Executive Summary
- The single most expensive pre-deployment decision in enterprise AI is not the tool — it is whether a workflow requires a full data foundation investment before the AI can function, or whether it can operate on existing data structures with minimal preparation. Getting this wrong costs 6–18 months and $300K–$1.5M in remediation that wasn’t budgeted.
- The corpus contains both extremes. Nebraska Medicine built a new revenue cycle automation in 10 hours — on top of a 6-month Ontology that cost a full enterprise implementation first. SlickDeals reduced deal-scoring latency 360x and grew merchant revenue 7% — by rebuilding a pipeline within a single data domain (user-generated deal content + community engagement signals) with no semantic integration layer required.
- A clear decision rule separates these two outcomes: single-domain, structurally coherent workflows with one input type and one output type rarely require a data reset. Multi-domain workflows — those that cross patient/procedure/payer, or order/inventory/supplier — require semantic integration before AI can operate reliably. Domain count is the primary decision variable. Data structure type is the secondary variable. Company size and data maturity moderate the timeline and cost.
- The 60% AI project abandonment rate (Gartner, Feb 2025, n=248) is not evenly distributed across workflow types. It concentrates in multi-domain deployments where the data pre-condition was not assessed before the project was approved. The SlickDeals outcome is reproducible at mid-market scale. The Nebraska Medicine outcome requires the Ontology investment first — and that investment does not transfer from vendor to vendor.
- A four-variable diagnostic — domain count, data structure type, decision-point density, and data maturity score — classifies any planned AI workflow as Proceed, Prepare, or Reset. This classification should be completed before vendor selection, not after.
The Two Poles: Nebraska Medicine vs. SlickDeals
Understanding the data reset question requires holding two real deployments in tension simultaneously.
Nebraska Medicine: Multi-Domain Requires Foundation First
Nebraska Medicine deployed Palantir AIP to automate revenue cycle medical necessity validation — a workflow that requires cross-referencing patient demographics, procedure codes, payer rules, and physician credentials simultaneously. The AI needed to evaluate whether a specific procedure was medically necessary for a specific patient under a specific payer’s current rule set.
The result: a new guideline evaluation workflow built in 10 hours. That headline is accurate. What the headline omits: the 10-hour build sat on top of a 6-month enterprise-scale Ontology deployment. The Ontology unified patient records, procedure data, payer contracts, and physician data into a single semantic representation that the AI could query coherently. Without the Ontology, the 10-hour build was not possible — the data existed in separate systems with incompatible schemas, separate ownership, and no shared identity layer.
Nebraska Medicine’s deployment is not a testimonial to AI speed. It is a testimonial to what compounding returns from data infrastructure investment look like. The first use case cost 6 months. The second cost 10 hours. Organizations that evaluate only the 10-hour figure without asking “what came before it” will budget AI projects that stall for the missing foundation.
Data signature: Patient + procedure + payer + physician = four domains, four separate source systems, four different data ownership functions, one required semantic integration layer.
SlickDeals: Single-Domain, Structured Coherence, No Reset Required
SlickDeals rebuilt its AI deal-scoring and personalization stack between 2023–2025, replacing a 3-hour batch ETL pipeline with a real-time event processing architecture. Deal scoring latency dropped from 3 hours to 30 seconds (360x). Merchant outbound clicks and revenue increased 7%. Homepage personalization produced 70%+ higher product detail page views.
The key structural fact: SlickDeals’s entire AI stack operates within a single data domain. The input is deal content — text, category, discount percentage, brand, community votes, engagement velocity. The output is a ranked list of deals for a given user. All data lives in one coherent system with a single ownership function (engineering). No cross-domain identity resolution was required. No semantic integration layer was needed because there was nothing to integrate.
The architecture decision that unlocked results was infrastructure modernization (SQL Server → Databricks, LAMP VMs → EKS), not a semantic data layer. The data was already structurally coherent. The AI needed latency and real-time signal, not a new data model.
SlickDeals’s outcome is directly reproducible at mid-market scale — by any company with a single, coherent, high-velocity data domain. A law firm’s document review queue. A manufacturer’s quality inspection images. A financial services firm’s transaction approval log. Each of these is a single-domain, structured problem that does not require a data reset — it requires infrastructure modernization and model deployment.
Data signature: One data type (deal content + user behavior signals), one source system, one ownership function, one output type. No cross-domain resolution required.
