--- title: "The Godfather Problem" url: "https://atlan.com/context-and-chaos/issue/the-godfather-problem/" description: "Your AI's context has a floor. Most teams have never measured theirs. Here's the one number I use." keywords: "Context Engineering, Enterprise AI, AI Readiness, Data Governance" --- > 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 deep dive by Landon Pangburn (Enterprise Account Executive, Atlan), published June 18, 2026 (11 min read). When enrichment agents produce accurate column descriptions, data leaders ask "what's the catch?" The answer: the good output is real, but context is not a yes-or-no property. It has a floor, almost nobody measures where theirs sits, and one number (70% steward acceptance on a scoped domain within 30 days) tells you. About the author: Landon Pangburn helps AI and data leaders establish scalable and trustworthy AI. He has held multiple roles at Atlan over four years and partnered with hundreds of AI and data leaders. LinkedIn: https://www.linkedin.com/in/landon-pangburn/ ## The Godfather Problem Ask a general-purpose model to describe `revenue_usd` and you get "Revenue in US Dollars": correct the way calling The Godfather "a movie about a family business" is correct. Feed an agent lineage, query history, glossary and downstream consumption, and it returns: recognized USD revenue, computed from `orders.amount x fx_rate`, filtered to shipped and delivered orders, consumed in the Finance team's MRR dashboard. Same model, different diet: not smarter, better fed. The catch: that output is only available to teams whose signals are good enough, and most teams can't tell whether theirs are. ## How the descriptions actually get built The agent runs a sequence where each step narrows ambiguity for the next. A column's meaning is the union of four signals; schema alone is too thin: - **Lineage:** how it was built. - **Query history:** how it is used. If `customer_tier` is always grouped with `SUM(revenue)` and filtered to enterprise and mid-market, it is a revenue segmentation dimension whatever the docs say. - **Glossary:** the agreed business vocabulary. - **Downstream consumption:** the anchor. A column feeding a Finance-owned MRR tile has its meaning fixed by that usage. Columns with semantic weight cluster by meaning, not name: `cust_id`, `customer_sk`, `party_id` and `AccountId` can be the same concept, which string matching misses. Resolved, the domain map becomes a typed graph with known join keys, and a KPI becomes an aggregation over a known semantic scope (MRR is the right revenue, over the right products, with the right recognition date, in the right domain). Otherwise you face forty-seven revenue columns all raising a hand. That dependency is where the floor lives. ## The diet has a floor Context-fed AI only works above a coverage threshold. Below it, the failure is confident errors, not gaps: a plausible-but-wrong description certified by a steward skimming a batch of two hundred is worse than a blank field. The blank says "unknown"; the wrong one says "trust me". Two places the floor gives way: - **Lineage coverage.** Transformations outside the lineage graph (uncaptured Python pipelines, ad-hoc notebook transforms, transformation models whose lineage was never exposed) force the agent to infer construction from naming patterns and column shapes: plausible, sometimes wrong. - **Query history depth.** New warehouses, freshly migrated datasets and low-usage domains have thin behavioral signal, and output drifts back to the generic. ## A number that answers it **70% steward acceptance on a scoped domain, inside a thirty-day window:** of the descriptions the agent proposes for a bounded domain, how many your own stewards accept without edits. Run it on one domain you know cold. - Why 70%: in the author's projects, that is roughly where the governance overhead of catching errors stops exceeding the productivity gain. Below it, stewards spend more time correcting than writing from scratch; above it, they are in faster editing mode and the cycle accelerates. - The author's own numbers (not an industry average): enrichment agents applied descriptions at six-figure scale within weeks, with steward acceptance above 90% once a domain was above the floor. Rollouts that used to take most of a year compressed dramatically when the substrate was already in good shape. - The compression is a property of the substrate, not the agents. Teams that hit it had lineage and query history above the floor before the agents arrived, not newer models. ## The flywheel, and why sequence matters Every approved description becomes a signal for the next run, so each pass starts from a higher floor. Enrich and certify your cleanest, most-used domain first, and the agent carries that resolved vocabulary into messier neighboring domains. Starting in sparse, low-traffic corners gives the agent its hardest work with its thinnest signal. Cold-start advice: **start with your most-queried assets**, not the messiest or most neglected. They have the richest behavioral signal and clear the 70% bar fastest. ## How to keep it honest above the floor Four conditions any system needs, built or bought (design constraints if building, evaluation criteria if buying): 1. **The agent proposes; a human certifies.** Nothing publishes automatically; the steward shifts from writing to certifying and handling exceptions. 2. **Generated content stays in its own lane.** Machine-written descriptions stay distinguishable, and human authorship always wins. 3. **Everything is marked, logged, and gap-fill only.** Generated output is flagged, every run is auditable, and empty fields are filled rather than existing work overwritten. 4. **Runs are deliberately scoped.** Enrichment targets a bounded, chosen set of assets, runs when the team triggers it, reads metadata and usage logs rather than raw data, and inherits the warehouse's existing permissions. Skip one and a tired steward batch-approves a thousand confident, wrong descriptions. Below-floor coverage and ungoverned above-floor coverage fail the same way: a catalog full of plausible wrong answers. The 70% bar tests the first; these four conditions test the second. ## The catch It is usually neither a rigged demo nor an unreachable level of upstream cleanliness. "Feed it context" was never a switch; it is a coverage threshold most teams have never measured. Before evaluating another model, run the 70% test on one domain you know cold. It tells you whether the problem is the model or the diet, and it is almost always the diet. ## About Context & Chaos A community newsletter where practitioners, builders and thinkers share lessons on context engineering, governance, architecture, discovery and the human side of data and AI work. Contribute: https://atlan.com/context-and-chaos/contribute/ Related reads: - [Snowflake Summit 2026: Everyone Owns Context Now.](https://atlan.com/context-and-chaos/issue/snowflake-summit-2026-everyone-owns-context-now/) - [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/) - [BI-Ready Is Not AI-Ready](https://atlan.com/context-and-chaos/issue/bi-ready-is-not-ai-ready/) - [The Onboarding Gap Is Killing Your AI Agents](https://atlan.com/context-and-chaos/issue/the-onboarding-gap-is-killing-your-ai-agents/) - [Browse all Context & Chaos articles](https://atlan.com/context-and-chaos/)