The DeployCo Teardown: How AI Forward-Deployed Startups Actually Work

What a DeployCo is, and the loop it runs

Strip away the marketing and an AI forward-deployed deployment startup is a single repeatable loop:

1. Embed. Put engineers physically (or virtually) inside the customer from "day zero," owning a business problem rather than advising on it.
2. Harvest context. Ingest the enterprise's proprietary data, standard operating procedures (SOPs), systems of record, and the tacit judgment of its subject-matter experts (SMEs).
3. Build agentic logic. Stand up production decision systems and agents on that context — typically on a platform that turns SOPs into auditable, runnable workflows.
4. Bill on outcomes. Tie a meaningful share of the fee to hitting the customer's objective (resolution rate, cost saved, turnaround time), explicitly positioned against time-and-materials consulting.

This is the Palantir forward-deployed engineer (FDE) model, re-pointed from deterministic data integration onto probabilistic LLM/agent workflows. Distyl's own framing — "embed engineering talent, and deliver outcomes within three months"[2] — is the loop in one sentence.

The two anchors, and the one structural difference that matters

Distyl and Northslope run a functionally identical motion — ex-Palantir FDEs embed, harvest proprietary context, stand up custom agentic logic, bill on outcomes — but they finance the stack differently, and that difference is the whole strategic question for a founder.

The lesson in one line: Distyl owns its substrate and carries a platform-company valuation; Northslope rents Palantir's substrate and captures the layer above it. Same loop, different place on the value stack — and, as Section 9 argues, different defensibility.

How a deal actually runs, phase by phase

The defining claim of the model is compression: what legacy consultancies delivered in "quarters-to-years," DeployCos claim to ship in "weeks-to-a-quarter." The evidence is real but uneven — it depends heavily on whether the workflow has a clean "done" state.

Discovery & scoping

Engagements start by finding a high-volume, currently-expensive, checkable workflow. Northslope's founder describes the opening move literally: "Monday of week three, you'll be on your first customer project… you'll show up… and ask: 'What is the hard problem you have?'"[5] No months-long discovery deck — the FDE is in the building, building, almost immediately.

FDE embed & build

The FDE co-builds with the customer's SMEs, shipping a first working version in days. Northslope: "Within two to three days… you'll ship something, get it in front of a user, and then build the next version."[5] Distyl's Distillery is designed so the customer's SMEs refine the AI logic in natural language directly — the FDE wires the substrate, the SME tunes the judgment.[7]

Validation, handoff & "owning the outcome"

"Owning the outcome" is operational and contractual. Operationally, the system runs in production with full observability — Distillery "captures every input, output, tool call, and reasoning step," routes failed tasks to human-in-the-loop review, and feeds analyst corrections back into the system.[7] Contractually, part of the fee is tied to the customer's objective. Distyl claims a "100% production record" across deployments[2] Company claim — a number only a vendor confident in its verification layer would dare put in a press release.

Maintenance & expansion (the land-and-expand)

The motion mirrors Palantir's: services-heavy land (FDEs build the first workflow), then platform-centric expand (more workflows, more departments, recurring license). Sacra characterizes Distyl's contracts as multi-year deals bundling Distillery access with embedded services, "tens of millions of dollars over several years."[4] Inferred — no rate card

2nd order Because every run produces verified pass/fail traces wired into that customer's systems of record, switching costs compound: ripping out the DeployCo means rebuilding the verifier integrations and losing the accumulated correction data. 3rd order This is why the model can sustain outcome pricing — the vendor accumulates a proprietary, per-customer dataset that the model labs do not have, making the next workflow cheaper to stand up than the last.

Who you actually hire — and why it determines the cost structure

The headcount mix is the business model. A services-heavy org with low product leverage looks like a consultancy; a thin FDE layer over a reused platform looks like software. DeployCos try to start as the former and migrate to the latter.

What an FDE actually is

Not a pure engineer and not a pure consultant — a hybrid who owns a customer problem end-to-end. Palantir's internal term was "Deltas," the "tip of the spear" embedded with clients.[20] Everest Group classifies the FDE model as a "category of one": it "operate[s] like a product company in its platform approach but deploy[s] like a consulting firm in its proximity to the customer."[21] Distyl's job spec asks an FDE to design and deploy end-to-end AI workflows, build compound systems (models, agents, retrieval, evaluation), integrate customer data/APIs, and own architecture, security, and observability.[22]