The Decision Framework: Four Variables
Variable 1: Domain Count (Primary)
Count the number of distinct business entities the AI workflow must cross to function correctly. A “domain” is a category of business data that has its own source system, its own ownership function, and its own data quality standards.
| Domain Count | Implication | Likely Data Requirement |
|---|---|---|
| 1 domain | Single-entity input → single-entity output | Infrastructure modernization only; Proceed if data is structurally coherent |
| 2 domains | Cross-domain join required (e.g., customer + transaction) | Prepare: entity resolution between two systems, typically 4–8 weeks |
| 3–4 domains | Complex integration (e.g., patient + procedure + payer + physician) | Reset: semantic integration layer (Ontology or equivalent) required; 3–6 months minimum |
| 5+ domains | Enterprise-wide data model required | Full reset; 12–24 months for most mid-market organizations |
Corpus anchors for this calibration:
- 1 domain → Proceed: SlickDeals (deal content + user signals, one system, 360x latency improvement with infrastructure modernization only)
- 1 domain → Proceed: TTEC call center AHT reduction (structured call logs, one system, 15–20% handle time reduction per IBM case study data)
- 2 domains → Prepare: Vodafone VOXI customer service (conversation data + intent classification layer; required NLP pre-processing before conversational AI could function)
- 3+ domains → Reset: Nebraska Medicine revenue cycle (patient + procedure + payer + physician; 6-month Ontology before 10-hour build)
- 3+ domains → Reset: Global bank transaction monitoring (customer + transaction + counterparty + fraud signal; 60% faster resolution after semantic integration layer per Palantir case library — timeline not disclosed)
- 5+ domains → Full reset: Fortune 100 CPG 7-ERP digital twin (7 separate ERP systems; 5-day Palantir Bootcamp build preceded by enterprise data unification — the Bootcamp was the last step, not the first)
Variable 2: Data Structure Type (Secondary)
After domain count, the type of data determines whether preparation is linear (moderate effort) or non-linear (high effort regardless of tools applied).
| Data Type | Preparation Required | Typical Timeline |
|---|---|---|
| Structured, single schema | Schema validation + dedup | 1–4 weeks |
| Structured, multi-schema | Entity resolution + schema normalization | 4–12 weeks |
| Semi-structured (JSON, XML, logs) | Parsing + field standardization | 2–8 weeks |
| Unstructured, consistent format (call transcripts, legal documents of same type) | NLP pre-processing + embedding | 4–12 weeks |
| Unstructured, mixed formats (mixed document types, handwritten + digital) | Classification + OCR + normalization | 3–6+ months |
| Unstructured, undocumented business rules | Classification + rule elicitation + validation | 6–12+ months |
The critical frontier is undocumented business rules. When the logic for a decision lives in the head of a senior analyst — not in a documented policy, not in code, not in a data dictionary — it cannot be automated until it is documented. This is the bottleneck that tools cannot solve. IBM and Dataversity have confirmed that LLM-based data quality assistants hallucinate when source metadata is sparse — exactly the condition that exists when business rules are tribal.
The 26% figure from IBM’s Chief Data Officer survey is the most direct marker of this constraint: only 26% of organizations can use unstructured data in a way that delivers business value. For the 74% that cannot, unstructured workflows require either a Reset or a scope reduction to the structured sub-portions of the workflow.
Variable 3: Decision-Point Density (Tertiary)
High-decision-point workflows — where the AI must make or support many distinct judgment calls within a single workflow — amplify the data requirements of both Variables 1 and 2. Low-decision-point workflows can tolerate lower data quality because errors in one step do not cascade.
| Decision Type | Data Requirement |
|---|---|
| Binary classification (fraud / not-fraud, approve / deny) | Lower data quality tolerable; errors are discrete |
| Ranked scoring (deal quality, lead priority) | Moderate — ranking errors are graceful degradations |
| Multi-step reasoning (medical necessity requiring rule lookup + patient history + payer policy) | High — each step error multiplies into next step |
| Agentic action (AI takes downstream action based on decision) | Highest — errors have real-world consequences; requires highest data quality |
The Nebraska Medicine workflow is high-decision-point: the AI evaluates a specific procedure code against a specific patient history against a specific payer’s current clinical policy. Each element requires current, accurate data. An error in the patient domain propagates into a false approval or false denial.
