Share with your CDO
Enterprise AI spending has quietly split into two camps: companies still racing to deploy more agents and consume more tokens, and a smaller group asking why their costs keep climbing while outcomes stay flat. Salesforce VP Clara Shih frames this as an architectural problem, not a pricing one. The leaky AI pipeline argument identifies five compounding drains, from bloated context windows to stateless agents reprocessing identical requests, and proposes that master data management and governance infrastructure are the actual fix. Uber’s runaway AI spend gets a brief mention as the cautionary case.
What this means for your business
If your organization is measuring AI maturity by deployment count, this argument is aimed squarely at you. The five failure points described here, bloated prompts, ungoverned data access, defaulting every task to frontier models, stateless memory, and redundant retrieval, are not edge cases. They are the default state of an enterprise that moved fast on pilots and deferred the data foundation work. Companies already sitting on mature MDM infrastructure and governed data catalogs are quietly running cheaper AI than their token-volume peers, even before any model price cuts hit.
The argument holds structurally, but the framing carries a Salesforce and Informatica shaped thumb on the scale. The specific fixes proposed, master data management, a unified customer data platform, pre-curated golden records, happen to map exactly onto the product portfolio the author sells into. That doesn’t make the diagnosis wrong. Context bloat from dirty, duplicated data is a real and measurable cost driver. But the prescription skips routing logic, fine-tuning, and agent framework design as efficiency levers, all of which matter and none of which require a CDP purchase. Treat the five pillars as a useful diagnostic, not a complete one.
The decision this actually reframes is whether your AI cost conversation is happening in the finance function or the data function. If your CFO owns the token budget and your CDO owns the data quality roadmap, those two conversations are probably not connected yet. The companies that close that gap first will find that model price compression from Anthropic and OpenAI becomes a bonus rather than a lifeline. I’d revise this view if frontier model costs fall fast enough that architectural waste simply stops mattering at the margin, but at current enterprise consumption scales, that threshold is still years away.
Concept deep-dive: Master Data Management
Master data management, or MDM, is the discipline of creating a single authoritative record for key business entities like customers, accounts, and products by deduplicating, merging, and reconciling data from multiple source systems. Think of it as the difference between giving an AI agent a clean, verified customer file versus handing it three conflicting spreadsheets and asking it to figure out which one is current. For AI pipelines, MDM directly reduces the volume of raw context that must be stuffed into every prompt, which is where a large share of token waste originates.
Based on reporting from Salesforce VP on the leaky AI pipeline: why cheaper tokens won’t fix enterprise AI, originally published 2026-07-20 07:07:00.