Team shape & the org structure

• Northslope: ~95+ people, the "vast majority" titled Forward-Deployed Engineer; at one point >90% staffed by Palantir alumni.[5] Reported Flat hierarchy — the founder lists himself as an FDE. Pods of small, high-velocity teams ("we sent two people for three months") rather than large consulting squads.[5]
• Distyl: leadership pulled from Palantir, plus AI engineers/researchers from OpenAI and Apple.[3] The structure splits into forward-deployed talent (embed, own outcomes) and a central platform/research team that builds and hardens Distillery so field-learned patterns become reusable product.[2]

The recruiting pipeline: the Palantir "founder factory"

The hiring edge is a specific, documented talent pool. Business Insider calls Palantir's FDE role a "founder preparation bootcamp"[20]; the WSJ ran a feature on "The Palantir Mafia."[23] The skillset — embedded, outcome-owning, full-stack, client-facing — maps almost 1:1 onto both founding a DeployCo and staffing one.

What the platform actually does — and where the real IP is

Distyl's Distillery: SOPs → Routines → tasks

Distillery converts standard operating procedures into auditable AI workflows it calls Routines.[4] The pipeline, per Sacra's technical breakdown:

1. Ingest the SOP. A user uploads a written procedure ("approve refund requests," "process insurance claims").
2. Decompose. Distillery auto-breaks it into discrete tasks — validate order number → check warranty status → issue refund.
3. SME refinement. Subject-matter experts use a visual no-code builder to review and tune the auto-generated prompts and tool chains, attach test cases, and provide feedback examples — without programming.
4. Run with full auditability. In production the platform captures every input, output, tool call, and reasoning step; failed tasks trigger alerts or human-in-the-loop review; analysts download execution traces and feedback flows back as prompt adjustments and A/B tests.

It connects to enterprise data via SQL, APIs, and document repositories, with role-based access, multi-tenancy, and SOC-2 compliance.[4] Reported (Sacra)

Northslope's Palantir-native stack

Northslope inverts the build/buy decision: rather than own a substrate, it builds mission-specific AI applications on top of Palantir's Ontology (Foundry/Gotham) and AIP, plus Google Cloud's Gemini.[6] The FDEs "conceive, architect, build, maintain" the agentic layer; Palantir provides the deterministic data foundation. The trade: faster to credibility (you inherit a $400B+ platform's governance and data integration) but you don't capture the substrate margin — Palantir does.

The verification layer — the enabler of outcome pricing

This is the genuinely load-bearing piece, and the most misunderstood. Generating candidate answers from an LLM is cheap and commoditizing. The scarce asset is the verifier — the thing that turns a pile of cheap candidate answers into one trustworthy answer reliable enough to bill on. The verification hierarchy, strongest to weakest:

Distyl's case studies surface this layer in production language: prior-auth decisioning is "auditable… LLM writing down its own step-by-step logic… & source citations"[10]; the supply-chain tool uses "heuristics to triage… escalating complex cases to more powerful LLMs"[11]; the CPG assistant catches "errors mid-retrieval and route[s] to the right next action."[12] Distyl's platform explicitly markets evaluation (accuracy, latency, cost, robustness, explainability), observability, and guardrails against hallucination.[13]

Present both sides. The honest limit: a huge class of enterprise work — strategy, design, open-ended judgment — has no objective verifier and never will. There the only "verifier" is another model with no ground truth (Tier 3), the coverage curve can't be cashed in, and "own the outcome" becomes a liability. The durable versions of this business concentrate on the checkable quadrant (SQL/analytics, reconciliation, document-to-schema extraction, claims with system-of-record ground truth) and are explicit about what they won't guarantee.

The three revenue legs — and the metrics, labeled

DeployCos sell three things, and the strategic arc is shifting the mix from the first toward the second and third:

1. Forward-deployed services — embedded FDE teams. Services gross margins, typically ~30–50%. Estimated — not disclosed
2. Platform license — recurring access to the substrate (Distillery). Software gross margins, ~70–90%. Estimated — not disclosed
3. Outcome fees — a share of the fee tied to the customer's objective.

The metrics ledger

How outcome-based pricing actually gets structured

The clearest publicly-described mechanics come from the CX peers, which sit at the same outcome-pricing frontier. Sierra bills per successful resolution — the customer pays only when the agent resolves the issue end-to-end, and an escalation to a human is typically free.[24] Decagon offers a per-resolution option (billed only when the AI fully resolves with no human handoff) but notes most customers still choose per-conversation for budget predictability.[15] The DeployCo version is the same logic applied to back-office workflows: define the objective (cases approved, disruptions root-caused, dollars saved), instrument it with the verification layer, and tie a fee component to it. You cannot offer outcome pricing without the verifier — that is the dependency that makes the entire business model possible.