The SlickDeals workflow is ranked scoring: errors in deal ranking are graceful — a 30-second-old score for a deal with 48 hours of inventory is directionally correct even if not perfectly precise. The system tolerates moderate error in exchange for latency.
Variable 4: Data Maturity Score (Moderating)
The Innoflexion Data Readiness Index provides the most operational threshold calibration in the corpus: above 0.75 supports autonomous AI; 0.50–0.74 requires human review checkpoints; below 0.50 demands structured remediation before production use.
For organizations without a formal DRI assessment, the following proxy questions approximate the score:
- Can the relevant data be accessed programmatically from one location without manual extraction? (Yes = proceed signal)
- Does the data have documented field definitions, data types, and business rules? (Yes = proceed signal)
- Has the data been used in a production system that other teams rely on? (Yes = moderate maturity signal)
- Does the data require a subject-matter expert to interpret correctly? (Yes = structured preparation required)
- Has the data ever been profiled for completeness, uniqueness, and format consistency? (No = Reset signal)
A “No” on question 5 combined with 3+ domains is the strongest predictor of an unbudgeted data reset in the corpus.
The Three Classifications: Proceed, Prepare, Reset
Combining the four variables produces three deployment classifications.
Proceed: Deploy with Infrastructure Modernization Only
Profile: Single domain, structurally coherent data, low-to-moderate decision-point density, DRI ≥ 0.50.
What it requires: Infrastructure work — latency reduction, real-time processing where batch existed, model serving infrastructure — not data model redesign.
Corpus examples:
- SlickDeals deal scoring (1 domain, structured, ranked output, 360x improvement)
- Legal document review within one document type (contract clauses → risk flags; same structure throughout)
- Manufacturing quality inspection images (one data type, one decision type: defect / no-defect)
- Call center handle-time reduction on structured call logs (TTEC profile; one input type)
Budget and timeline: Infrastructure modernization ($50K–$300K depending on scale), model selection and deployment ($20K–$100K). Total: $70K–$400K. Timeline: 6–16 weeks to production.
Caution: Proceed classification assumes the data is structurally coherent. If the data has never been profiled and the business rules are undocumented, run a 2–4 week data assessment before committing to this classification. The 60% abandonment rate concentrates in cases where Proceed was assumed without verification.
Prepare: Targeted Data Remediation Before Deployment
Profile: Two domains or one domain with mixed schemas, semi-structured or mixed-format data, moderate decision-point density, DRI 0.40–0.74.
What it requires: Targeted entity resolution between two systems, schema normalization, NLP pre-processing for unstructured data, business rule documentation for one domain. This is a scoped data project, not a company-wide data program.
Corpus examples:
- Customer service AI requiring customer profile + transaction history (two domains, entity resolution required)
- Vodafone VOXI conversational AI (one domain — customer service transcripts — but unstructured, requiring NLP intent classification before routing AI could function)
- Marketing campaign AI requiring customer data + campaign performance data (two domains, different ownership)
- Contract AI pulling from multiple document types within one function (legal; mixed formats but single ownership function)
Budget and timeline: Targeted data remediation ($80K–$400K) + AI deployment ($20K–$150K). Total: $100K–$550K. Timeline: 3–5 months to production.
Common mistake: Scoping Prepare as Proceed and discovering the second domain at implementation time. The entity resolution problem — matching customer records across a CRM and a transaction system that don’t share a common key — is a 4–8-week engineering project that is not visible until someone tries to join the tables. The Gartner finding that 83% of data migrations fail or overrun traces partly to this scoping failure.
Reset: Semantic Integration Layer Required
Profile: Three or more domains, multi-schema structured or unstructured data, high decision-point density, DRI below 0.50 on one or more required domains.
What it requires: A semantic integration layer — an Ontology (Palantir model), a data mesh with domain ownership (Thoughtworks/Dehghani model), or a data fabric (enterprise only) — that creates a unified representation of multiple business entities. This is a 6–18 month foundational investment before the AI deployment begins.
Corpus examples:
- Nebraska Medicine revenue cycle (patient + procedure + payer + physician; 6-month Ontology)
- Fortune 100 CPG 7-ERP digital twin (multi-system integration; 5-day Bootcamp was final step after data unification)
- Supply chain AI requiring supplier + inventory + demand + logistics (four domains with different ownership functions)
- Any workflow requiring real-time cross-system decision-making with high consequence per decision
Budget and timeline: Semantic integration layer ($300K–$2M+ for enterprise, $150K–$800K for mid-market with reduced domain count) + AI deployment ($50K–$300K). Total: $450K–$2.3M. Timeline: 9–24 months to full production.
