Reading time: 6 minutes | Issue #40 | Book a Discovery Call

Happy Tuesday. Mark here.

Cedar launched Kora last week: three AI agents that handle patient billing calls, collect payments, and resolve denied claims. Trained on 1.5 billion patient interactions across 63 million patients. 40% of inbound billing calls resolve without a human. I read the announcement and felt admiration first, then frustration, because the adoption data underneath tells a different story.

Inside the Issue

88% of AI Pilots Never Reach Production. We Shipped Seven in Four Months.

IDC's CIO Playbook research, conducted in partnership with Lenovo, tracked what happens to AI proofs-of-concept inside enterprises. For every 33 POCs a company launched, four graduated to production. The other 29 stalled or got killed. S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The abandonment rate doubled in twelve months.

I've watched the pattern from inside dozens of engagements. A team bolts a model onto the product, watches it hallucinate over unstructured inputs, and concludes that AI isn't ready for their industry. The model worked. Nobody structured the inputs it needed to work with.

A PE operating partner asked us to build production AI agents inside a portfolio company's billing system, processing real healthcare claims under HIPAA. Two people had been hand-writing every rule in their claims engine across 300+ denial codes and payer logic that changes with quarterly CMS updates. Four months later, seven production agents handle it with zero patient data exposure. One senior engineer, full-time, with a fractional AI lead and delivery coordinator.

The model was the shortest line item in the build.

The 4x gap

Gartner surveyed 353 data and AI leaders and published the results in April 2026: organizations with successful AI initiatives invest up to four times more in foundational areas (data quality, governance, AI-ready people and change management) than those with poor outcomes. Not a marginal difference. 4x. Gartner also predicted that organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026.

Our first month on the healthcare engagement looked like nothing was happening. We mapped the data: where it sits, what's missing, what's structured versus inferred, what changes when a payer updates their rules. No model, no agent code, no demo for anyone.

Data mapping means cataloging schema definitions, tracing join logic across legacy tables, documenting which fields update on which cadence. The alternative is building an agent that hallucinates because it reasons from incomplete facts, then spending months debugging symptoms instead of causes.

The enrichment layer

We built a layer that assembles 34 dynamic variables per claim before any LLM sees it, pre-computed and versioned so the agent receives ranked facts instead of searching for context: claim type, denial history, payer-specific requirements, procedure codes, prior authorization status, resubmission windows, all structured and queryable before the model touches anything.

A Towards Data Science analysis of 100+ production deployments found that practitioners trace most production failures to retrieval, not generation. The model can reason. It can't reason from garbage inputs. The enrichment layer fixes the inputs, and the model becomes the easy part.

A paper from Commure (Saikat Maiti, VP of Trust) put the architectural principle in sharper terms: "infrastructure-level controls operate independently of LLM compliance." Even if someone manipulates the model through prompt injection, which the paper calls "a structural feature" of LLMs rather than a solvable bug, the infrastructure prevents damage. You strip patient data before the model sees it. You validate the output against a strict schema. Deterministic code accepts or rejects the action. If the output falls outside the allowlist, the system fails closed.

Every agent in our build follows that pattern. Seven agents, each locked to a single workflow (denied claim follow-up, billing reconciliation, payer-specific resubmission), each running its own enrichment payload. None of them pass patient-identifiable information to the model layer.

The eval harness

Then we built the part most teams skip.

Every agent runs against a curated test suite before any update reaches production. When a model provider ships a new version or payer logic changes with CMS quarterly updates, the harness catches regression before a single live claim is affected. The flagship agent reconciles denials to the penny: 59 out of 60 on the eval set against real claim scenarios from the client's data.

We route calls across two model providers. Swapping one changes nothing in the output because the eval harness verifies it. The infrastructure runs on Azure Foundry at $200-$300 per month.

Our read: the 88% that never reach production share a common architecture. The model connects to raw, unstructured data. Compliance lives in a system prompt. There's no regression testing. The team is locked to a single provider. Fix those four things and you move from the 29 that stall to the four that ship.

The operating partner who brought us in now benchmarks the rest of the portfolio against this system. That's the line between a portfolio company running AI and one still running demos.