2nd order Outcome pricing transfers performance risk from buyer to vendor, which is exactly why it wins deals against incumbents who bill regardless of result. 3rd order But it also caps the addressable surface to verifiable work — and concentrates existential risk in the eval layer being right. A mis-calibrated verifier doesn't just lose a deal; it means the vendor guaranteed an outcome it can't actually measure.

Where the business actually concentrates — and the granular work

Sector concentration (not generic "F500")

Distyl's public footprint is over-indexed on healthcare payers — three of its six public case studies are health-insurer engagements — followed by telecommunications, financial services, manufacturing/supply chain, and CPG.[16] Its disclosed sector list: healthcare, telecommunications, insurance, manufacturing, and financial services[2], plus retail and federal/HHS per prior reporting. The named/credibly-reported logo is T-Mobile USA (the F100 telecom case)[3]; the other case-study customers are disclosed only by tier and industry ("F20 healthcare payor," "F50 hardware manufacturer") — named tiers, customers not disclosed. Northslope's reported verticals: solar/renewable energy, aerospace ("building rockets"), healthcare/hospitals, financial services (payments), retail, and defense/aviation.[5]

Why these sectors first? They share two traits the model needs: (1) high-volume, expensive, rule-governed workflows, and (2) a system of record (claims database, ledger, ERP, contract repository) that can serve as an objective verifier. Regulated industries also value the auditability the platform provides — which is precisely why prior-auth and contract decisioning, not open-ended strategy, are the beachhead.

The granular automations — before-state → what the agent does → measured result

These are the actual tasks, not abstractions. All metrics are company-disclosed via the vendors' own case studies Company-disclosed unless otherwise labeled.

① Prior-authorization review for a Fortune-20 health insurer [10]

• Before: Humans manually review each PA request, comparing it against evidence-based clinical policies — "manual, expensive, and time-consuming," with compliance risk and inconsistent quality.
• What the system does: Reviews the request plus associated medical-record data, compares against clinical policy, reasons to a decision — approve, or escalate to human review. Clinical SMEs refine the AI logic in natural language; the LLM writes down its step-by-step reasoning and source citations for an audit trail.
• Result: $200M+ estimated cost savings; 200k+ cases/month with accelerated approval; 90%+ decisioning accuracy; deployed in one quarter.

② Income verification & fraud detection for a public auto lender [8]

• Before: Analysts manually reviewed 50k+ dealer applications monthly, cross-checking W2s, 1099s, address proofs, IDs, and bank statements — 3 human touchpoints per approval, 8-hour manual turnaround, slow/inconsistent, missing fraud patterns.
• What the system does: Ingests applications via API, extracts income data from unstructured documents, verifies consistency across sources, flags fraud/compliance issues, and learns from analyst decisions using the lender's own job aids.
• Result: 93% cost reduction for loan origination; 5-minute agentic turnaround (from 8 hours); live in 1 week.

③ Supply-chain root-cause analysis for a Fortune-50 hardware manufacturer [11]

• Before: Planners and procurement teams manually search multiple systems, exchange emails, and hold calls to diagnose disruption root-causes — delays and inconsistent decisions.
• What the system does: Consolidates and explains demand-driven changes in natural language; uses heuristics to triage and escalates complex cases to more powerful LLMs; aggregates hundreds of unrelated alerts into a single actionable root cause.
• Result: 80% targeted reduction in root-cause-analysis time; 1,500+ disruptions root-caused daily; 30+ scenarios; built within a quarter.

④ Provider-contract decisioning for a Fortune-20 health insurer [28]

• Before: Network teams manually research contract terms, termination clauses, and renewal requirements one-by-one across dozens of tools — fragmenting provider intelligence and delaying action.
• What the system does: Converts 600k+ static contracts into structured, searchable data (rates, reimbursement methods, clauses, amendments); aggregates current state across amendments; links reimbursement logic to performance data.
• Result: $16M annual savings; 600k+ contracts made queryable.

⑤ Order-incompletion resolution for a Fortune-50 CPG brand [12]

• Before: Supply-chain specialists did time-consuming structured/unstructured data analysis to resolve order incompletions, slowing customer responses.
• What the system does: An agentic assistant translates natural language into queries across structured/unstructured/semi-structured stores (advanced RAG), navigates multi-step workflows, and catches errors mid-retrieval.
• Result: 47% improvement in resolution time; 100+ non-technical users enabled; junior operators onboard in weeks instead of quarters.