Critical context for mid-market: Most 200–2,000 person companies do not have the data infrastructure that AIPCon case studies presuppose. The AIPCon customer list skews Fortune 500 with mature data teams. For a 500-person company facing a Reset classification: the correct response is not “we need Palantir” — it is “we need to scope the data project first, determine whether this workflow is worth the foundation investment, and pick a smaller workflow as the first deployment while the foundation is built.”
The IBM Tech Debt Reckoning finding is directly applicable: organizations that scope tech-debt remediation into the AI business case from the start project +29% higher ROI than those that discover it mid-project. A Reset classification discovered at pilot time costs 18–29% of total budget in unplanned remediation.
The Four Workflow Archetypes That Succeed Repeatedly
Across the corpus, four workflow types succeed at higher rates than others. Each has a predictable data profile and a predictable data pre-condition.
Archetype 1: High-Volume Structured Decisions
Examples: Fraud detection, credit scoring, insurance underwriting triage, deal quality scoring, inventory reorder triggering.
Data pre-condition: Structured transaction data within one or two owned systems. Entity resolution between two systems at most.
Why it works repeatedly: The input is numeric or categorical, the decision is binary or ranked, the ground truth is observable (fraud confirmed, credit paid back), and the error cost is bounded. BCG data (n=~900 workflows) shows insurance claims validation at 50% scaling penetration and 25–32% cost savings — the highest penetration and most consistent outcome in the workflow dataset.
Data classification: Proceed (if single source system) or Prepare (if cross-system join required).
Archetype 2: Knowledge Retrieval Bottlenecks
Examples: Legal research, customer support response, technical support knowledge base, internal policy lookup, clinical guideline retrieval.
Data pre-condition: Structured or semi-structured documents within one function’s ownership. No cross-entity joins required for the retrieval step — only document corpus.
Why it works repeatedly: RAG (retrieval-augmented generation) requires only document indexing, not semantic integration across business domains. The AI retrieves relevant passages; humans validate and act. The Mallesons hybrid HOTL model (96% adoption, 20% cycle-time reduction) is this archetype — the AI retrieves, the lawyer decides.
Data classification: Proceed if document corpus is coherent and permissioned. Prepare if documents are mixed formats or lack metadata. Reset is almost never required for pure retrieval.
Caution: Legal AI hallucination rates of 17–33% (Stanford/Magesh, JELS, 2025) apply when retrieval is misconfigured or the knowledge base is incomplete. The failure mode is not the workflow archetype — it is insufficient knowledge base curation.
Archetype 3: Document Synthesis Before Human Decision
Examples: Revenue cycle medical necessity validation, contract review before negotiation, insurance claims assessment before adjuster review, financial close variance analysis before CFO review.
Data pre-condition: This is the most variable archetype. When documents come from one function (contracts in legal, claims in insurance operations), Prepare is often sufficient. When the AI must synthesize across entities (patient + procedure + payer), Reset is required.
Why it works repeatedly when data is ready: The AI performs the time-consuming preparation work; the human makes the binding decision. Error consequence is bounded by the human review step. The Anthropic legal team case (marketing review from 2–3 days to 24 hours) is this archetype — single domain (marketing content), structured review process, human approves.
Data classification: Proceed or Prepare for single-function document synthesis. Reset for cross-entity synthesis (the Nebraska Medicine pattern).
Archetype 4: Monitoring at Inhuman Scale
Examples: Supply chain disruption signals, transaction fraud pattern detection at high volume, infrastructure anomaly detection, competitive pricing monitoring.
Data pre-condition: Typically single or two-domain; the challenge is volume and latency, not semantic integration.
Why it works repeatedly: Humans cannot monitor at the required scale. AI enables a monitoring function that literally did not exist before. Error cost is bounded by low false-positive rates for high-severity signals and human review for medium-severity signals.
Data classification: Proceed if event stream is structured and coherent. Prepare if data sources are heterogeneous (e.g., multiple feed formats in supply chain). Reset is rarely required for monitoring use cases — the domain is usually defined by the asset being monitored.
Where Data Reset Is Waste vs. Necessity
The BCG finding — companies that spread AI across 100 use cases fail while those concentrating on 1–3 domains succeed — directly implies a prioritization principle for the Reset decision.
