The Engineer Isn't the Hard Part: Six Commitments PE-Backed Practices Must Make Before AI Moves Their RCM

On May 4, Anthropic and a consortium of Blackstone, Hellman & Friedman, Goldman Sachs, and others announced a $1.5 billion AI services firm. Forward-deployed engineering teams from the new firm will embed directly inside the operations of PE-backed companies whose sponsors funded it. For everyone below that economic floor, the announcement does not change the question that has to get answered first. Whether an AI investment in your RCM operation moves your KPIs or feeds the 95% of pilots that never reach production comes down to six commitments the platform has to make before any engineer shows up.

The forward-deployed engineering model works for the operations of PE-backed healthcare MSOs large enough to absorb a small engineering team embedded for months at a time. Sub-platform PE-backed practices sit below that floor: the rollups still running multiple EMR and PM systems bootstrapped together, or different instances of the same platform inherited from prior partial acquisitions. Even if the economics worked for them, the engineer is not the bottleneck. The engineer does not know how a telehealth platform handles eligibility verification across dozens of payer feeds when half reject the telehealth modifiers, why an ophthalmology MSO's central billing office is at war with the front desk at site seven over eligibility routing, or which site-level write-offs the platform CFO classifies as recurring versus non-recurring at the consolidation layer. None of that lives in documentation. It lives with the operators running the work, and it has to be surfaced and translated before any tool can be designed against it.

The Six Commitments

The first is honest data infrastructure assessment: a current-state inventory of which PM and EHR systems feed which billing platform, which sites are still on hybrid workflows, where the data warehouse exists or does not, and which payer feeds reconcile cleanly. Aspirational data maps do not count.

The second is process documentation that captures what actually happens, not what the SOP says happens. Site-level workflows have almost always diverged materially from stated standards, and the divergence is the source of the noise the AI is supposed to clean up.

The third is KPI baselining. A platform cannot measure whether AI moved its denial rate, days in AR, first-pass clean claim rate, or cost-to-collect without a defensible baseline, and producing that baseline almost always surfaces reporting inconsistencies across sites that have to be resolved first.

The fourth is vendor evaluation discipline. The KLAS December 2025 data identified 657 vendors in the healthcare AI space, more than half of them mentioned by only one survey respondent. The field is dominated by undifferentiated point solutions making functionally identical claims, and most platform leadership teams do not have an internal framework for distinguishing genuine workflow integration from demo theater. The cost of that gap is paid in failed pilots.

The fifth is compliance and audit trail design from day one, not as a retrofit. Retrofitting audit logs, PHI redaction, and explainability features into a finished AI product is significantly more expensive than building them in. For PE-backed platforms where compliance scrutiny intensifies during exit diligence, this corner cannot be cut.

The sixth is change management bandwidth at the site level. AI tooling that the people who are supposed to use it do not adopt produces none of the projected ROI, and adoption requires sustained operational engagement, not a vendor training session.

Why PE-Backed Platforms Are Structurally Disadvantaged

Two structural realities make these six commitments harder for PE-backed platforms specifically. The first is system fragmentation. A urology rollup, oncology platform, or multi-specialty MSO assembled over the past three to five years is operating on a patchwork of practice management systems, EHRs, billing platforms, and clearinghouses inherited from each acquired practice. The Reveleer and Harris Poll 2025 State of Technology in Value-Based Care report found only one in three providers rate their data integration capabilities as excellent, and fewer than half are highly confident in the accuracy of the patient data they are using.

The second is hold-period pressure. Three-to-seven-year hold periods compress the timeline for showing EBITDA improvement, which pushes platform leadership to deploy point solutions on top of fragmented data infrastructure rather than invest in the unification work that would make those point solutions perform. The AI tooling gets pointed at a data foundation it was not designed to operate on, the pilot underperforms, and the platform either kills the project or keeps it alive in pilot purgatory long enough that the next acquisition disrupts whatever standardization had begun.

The Numbers

KLAS Research's December 2025 update found more than two-thirds of healthcare organizations were using some form of AI by year-end. Within RCM specifically, claims adjudication and coding automation each had 24% adoption, denials management 17%. Bain and KLAS named RCM the top provider AI investment priority for 2025. CFOs are looking for measurable financial returns inside a twelve-month horizon. What happens after the capital is committed is where the picture gets uglier. MIT Sloan's 2025 research found 95% of generative AI pilots fail to scale to production. Deloitte's 2025 enterprise AI survey reported 42% of companies abandoned at least one AI initiative during the year, with average sunk cost between $4.2 million and $7.2 million per abandoned project. Within healthcare specifically, industry analysts place the failure rate at 78.9%, with the dominant abandonment causes being data quality issues judged insurmountable (38%), business case no longer viable (29%), loss of executive sponsorship (21%), and technical infeasibility (12%).

For sub-platform PE-backed practices evaluating AI investment in the second half of 2026 and into 2027, the practical question is not whether the technology has matured. It has, and the use cases are well-defined. The practical question is whether the platform has the data infrastructure, the documented processes, the baseline metrics, and the operator-led translation capacity to make any AI deployment perform. The engineer doesn't determine whether the tool performs. The work that happens before the engineer arrives does.