⑥ Retail-chain tax reporting (Northslope) [5]

• Before: A massive retail chain paid a high-seven-figure annual contract to a Big Four firm for tax reporting — still leading to mis-filed taxes and fines.
• What the team did: Sent two FDEs for three months and built custom, mission-specific software the internal team now runs itself — integrating all data, filing on time, responding to state disputes.
• Result: Eliminated the high-seven-figure annual contract; avoided penalties; dramatically reduced operational overhead. Founder interview

GTM tactics & where it breaks

GTM: founder-led brand and the ex-Palantir network; outcome guarantees as the wedge against incumbents; lab partnerships (Distyl's formal Services Alliance with OpenAI[17]) and SI/platform partnerships (Northslope as Palantir's first "Vanguard: Elite" partner[6]); and a deliberate focus on regulated industries where auditability is a feature, not overhead.

Where it stalls: (1) Fuzzy work — anywhere the outcome can't be objectively verified, the model can't honestly guarantee it. (2) Long-tail integration — wiring verifiers into a legacy estate is real, slow engineering; implementation complexity is Sacra's headline risk for Distyl.[4] (3) Outcome disputes — when the measured result is contested, "own the outcome" becomes a liability rather than a selling point.

Other companies running a similar embed-and-build motion

Six-plus independent companies run adjacent versions of the loop. The key bifurcation across the field: outcome-based billing clusters in productized single-vertical CX (Sierra, Decagon); FDE-embed clusters in services build-shops (8090, Northslope); Distyl is the rare one at clear unicorn scale that does both.

The ancestor, and who else wants this budget

Palantir as the FDE-model ancestor — and the genuine technical difference

Palantir pioneered the forward-deployed model in the mid-2000s and, per Everest Group, at one point ran more forward-deployed engineers than traditional product engineers.[21] But there's an architectural distinction the new wave must reckon with. Palantir's moat is the deterministic Ontology — a governed map of an enterprise's objects, relationships, logic, permissions, and actions (Foundry/Gotham), representing the enterprise's decisions, not merely its data.[34] LLM agents (AIP) are a probabilistic reasoning layer that operates on top of that deterministic substrate.

The competitive field (brief context — not the focus)

Lab arms — mostly coopetition. The model labs both enable these startups and build competing delivery muscle. Distyl runs a formal Services Alliance with OpenAI, whose COO called it "a key enabler for enterprises deploying OpenAI technologies at scale."[17] Yet OpenAI also lists its own Forward Deployed Engineer roles ($180K–$280K)[18], and a16z counted 22 of 311 open OpenAI roles as forward-deployed/solutions engineering.[17] The labs are partner and potential competitor at once.

Consulting / SI response. The incumbents are moving fast. EY named 8090 a founding partner of its AI Product Development Lifecycle, powered by 8090's agentic "Software Factory" (March 2026).[29] And the scale is real: Accenture disclosed $5.9B in GenAI new bookings in FY2025 (~7% of total bookings) and $2.7B in advanced-AI revenue.[35] Confirmed (filings) The Big Four/SIs are the deepest-pocketed competitors for the same enterprise budget — and the EY↔8090 deal shows their preferred answer is to partner with an AI-native build shop rather than build one from scratch. (Separately, Thrive Holdings is a different model — a Thrive Capital/OpenAI-backed AI-industrialized services roll-up, not an FDE integrator.[36])

The honest synthesis

The machine works, it's venture-scale, and it's increasingly understood. That last fact is the problem: a new entrant cannot win on the motion alone. Here is the founder's decision framed cleanly.

The three distinct ways to build one

What's commoditizing

• The motion itself. a16z has published the playbook[17]; ~224 FDE roles are open across ~39 companies[19] single-source snapshot; the ex-Palantir pipeline is well-mapped. "We embed FDEs and build on your data" is no longer differentiated.
• The orchestration glue. Routers, prompt chains, retrieval — open-source covers most of it. The frontier models keep getting cheaper and better, compressing the cost arbitrage on any single task over time.

Where the open lanes are

• Vertical wedges with a built-in verifier. The durable opportunity is a specific vertical where (a) an objective verifier already sits in the customer's stack (SQL/analytics, reconciliation, code/IaC, claims with system-of-record ground truth), and (b) the founder has unfair access or domain credibility. Distyl is over-indexed on healthcare payers; the un-embedded verticals are the opening.
• Substrate, if you can fund it. The only structurally defensible version owns the platform — but that's a Palantir-shaped, capital-intensive, services-heavy-early company. Be honest about whether you can finance the margin trough.
• The verifier as a wedge. Selling the evaluation/verification layer to enterprises building their own agents is a more product-led, possibly better risk-adjusted bet than a horizontal Distyl clone.