The Production Readiness Audit for Regulated AI

Four axes. Score each 0 to 3 before your AI agent touches production in a regulated environment. Built from the healthcare billing deployment above and the Commure Zero Trust framework.

Axis 1: Data Foundation

How many variables does your agent receive per input, pre-computed and structured? Is the enrichment layer versioned so you can trace which data the agent used for any given decision? Does your data update on the same cadence as the rules it represents? Payer logic changes quarterly. If your data layer doesn't, your agent reasons from stale facts. Can a new engineer understand the data pipeline from documentation alone?

The red flag here is the agent connecting to a database and writing its own queries. The TDS production framework sets retrieval quality targets at >0.85 context relevance and >0.90 context recall. If you can't measure yours, you don't have a data foundation.

Axis 2: Compliance Architecture

Does any patient-identifiable information reach the LLM? Are compliance controls enforced by infrastructure (network policies, credential proxies, egress rules) or by instructions in a system prompt? If someone manipulates the model through prompt injection, does the architecture still prevent data exposure?

Compliance that depends on telling the agent not to expose data is not compliance. The Commure paper calls prompt injection "a structural feature" of LLMs. If your only defense is an instruction the model might ignore, your compliance posture is a document, not an architecture. The paper recommends four layers: kernel-level isolation, credential proxy sidecars, network egress policies, and prompt integrity frameworks, each working on its own so a failure in one layer doesn't compromise the others.

Axis 3: Eval Coverage

Do you have a curated test suite of real scenarios from your own data? Does every agent run against that suite before any update reaches production? When the model provider ships a new version, how do you verify your agent still works?

Our flagship agent scores 59/60 against real claim scenarios, and we run that suite on every deployment, every model update, every payer logic change. The TDS framework recommends answer faithfulness >0.95 in regulated industries and hallucination rate <2%. If you can't produce those numbers for your agent, the eval coverage isn't there yet.

Axis 4: Model Independence

Can your agent run on a different model provider without changing the enrichment layer or output validation? If your primary provider raises prices 5x, what happens to your margins? We route across two providers. Swapping one changes nothing in the output because the eval harness verifies equivalence. The harness is the proof, not a trust assumption about how the models behave.

Score each axis. More than one score above 7 means your deployment carries risk you haven't priced. Any red in Axis 1 or Axis 2 means production in a regulated environment is premature.

01 Blackstone, Anthropic, and Hellman & Friedman launched Ode: $1.5 billion, 100 engineers. The venture deploys AI implementation teams inside PE portfolio companies. Ode is already operating at 6 of Blackstone's 270+ portfolio companies, with plans for 25. Chief Technologist Eddie Siegel: "I think model selection matters, but it's not where the majority of calories are spent. It's one ingredient in a system." The PE industry's largest bet yet that implementation is harder than model selection.

02 Cedar launched Kora: agentic AI for patient billing, trained on 1.5 billion interactions. Three purpose-built agents handle inbound voice, outbound voice, and bilingual text for billing resolution. 40% inbound call containment across 600,000+ patient conversations. 21% relative payment rate increase within 7 days of outbound contact.

03 FTI's 2026 PE AI Radar found a performance tier forming in PE AI. 95% of PE funds report AI initiatives meeting or exceeding their business case, but only 7% have reached enterprise-scale. An "AI Alpha" tier sees 6% stronger ROI, 4% more revenue growth, and 18% more AI-related exits. The gap between tiers has nothing to do with how much they spend (all invest ~11% of revenue in AI) and everything to do with deployment discipline.

04 Gartner predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026. Salesforce/MuleSoft's 2026 Connectivity Benchmark adds that 96% of IT leaders agree AI agent success depends on seamless data integration across all systems, and 86% warn that without it, agents create more complexity than value.

The Production Readiness Scorecard from Worth Your Time, formatted as a one-page visual checklist.

We build production AI for healthcare and regulated industries.

If your team is working through how to get from pilot to production, we do discovery calls where we talk through your roadmap and see where we can help you ship better.

Until next Tuesday,

— Mark Ajzenstadt, Founder @ Limestone Digital

P.S. We just signed on another PE-backed healthcare company. Slots are limited.