Palantir AIPCon 11: How Ontology Turns AI Agents Into Operational Handoffs Across 12 Enterprise Cases

The article examines 12 public case studies presented at Palantir's AIPCon 11 on September 10, 2026, distinguishing between live business applications, live demos, research prototypes, and partner deployment designs. The central theme: Ontology — Palantir's layer of objects, properties, relationships, actions, functions, and dynamic security — enables AI agents to operate inside core workflows while remaining authorized, constrained, verifiable, and traceable.

1. Customer Demand Changes: Insurance and Manufacturing Handoffs

Case 01 Acrisure: From Coverage Gap to Policy Service

Acrisure, grown through nearly 1,000 acquisitions, built Auris AI on Palantir Ontology and AIP to unify customer, carrier, contract, task, and document objects across advisory, underwriting, and service roles. A demo traced a florist adding two stores, three vans, and new staff: advisor Dominic sees renewal and growth prompts, reviews peer experiences, and decides to proceed; underwriter Luke receives bundled legacy system data and documents, gets flagged that the employee roster lacks new hires, completes the submission, and compares limits, premiums, deductibles, and carrier appetite side by side, recording the selection rationale; service rep Katie then receives endorsement, benefits, and document reconciliation tasks, with background agents handling duplicate paperwork and escalating exceptions. The key observation: a single customer object maintains state across three roles — what the advisor approved, what materials were missing, why the final option was chosen — without re-assembling context at each handoff. Acrisure notes the platform targets internal operations; cross-company interconnection remains a future direction. The demo shows workflow organization, not proven improvements in underwriting cycle, loss ratio, or profit.

Case 02 Eaton: Customer Requirements Drive Configuration, Materials, and Lead Time

Eaton builds custom electrical equipment. Its Equipment Forge links customer requirements, engineering standards, BOMs, suppliers, factory capacity, and committed lead times into a traceable digital chain with specialized agents. In the demo, a special arc-safety requirement triggers an engineering deviation agent that identifies the deviation from Eaton standards and historical decisions, passes constraints to a BOM availability agent, which finds a critical component with a potential one-year lead time. The owner evaluates a second supplier; the system recalculates quote and lead-time options along the requirement-configuration-supplier graph. The human makes the supplier switch; agents continue tracing downstream effects; the employee owns the delivery commitment. The incremental value is exposing a material shortage caused by a safety requirement before quoting, not merely helping engineers read PDFs. The demo shows how lead time and price change with alternatives; no average cycle-time improvement across all orders was disclosed.

Case 03 Hexion: Consumption Changes Propagate to Procurement and Pricing

Chemical maker Hexion organizes customer, product, inventory, order, transport, plant, raw material, supplier contract, and cost into an ontology so demand changes propagate along business relationships, not just as departmental alerts. An agent adjusts delivery plans based on tank level, consumption rate, delivery windows, vehicle scheduling, and order consolidation; confirmed decisions write back to the ontology and downstream systems. Increased consumption drives raw-material re-evaluation: a procurement agent reads phenol contract constraints, supplier quotes, global price indices, and even price updates buried in supplier emails, producing several sourced purchase options for human cost-risk comparison. Cost changes then feed finished-goods pricing and customer priority decisions. Hexion's CEO cited a $300M EBITDA improvement over three years from a broader transformation, explicitly not attributable to the demonstrated agent flow. The observable change is a cross-functional state-propagation chain across sales, supply chain, procurement, and commercial decisions; chemistry, production, and trade-offs remain with specialists.

2. Supply-Chain Decisions: Preserving Rationale for Reuse

Case 04 NVIDIA: Turning Planners' Tacit Experience into Reusable Decision Records

NVIDIA's multi-site factories compete for scarce GPU, CPU, and memory components. Materials, sites, committed volumes, capacity, allocations, and actual production results sit in a unified Palantir Foundry view. The cuOpt optimizer runs weekly key-material allocations and what-if scenarios (e.g., -10% memory, +1 site), tracking Time of Ownership — material dwell time after arrival. Optimizers have blind spots: planners read partner emails, weather forecasts, supplier call notes to spot risky capacity promises. NVIDIA and Palantir began recording planners' final allocations, why they accepted or overrode optimizer suggestions, expectations at the time, and actual outcomes. The ontology provides context for today's decisions and turns human judgment into auditable training material. These records trained Nemotron 3.5 Lightning, evaluated via point-in-time backtesting using only information then available. On a specific development set, the specialized model reached 86.7% accuracy on allocation decisions, Nemotron 3 Ultra 55.5%, base Lightning 17.5%. NVIDIA acknowledges future production risk prediction remains hard; final allocations stay with planners who accept, modify, or reject, with results written back; controlled retraining triggers only after sufficient accumulation; models do not self-retrain online. The transferable lesson: first let experts leave traceable decisions in the same flow, then train specialized models; without that decision log, larger models alone cannot replicate expertise.

