself is a thin layer on top.

Give your agent the customer context

Gorgias's lesson is that the context, not the model, makes an internal agent useful. BuildBetter turns your calls, tickets, Slack and surveys into structured, cited customer evidence that any agent can read over MCP.