Forward deployed engineering for healthcare: the operating model behind AI that actually ships.

Only 38% of healthcare leaders said their most recent EHR implementation hit the mark. About 40% reported significant misses, according to the KLAS Arch Collaborative 2025 EHR Implementations report. AI is repeating the pattern: a budget gets approved, a pilot gets stood up, a vendor demos a model, and 18 months later nothing is in production at scale.

The reason is almost never the model. It is the gap between the model and the operating reality. Healthcare workflows are messy by design. Specialty coding, payer rules, EHR diversity, HIPAA, prior auth pathways, scheduling templates, and the staff who actually run those workflows are not problems an off-the-shelf SaaS can absorb. They are problems someone has to sit with.

That is what Forward Deployed Engineering is for.

A Forward Deployed Engineer, or FDE, is a software engineer embedded inside the customer's workflow, on the vendor's payroll, building capabilities the customer would otherwise wait quarters for. Palantir built the model. Pragmatic Engineer described it for an engineering audience. Anthropic, OpenAI, and Google now hire for it. The New Stack reported that FDE roles grew 42 times between 2023 and 2025, the fastest-growing job category created by AI.

An FDE is not staff augmentation. Staff-aug ships a body without context, without IP transfer, and without a shared incentive. An FDE is not a SOW-driven consultancy. Big-3 firms ship process, slides, and a sub-contracted team that exits at go-live. An FDE is not a customer success manager. Cara's working definition is one step more specific: an FDE plus a project manager, both Cara employees, sized as a fractional or full-time pod, travelling to the customer.

Palantir expects FDEs to spend roughly a quarter of their time on-site with customers. In healthcare, peers like Commure run closer to half. The shape that fits a medical group, an MSO, or a digital-health operator is the same: same people, on the ground, for as long as the work warrants.

Fractional CTO. Strategy without hands. Useful for early-stage planning, expensive for the operational engineering that actually changes throughput. Typical market rate is $5,500 to $14,000 a month for advisory time, which is the wrong unit for the work in question.

Big-3 healthcare consulting (Tegria, Nordic, Chartis, Huron). Strong at process design, slow at production code, expensive at scale, and structurally incentivized to leave at go-live. The work product is a deliverable, not a system. Many customers report needing a second engagement to actually operationalize the first.

Dev shop. Builds, leaves, no clinical context. Functional for one-off websites or microsites. Wrong for anything that touches an EHR, a payer portal, a clinical pathway, or PHI.

FDE pod. Clinical context, builds, stays, owns the outcome. Cara employees on Cara payroll. Same people through observation, build, and the years afterwards. This is the missing shape in healthcare technology buying, and it is the one buyers describe when they say they want a partner who shows up.

Three forces are converging. First, Software 3.0, in Andrej Karpathy's framing from his June 2025 YC AI Startup School talk, is making natural language the new programming interface. That moves the bottleneck off model capability and onto domain translation. Healthcare workflows are exactly the domain where translation is hardest, because the people who hold the workflow do not write code and the people who write code rarely sit in the workflow.

Second, the cost of getting it wrong is becoming visible. Primary care physicians spend 36 minutes in the EHR per roughly 30-minute patient visit, and family doctors log 86 minutes of nightly pajama time inside the chart, according to AMA reporting. That is a clinician-burden cost any FDE pod is paid to eliminate.

Third, MSOs and CINs are consolidating but inheriting incompatible stacks. The MSO operator inherits an AdvancedMD here, an Athena there, a homegrown intake tool, three different scheduling templates, and a marketing site no one has touched since 2022. SaaS cannot fix this. A consulting firm cannot fix it inside the timeline. An FDE pod can, because the work is exactly the high-context, repeatable, evidence-heavy engineering FDE is built for.

The global healthcare IT consulting market is roughly $65 billion in 2025 and projected to exceed $180 billion within the decade. That budget is moving. The question is whether it moves to firms that bill for slides or firms that ship code.

One engineer plus one PM, both Cara employees, embedded onsite at meaningful intervals during ramp. Two weeks of discovery sitting next to the front desk, the MA, the biller, the ops lead, and the physician. The first production agent or workflow live inside 30 to 45 days. Weekly demo cadence to clinical and executive stakeholders. Monthly business review against the KPIs the customer agreed to. Quarterly on-site planning. Written decision log and status, every week, signed by the PM.

Two real-world patterns illustrate the shape. The first is the anchor-customer pattern: Hy-Vee Health Exemplar Care, a DPC group in Iowa where a Cara FDE drove rollout across patient acquisition, scheduling, onboarding, and clinical workflows. The second is the scaling-MSO pattern, in flight at Thrive Medical in Austin, where the FDE plus PM are embedded with a five-clinic MSO ramping toward a 100-clinician Texas CIN on AdvancedMD.

The same people stay on the account. That is the operating commitment that separates Embedded Services from a Sprint, a SOW, or a vendor-success function.

Are the engineers your employees, or sub-contracted? Will they sit with our staff, or call in? What is the post-go-live ownership model in writing? Who keeps the IP, and is the transfer language in the MSA or only in the SOW? What is the exit ramp if we want to take it in-house in year three? Do you take outcome-based pricing on any portion of the engagement? What is the named-person commitment, and what happens when that person rotates? What is the security posture, what is the BAA scope, and where does PHI live? What is the cadence of written status, and who signs it? What is the first thing you will refuse to do, and why?

Cara's view: a vendor that cannot answer all ten in writing, in advance, is not running an Embedded model. They are running a staff-aug or consulting model with FDE language pasted on top.

Be honest about the cases where FDE is overkill. If the workload is generic, appointment reminders, eligibility checks, basic patient comms, buy SaaS. The market for these is mature and cheap. If the organization has a deep internal engineering bench, a Providence or a Kaiser, license a platform and skip the embed. The internal team will execute faster than any vendor pod.

FDE is the right answer for the middle: independent medical groups, MSOs and CINs, regional health systems, digital-health operators scaling past their first ten customers, payers piloting in a single market, and pharma medical-affairs teams that need a real software loop instead of another deck. Gartner has cautioned that 70% of enterprises will be forced to abandon agentic-AI deployments delivered through FDE-led engagements by 2028, because of vendor-lock-in and a lack of internal skill to evolve them. That risk is real. The mitigation is a contractual IP-transfer and exit ramp, written into the engagement at the start. Cara does it that way on purpose.

FDE job postings grew 42 times in two years. Anthropic, OpenAI, and Google are hiring against the same template Palantir wrote. CIOs are quoting Gartner research naming FDEs as the new limiting factor for enterprise AI. The model works because the bottleneck for production AI is not the model. It is the loop between the model and the operating reality of the customer.

The question for healthcare buyers in 2026 is not whether to use an FDE-style partner. It is whether yours is on the vendor's payroll, on your floor, and contractually obligated to leave the IP behind.