---
title: "The Operating Model for Shipping AI-Generated Context"
url: "https://atlan.com/context-and-chaos/issue/the-operating-model-for-ai-generated-context/"
description: "Generating context at scale is no longer the bottleneck. The business signal was already in the estate. Shipping it, and the operating model around it, is the work."
keywords: "Context Engineering, AI Agents, Data Governance, AI Readiness"
---

> Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/

A Context & Chaos article by Nandini Tyagi (Strategic Initiatives, Atlan), published October 1, 2026 (13 min read). It draws on [a conversation with three practitioners](https://atlan.com/context-in-practice-can-ai-document-your-data/) who used AI to document their company's data at a scale their teams could never have staffed by hand: Zhenni Hu (manager, AI & data governance, Mastercard), Rupal Sumaria (head of data management, ASOS) and Keith Guyett (principal solutions architect, Sophos). Generating context at scale is no longer the bottleneck. Shipping it, and the operating model around it, is the work.

Zhenni opened by asking the audience two questions. Have you used AI at work this week? Plenty of thumbs-ups. Would you be comfortable letting AI make a decision you were held accountable for? None. Everyone uses AI, but few data practitioners will sign their name to what it produces. The models were not the question; whether the material underneath could be vouched for was.

Metadata was built for humans, and now AI uses it just as much, so the data catalog is changing (as Prukalpa [recently argued](https://atlan.com/know/is-the-data-catalog-finally-dead/)). Agents write context, read it back at runtime, and quality compounds: better evidence at one layer produces better context at the next. All three agreed agents can write context. The question is what it takes to get context you would put your name on.

## The problem: Metadata built for speed, not agents

- Sophos had more than 600,000 assets in its catalog by March. The first nine thousand enriched quickly, and speed turned out to be the wrong target. "Doing it fast just gave us static restatement of the underlying data model. It did not provide any business purpose." (Keith)
- A restatement is what the model returns when it can read only the schema: the customer_id column contains customer identifiers. Nothing in it is wrong, and it helps no one decide anything.
- "A good enough description for a human is not necessarily a good enough description for an AI agent, because the human has that business knowledge ingrained. The AI agent has to reason through it." (Keith)
- A person reading a thin description already knows which of three near-identical order tables finance trusts, that a field changed meaning after a migration, and what risk means by "exposure". The description was a pointer; the reader supplied the rest from memory. An agent reasons only over what is written down.
- The standard moved because the reader changed, and most estates are full of documentation that doesn't make sense to AI.

## Widen the control scope to add tribal knowledge

The [first phase](https://atlan.com/know/is-the-data-catalog-finally-dead/#rxc-2) of an agentic data catalog is creating context for agents. The first idea: **control scope**, what the agent sees and uses to write descriptions.

- Sophos fixed its first run by changing the inputs: give each description a business purpose, state the grain and scope, consider metric relevance, use examples where they exist, and read the lineage and source context (Keith's "metadata around the metadata"). The same model on the same assets then produced something a reviewer could work with.
- Mastercard reached the same conclusion: column names alone yield a polished restatement of the schema. Usable drafts came from reasoning over real signals: how the data is queried, what it connects to upstream and downstream, and existing glossary terms, classifications and data products.
- The difference between a generic and a useful description is what the model had to work from, not the model. Correcting a close draft takes minutes; rewriting a generic one takes as long as writing it from scratch.
- Sophos re-enriched some of the first nine thousand and reached almost 43,000 enriched assets between April and early May. They are near 100,000 now.

## Think left to right when building a context pipeline

The second idea is the **order** in which assets are enriched before agents consume them, the [second phase](https://atlan.com/know/is-the-data-catalog-finally-dead/#rxc-3) of an agentic data catalog.

- "I prefer following the chain of my pipeline because I think you build knowledge and transformation on top of each other... For me, it's left to right." (Rupal, ASOS)
- Source descriptions (business purpose, grain and scope, usage and lineage) feed the semantic model (entities, relationships, definitions), which grounds the metric layer (approved formulas, ownership, business scope), which the agent uses to produce an answer, decision or action. Each stage is built from the one before, so the order cannot be rearranged.
- Mastercard ran the same sequence as program order: build the foundation, enrich it, make it discoverable, and only then hand it to agents.
- "Handing organizational context to an autonomous agent where you have not verified that context yourself is a fast way to lose trust in the whole program." (Zhenni)
- An agent enriching a semantic layer over restated source descriptions produces a more confident restatement one level up. Errors become the input to the next layer; the compounding has a direction.