Reset is a necessity when:
- The workflow crosses 3+ business domains and the AI’s decision quality depends on all of them simultaneously
- The workflow affects high-consequence decisions (patient care, large financial transactions, regulatory compliance)
- The company is building a platform that multiple future workflows will share — the Palantir Ontology model compounds across use cases, so the Reset investment is amortized
Reset is waste when:
- The workflow can be scoped to a single domain and still deliver 80%+ of the intended value
- The company is in Year 1 of AI deployment and the Reset investment would crowd out budget for proven single-domain deployments
- The vendor requires a full data model to deploy but the workflow could be handled by a narrower tool without the integration requirement
- The Reset timeline (12–24 months) exceeds the executive patience horizon for the use case — a Reset that isn’t finished produces zero ROI, while a Proceed on an adjacent simpler workflow produces results in weeks
McKinsey’s domain concentration finding — that high performers concentrate AI in 1–3 business domains rather than spreading across the enterprise — is the empirical argument against company-wide Resets as a starting posture. The correct sequence for most mid-market organizations is: pick the highest-value Proceed or Prepare workflows first, generate real ROI, use that credibility and budget to fund targeted domain-specific Resets for the highest-value cross-domain workflows.
Deloitte’s future-built vs. laggard data shows this sequencing pays off: organizations that invest in data foundations before scaling reach production with 62% of AI initiatives vs. 12% for laggards, and compress time-to-impact to 9–12 months vs. 12–18 months. The Deloitte data does not say “do all the data work first” — it says organizations that treat data infrastructure as a staged investment rather than an afterthought reach production faster on the workflows they prioritize.
Key Data Points
| Data Point | Figure | Date | Source | Credibility |
|---|---|---|---|---|
| Nebraska Medicine workflow build time after Ontology | 10 hours | Sept 2025 | Palantir/Nebraska Medicine (co-published) | MEDIUM |
| Nebraska Medicine foundational Ontology deployment | 6 months | 2024–2025 | Palantir/Nebraska Medicine | MEDIUM |
| SlickDeals deal-scoring latency improvement | 360x (3 hrs → 30 sec) | 2025 | AWS re:Invent (named speaker) | HIGH |
| SlickDeals merchant revenue increase | 7% | 2025 | AWS re:Invent (named speaker) | MEDIUM-HIGH |
| Palantir Bootcamp to-contract conversion rate | ~75% | 2025 | Palantir-reported | MEDIUM |
| AI projects abandoned through 2026 without AI-ready data | 60% | Feb 2025 | Gartner (n=248) | HIGH |
| Data migrations that fail or overrun | 83% | 2025 | Gartner | HIGH |
| AI initiatives reaching production — data-ready vs. laggard | 62% vs. 12% | 2025 | Deloitte | MEDIUM-HIGH |
| Unplanned data remediation % of AI implementation budget | 18–29% | 2025 | IBM IBV (n=1,300) | MEDIUM |
| ROI uplift from pricing tech-debt into business case | +29% | 2025 | IBM IBV (n=1,300) | MEDIUM |
| CDOs who can use unstructured data for business value | 26% | 2025 | IBM CDO Survey | MEDIUM |
| Insurance claims validation: penetration and cost savings | 50% scaled, 25–32% savings | 2025 | BCG (n=~900 workflows) | MEDIUM |
| Data prep as % of AI project time | 60–80% | 2025 | O’Reilly | MEDIUM |
| Innoflexion DRI: autonomous AI threshold | ≥ 0.75 | 2025 | Innoflexion | MEDIUM |
What This Means for Your Organization
The data reset decision is not a technology question — it is a scope question. Before approving any AI deployment budget, three questions define the classification:
One: How many business entities does this workflow cross? Count the distinct systems, ownership functions, and data types the AI must touch to make one decision. If the answer is three or more, budget the Ontology or integration layer before budgeting the AI. If the answer is one, move to infrastructure modernization.
Two: Can the workflow be scoped down to one domain and still deliver 80% of the value? The highest-ROI deployments in the corpus are not the most ambitious ones — they are the ones that solved a specific, bounded problem completely. A law firm that deploys AI for contract clause extraction within its own document management system (one domain) will see results faster than one that tries to cross-reference client history, external case law, and internal precedent simultaneously before the data infrastructure supports it.