Case 05 L3Harris: 33 ERPs Unified into One Operational Picture

Defense contractor L3Harris, after many acquisitions, scattered operations across 33 ERPs, hundreds of software systems, homegrown tools, and spreadsheets. It built an enterprise unified data layer on Foundry with ~3.5M data connections and 5,000+ planned data feeds, feeding applications like Sector Control Tower for executives and Program Digital Cockpit for project teams, placing materials, people, and schedule in a single project view to assess ripple effects of a constraint change. Example: glass-fiber shortage for circuit boards. Previously ~30 people across businesses spent months gathering affected programs; now the connected data surfaces impacted projects in ~10 minutes. This measures impact scoping, not shortage resolution. A narrower experiment: country-of-origin monitoring for parts; a fine-tuned open-source model on proprietary data beat the prior frontier model after <48 hours development at 95% lower model cost; not generalizable to all defense tasks. Hard prerequisite: if the same material lacks a unified definition across ERPs, agents cannot reliably answer "which projects are affected," let alone replan across projects. Cross-system correlation first, then applications and models consume that business-fact layer — the deployment sequence in this case.

Case 06 Elmet: 10 Operational Apps on Legacy ERP in Two Months

Elmet Group makes tungsten, molybdenum, and precision components for defense and energy. Constraint: legacy ERP must stay; resources insufficient for full replacement; purchasing, costing, scheduling, and shop-floor tracking cannot wait. Four Palantir engineers engaged mid-June; ~two months later, 10 apps live in two divisions. Team prioritized with purchasing and production leads, delivered integrated data and apps in weeks. Work included exploding BOMs to discrete costs, judging order profitability, seeing which machine a work order runs on and how to sequence, and how material purchasing supports delivery. Demonstrates a "cover current process first, then incrementally extend" factory app build path. The 10-app count is Elmet's self-disclosed figure; no unified savings, per-app adoption rates, or proof of ERP replacement provided.

3. From Risk Analysis to Edge Execution: Who Acts, Who Verifies

Case 07 FAA: Safety Risk Conclusions Land on Mitigation Actions

FAA's aviation safety team deals with weather, incident reports, ATC audio, and multi-agency histories. Even on a big screen, analysts must link anomalies to specific flights, aircraft, facilities, and procedures. Demo scenario: rising Houston-area airborne spacing incidents. System threads event types, evidence, and related objects; an investigation agent assists clustering commonalities; personnel evaluate results and decide next steps. Critical second half: risks and hazards become trackable objects; mitigations become ontology objects assignable to ATC or other units. Managers track whether mitigations are assigned, implemented, and whether risk declines post-implementation. The system connects "detect → analyze evidence → determine mitigation → execute → verify effect" — not letting agents issue direct flight operations commands. No quantified accident-rate or risk-reduction results disclosed; workflow demo ≠ safety efficacy validation. Palantir's Action submission conditions illustrate the generic technical bar: e.g., swapping an aircraft requires the actor to belong to a designated group and the selected aircraft to be airworthy; conditions unmet → submission blocked. This official mechanism diagram is not the FAA's actual UI, nor does it imply FAA adopted identical rules; it shows enterprises must combine identity and business state in action permissions.

Case 08 Ondas Sentinel: Edge Devices Retain Full Decision Packages

Ondas previously showed stratospheric mission planning cut from weeks to minutes; this time scope widened. Enterprise side: post-acquisition system integration, policy doc comparison, finance-ops collaboration, CRM, leasing, insurance management sharing one ontology. Ondas claims integration cycles compressed from typical 12-24 months to ~3 months — self-reported, not a universal baseline. Edge side: SkyWeaver mission demo — suspicious vessel lead triggers task adjustment agent checking current assets and positions, offloading heavy compute to higher node, returning plan to edge; balloon visual model detects target, fuses imagery and RF signals, selects suitable uncrewed assets with remaining endurance and acceptable weather for further reconnaissance. Entire mission saved as a "decision package" containing scheduling, detection, inference, and device actions; offline, evidence persists locally; on reconnect, syncs for human review. Key differentiator: how operational evidence is preserved during disconnection, not "full autonomy." Ondas says operator feedback and mission logs evaluate models, then scene-tested models redeploy to edge. Video shows one mission chain; no systematic true-positive/false-positive statistics provided.