## Consider how much AI-generated context can be reviewed

The third idea, usually discovered last: **how many drafted descriptions** a human can really get through. It belongs to the [third phase](https://atlan.com/know/is-the-data-catalog-finally-dead/#rxc-4), learning from failures and compounding feedback.

- Running bootstrapping agents across a whole estate is one click, but hundreds of thousands of overnight descriptions build a review queue no team clears, and the first error a steward finds casts doubt on the whole batch.
- Mastercard instead took a bounded set of critical assets: a two-week pilot covering approximately 30,000 assets across roughly 700 tables and views, estimated at about 6,200 hours of steward effort saved.
- Its subject matter experts answered two questions: what percentage could you accept unchanged, and for the rest, were you editing or rewriting from scratch? Accept and edit mean the draft was worth having; rewrite means the time was never saved. Roughly half of the sampled critical assets came back untouched. Output was strongest on often-curated tables with real lineage and usage, weakest on assets nobody touches and on terms that mean different things to different teams.
- Sophos built a test harness in Cortex Code that graded all 43,000 descriptions from A to E. Around 80% came back A or B, so Keith could hand the business a pre-screened set instead of an open queue.
- "Perfection is not how your business runs. Your business runs on, can we get 80% there, so we operate the same way." (Rupal)

## Building these ideas into the context development lifecycle

- Sophos codified the three principles into its deployment lifecycle, from dev through QA into production. Descriptions and terms must be defined as part of a data product's rollout, or the product does not move forward. At each gate: dev requires a description and attached terms, QA requires verified lineage and reviewed definitions, and production is reached only once context is certified.
- It works like making tests a merge condition: nobody argues about writing them because the pipeline will not let you skip it.
- Bad context can no longer fail quietly. People used to work around a bad definition without filing a ticket; now it fails in public, fast enough to trace. At ASOS, business teams have started asking for metric ownership to be codified and settled like any other development process.

## How do we get the rest of the context in?

- Per-asset enrichment is not the end. "Enriching the individual asset might not be enough. The real context also comes with the relationships, whether you call it lineage or ontology, how the assets are meant to be used together, which ones to join and how, what the agreed metrics are, what guidance or prompts or instructions apply to them." (Zhenni)
- Agents need context that sits in no structured field: documents, threads and people's heads.
- On consumption, Sophos is building routing so a business question to an assistant goes to the governed catalog first and gets the real definition before the model reasons on its own.
- The open problems: how to give agents different types of context, and how to make sure it is the right context. Agents changed the economics of producing context; they did not remove the judgment that makes it worth trusting. Bootstrapping context has become dramatically easier, but shipping it to the business requires an operating model, and the companies that build that model faster will be the ones actually using and iterating on the context they generate.

Closing cartoon, "The Cats of Context & Chaos": a governance lead demanding all the metadata be written before launch beside exhausted stewards, versus reviewed, ready context shared with AI agents. Caption: "Now we can keep up."

## About the author

Nandini Tyagi leads Strategic Initiatives at Atlan, focused on how AI products are launched and adopted across enterprise customers. She works with data leaders on moving from AI strategy to implementation, and writes about what it takes to make AI stick inside large, complex organizations. [LinkedIn](https://www.linkedin.com/in/nandinityagi/).

## Related reads

- [Your AI Context Layer Is Being Built on Stale Metadata](https://atlan.com/context-and-chaos/issue/your-ai-context-layer-is-being-built-on-stale-metadata/) (May 2026)
- [BI-Ready Is Not AI-Ready](https://atlan.com/context-and-chaos/issue/bi-ready-is-not-ai-ready/) (March 2026)
- [Four Architectures That Make AI Work](https://atlan.com/context-and-chaos/issue/four-architectures-that-make-ai-work/) (September 2026)
- [Conceptual Modeling Is the Context Engineering Nobody Is Doing](https://atlan.com/context-and-chaos/issue/conceptual-modeling-is-the-context-engineering-nobody-is-doing/) (April 2026)

Context & Chaos is Atlan's community newsletter on context engineering, governance, architecture and discovery. Browse all articles at https://atlan.com/context-and-chaos/ or pitch a piece at https://atlan.com/context-and-chaos/contribute/.