Three: Does the Reset investment compound? The Nebraska Medicine Ontology is worth the 6-month investment because every subsequent use case at Nebraska Medicine builds on it. At 139% net dollar retention, Palantir’s commercial customers are confirming this with their wallets — they expand spend because each new use case is faster and cheaper than the last. If your organization is building a first AI application with no planned follow-on use cases, a full Reset is almost certainly the wrong investment. If you are building a platform for a function that will deploy 10+ AI applications over 5 years, the Reset cost is amortized across all of them.
The Atlan 200-deployment analysis holds: the median deployment that follows the right sequence (data ready first, workflow redesigned, tool scoped to match) achieves +159.8% ROI. The deployments that skip the sequence do not reach that outcome — they reach cost overruns and abandoned projects, contributing to the 60% Gartner predicts will be abandoned through 2026.
If this raised a specific question about your organization’s workflow candidates or data infrastructure state, I would welcome the conversation — brandon@brandonsneider.com.
Sources
All sources are from existing corpus research. No new primary source fetching was conducted for this synthesis pass.
-
Palantir AIPCon / Nebraska Medicine — Nebraska Medicine Guideline Evaluation workflow (10-hour build, 6-month Ontology). MEDIUM credibility (vendor + customer co-published). See:
research/01-ai-native-landscape/palantir-aipcon-enterprise-agentic-deployment-2026.md -
SlickDeals / AWS re:Invent 2025 — Mike Lively (SVP Engineering) on 360x latency improvement and 7% merchant revenue increase. HIGH credibility for operational metric (named internal speaker, public conference). See:
research/02-corporate-tools/slickdeals-ai-deal-scoring-personalization-2025.md -
Gartner — 60% of AI projects abandoned through 2026 without AI-ready data (Feb 2025, n=248 data management leaders); 83% of data migrations fail or overrun. HIGH credibility (independent analyst). See:
research/07-adoption-challenges/data-readiness-investment-roi.md -
Deloitte — Future-built organizations: 62% AI initiatives reach production vs. 12% for laggards; 9–12 vs. 12–18 months time-to-impact. MEDIUM-HIGH credibility (vendor caveat; consistent with independent findings). See:
research/07-adoption-challenges/data-readiness-investment-roi.md -
IBM IBV “The Tech Debt Reckoning” — n=1,300 AI decision-makers: +29% ROI uplift when tech debt is priced into business case; 18–29% of AI implementation budget absorbed by unplanned remediation. MEDIUM credibility (IBM Consulting vendor caveat; self-reported projections). See:
research/04-consulting-firms/ibm-ibv-tech-debt-reckoning-2026.md -
IBM CDO Survey 2025 — Only 26% of organizations can use unstructured data in a way that delivers business value. MEDIUM credibility (IBM vendor caveat). See:
research/04-consulting-firms/ibm-ibv-tech-debt-reckoning-2026.md -
BCG Widening AI Value Gap 2025 — 70% of AI value in core business functions; insurance claims validation at 50% penetration and 25–32% cost savings; companies spreading AI across 100 use cases fail vs. 1–3 domain concentrators. MEDIUM credibility (consulting vendor caveat). See:
research/07-adoption-challenges/bcg-widening-ai-value-gap-2025.md -
O’Reilly 2025 — Data preparation = 40–60% of AI project cost, 60–80% of project time. MEDIUM credibility (practitioner survey). See:
wiki/data-readiness.md -
Innoflexion Data Readiness Index — DRI thresholds: ≥0.75 for autonomous AI, 0.50–0.74 requires human review, <0.50 requires remediation. MEDIUM credibility (practitioner framework). See:
wiki/data-readiness.md -
Atlan 200-deployment analysis — Median +159.8% ROI conditional on workflow redesign and data preparation preceding deployment. MEDIUM credibility (vendor research, but cross-referenced). See:
research/07-adoption-challenges/ai-workflow-selection-roi-definition-framework.md -
Stanford / Magesh JELS 2025 — Legal AI hallucination rates 17–33% even with retrieval grounding. HIGH credibility (peer-reviewed). See:
wiki/data-readiness.md -
Palantir Q4 2025 Earnings — 139% net dollar retention; 571 US commercial customers (49% YoY growth); $4.38B remaining US commercial deal value. HIGH credibility (SEC-filed). See:
research/01-ai-native-landscape/palantir-aipcon-enterprise-agentic-deployment-2026.md
Brandon Sneider | brandon@brandonsneider.com April 2026