Why this is in the vault
The founder shared it with no comment. It is Lieberman's notes from a 60-minute call with the Head of AI at an unnamed $5B real estate company. It is one practitioner's field report, not a framework. Three lines matter to RDCO: a buyer saying "own your context" (the thesis of [[The Denominator]]), the "walls go up during the work" failure mode (phData delivery reality), and "inference is the new labor cost" (an ROI frame).
⚠️ Structural Bias
Glean and EndexAI are tagged by handle and praised. Lieberman's media business runs on sponsorships. No disclosure is visible in the post. Treat the Glean section as possibly paid placement. The interviewee is anonymous, so none of the claims can be checked.
The core argument (paraphrased)
- Sponsor and failure modes. The CFO drove the program. It fails in two ways: a C-suite that does not believe in it, or IT and security teams with no reason to learn AI. The blockers are data access, politics and getting people's time, not the technology.
- Agreement in the meeting, walls in the work. Everyone agrees, then the work stalls on "I don't know if I can give you access to that table" and "ask the business owner," even with a CEO directive.
- Sequence. First, one safe point of entry for everyone plus enablement that scales (they picked Glean, which keeps source permissions). Then watch who rises. Subject experts with "just enough training" build things the central team would not think of. The bottom-up use shows where top-down transformation can avoid political or legacy traps.
- ROI posture. Leadership accepted estimates that could be off by 20-50%, because the cost was too small to block a start.
- Ownership. Decentralized execution, central planning. The AI owner needs a founder's mindset, thinking across functions, and high EQ, because people assume the owner is coming for their jobs.
- Own vs rent. Own your context (how data is connected and fed to agents) and your strategy. Inference is the new labor cost: treat it as a cost-of-goods input and justify spend by use. Everyone gets a minimum level of access. All-inclusive per-seat pricing will not last. Homegrown models make sense only at very large scale or for proprietary decisions.
- Open infrastructure. No settled answer for how agents run in production (one platform or six), identity, or observability.
Mapping against Ray Data Co
- The Denominator / personal brand: "Own your context... how you connect it and feed it to models and agents is the real question" is the definitions-layer thesis from post #1 ("agents need your definitions"), told from the buyer side. It is a usable outside quote for the series. See [[2026-10-02-site-spec]].
- phData: "Walls go up during the work" names the access and ownership friction that sits between an approved use case and a running agent. It fits the "agents in production" credibility lane ([[project_credibility_for_phdata_sales]]).
- ROI framing: "Inference as cost of goods, justify by use" pairs with the metering gap in [[2026-10-20-secret-cfo-ai-spend-payback-mailbag]] (99% of value from 1% of usage).
Related
- [[2026-10-20-secret-cfo-ai-spend-payback-mailbag]]
- [[2026-10-02-site-spec]]
- [[2026-10-20-every-shared-agent-permissions-architecture]]