4. Readers, Customers, and Research Hypotheses Enter Workflows

Case 09 USA TODAY: Orchestrating Explainable Reading Experiences Around Readers

USA TODAY runs national content plus 200+ local outlets with reading, search, video, subscription first-party data previously siloed. Demo with pseudonymous reader Marcus: ontology links interests, reading history, subscription status, content, and interactions; agents select and rank personalized newsletter from existing news pool, place related recommendations and ad slots; editors see why each article was chosen. As reader continues reading, clicking, or giving feedback, behaviors write back; next email and next article page adjust accordingly. USA TODAY emphasizes reader data anonymized; AI handles discovery and presentation, not news writing; journalism standards unchanged. Business object: the evolving reader-article relationship. Conference delivered an explainable personalization flow; no independent retention or subscription lift metrics provided.

Case 10 Zeta: Placing Customer Insight Inside the Enterprise's Own Operating Context

Marketing firm Zeta has customer data and Athena platform for audience analysis (who, what they might like). But enterprises must decide which markets to enter, which budgets to allocate to whom, how to adjust existing campaigns — decisions requiring the enterprise's own product, customer value, budget, and placement results. Announced partnership connects Zeta's audience intelligence with Palantir-hosted enterprise operating context: identify growth opportunities, cross-reference existing campaigns and customer priorities, feed recommendations back into enterprise decision flows. Zeta disclosed its existing customer scale and marketing ROI figures — those reflect Zeta's current business, not gains from the new integration. Case is currently a joint solution: Zeta brings audience intelligence and activation; Palantir brings enterprise business objects and workflows; brand retains data, privacy, and decision control. Public demo insufficient to prove full deployment across all partner clients.

Case 11 Novartis: Research Agents Analyze Hypotheses Under Check Rules

Novartis' Data 42 on Foundry holds 3,000+ clinical trials, 1M+ patient records, genomics and proteomics. Challenge beyond "find data": trial design, measurement definitions, cleaning methods must be correctly understood by later researchers. Fractal research prototype gives agents curated, versioned knowledge context and breaks multi-omics analysis into steps with explicit inputs/outputs. Speaker said a manual 3-day analysis completed in 45 minutes; then showed a drug-indication hypothesis exploration prototype on the actual data lake covering 41,000+ candidate combinations. Every agent step has checks: output matches expectation, reasoning has scientific basis; full session saved; when researchers find dataset noise they add rules and reapply to broader analyses. Novartis explicitly calls this a prototype; scientists and review committees interpret evidence and make trade-offs. 45 minutes is one analysis duration; 41,000 is hypothesis space size; neither equals higher drug development success rates. Illustrates that business objects aren't only customers and orders — they can be evidenced, versioned, audited research hypotheses.

5. Cisco Addresses Deployment; The 12 Cases' Common Boundary

Case 12 Cisco: Delivering Security and IT Ontology in Controlled Environments

Cisco announced partnership with Palantir and NVIDIA: plans to deliver Palantir cybersecurity and IT ontology via "Cisco Secure AI Factory with NVIDIA" as the preferred full-stack foundation for "Palantir Sovereign AI OS." Cisco provides observable, secure, uniformly managed infrastructure; NVIDIA provides Nemotron models post-trainable on enterprise data; Palantir contributes Foundry, AIP, and ontology for business objects, permissions, and actions. Three parties aim to bundle compute, network, storage, security, observability, and AI workflows into a reference architecture. Unlike Acrisure or Hexion's customer flows, this is primarily a partnership and deployment design. Cisco's post gives no cost, throughput, or security gains realized at all customer sites. It raises unavoidable questions: who holds insurance data, defense supply chains, regulatory records; where models run; who accesses; how to trace issues. Agents that "write back to business" must first operate in enterprise-allowed environments.

Cross-Case Patterns and Limits

The 12 cases do not converge on a universal agent. Acrisure uses one customer object to bridge roles; Eaton and Hexion thread one demand change through configuration, supply, and cost; NVIDIA keeps planners' rationale in decision logs; FAA and Ondas leave mitigation or task end-state to responsible owners for verification; media, marketing, and research scenarios build their own objects around readers, customers, hypotheses. Common engineering sequence: first clarify objects and responsibilities, then design actions agents may submit, then verify results actually enter the next step. Boundaries to note: Elmet's 10 apps ≠ 10 autonomous agents; NVIDIA's 86.7% applies only to the allocation dev set; Novartis is a research prototype; Zeta and Cisco carry clear partnership-solution attributes. Palantir's ontology has long supported reads and actions; the new signal from this conference is more enterprises placing it at the center of long-running processes, letting people, models, and business systems share one continuously changing record. The real acceptance test therefore is: who approved what action, which business object did it change, can the next role pick it up, and how was the outcome later checked.