BP Accelerates Maintenance Decisions With Conversational Analytics With Genie

The story validates why AI analytics only works on governed data.

The speakers frame a clear lesson for enterprise AI teams: remove the analyst as the translation layer by making data itself speak the shared language of the business. Once that foundation is in place, conversational agents can be configured in minutes and deliver accurate, low‑latency answers at scale. That principle aligns with how leading companies are delivering trusted AI‑powered self‑service analytics by investing in data foundations before deploying conversational interfaces.

From fragmented sources to a shared business language

BP's maintenance data spans upstream, midstream, downstream, refining, logistics, and dozens of SAP and Maximo instances. Inconsistent semantics, such as multiple spellings for the same region, meant business questions triggered weeks of mapping and back‑and‑forth. Dirty data led to untrustworthy AI.

The fix was not more analysts. It was a unified model and vocabulary. BP centralized systems of record through Unity Catalog, then standardized entities so tables read as equipment, work order, functional location, and notification rather than cryptic source‑system names. Engineers now ask for data the way they think about operations, which cuts translation cycles and builds trust.

A seven‑step path to AI readiness

BP's pipeline follows a medallion pattern with explicit checkpoints for AI use:

• Raw data lands in cloud storage, then is governed in Unity Catalog.
• The silver layer enforces a proper data model with primary and foreign keys and replaces source‑specific jargon with human‑readable entities.
• The gold layer aligns to domain vocabulary and analytics use, shaped with subject matter experts.
   

Metadata is captured at each layer, including ownership, classification, lineage, quality definitions, tags, and support contacts. An AI data readiness scoring model evaluates five dimensions, with data quality weighted most heavily at 30 percent. Structure and standardization, metadata and lineage, accessibility and usability, and governance round out the score. Only datasets that score four or five advance to AI use.

Engineering and industry standards are embedded in the models. Lineage traces back to original sources, and data quality is continuously monitored, with BP testing Lakeflow Jobs for automated scoring.

Modeling once, conversing many times with Genie

With data AI‑ready, setting up a Genie conversational agent typically takes about ten minutes. Instructions teach the agent domain standards, metric definitions such as mean time between failure, and calculation rules. Current practical scope supports about 25 to 30 tables or views per agent. Larger domains are handled by composing multiple agents behind an agent supervisor so users query across a broader surface through one interface.

Good modeling pays back in speed and cost. BP observed identical questions answered around four times faster on well‑keyed, well‑defined tables because Genie could use known relationships rather than scan content to infer meaning. Recent product updates begin surfacing data model relationships directly, which reduces the need to re‑teach joins inside the agent and improves reliability.

Operating at scale: trust, cost, and governance

BP rolled the experience out to 400 users with strong feedback and plans to reach roughly 9,000. Trust remains fragile. A single wrong answer can set adoption back, which is why data quality carries the highest weight in the readiness score.

Usage cost is the other operational watch point. The plan described includes 10 dollars of free usage per user per month, roughly 150 DBUs or 80 to 100 questions, after which charges accrue on a pay‑as‑you‑go basis. Field engineers can exceed that quickly, so BP applies new governance controls that set cost limits at the agent level to keep consumption predictable. Accessibility and reuse also matter. Popular maintenance datasets feed digital twins, production, and risk management, so certification and sharing controls are part of the readiness process.

The speed came from data quality, not the chat UI

BP's weeks‑to‑seconds leap came from standard models, strong metadata, lineage, and governance. Conversational AI then became the natural interface to a well‑built foundation. Teams chasing similar gains should score AI readiness before measuring AI output, and only promote datasets that meet quality, structure, and governance thresholds.