On June 10, 2026, a four-person startup called Poetic announced a $50 million Series A at a $500 million valuation, led by Kleiner Perkins with participation from Founders Fund, First Harmonic, and OpenAI. The number that should keep every enterprise automation vendor awake tonight is not the valuation. It is 99%—the accuracy Poetic claims on compliance, fraud, and underwriting processes that Fortune 500 financial institutions have failed to automate for decades. In an industry where an 82% AI project failure rate is the norm and the average failed project costs $11.3 million, Poetic is making a contrarian bet: the future of enterprise automation is not more intelligence. It is more determinism.
The company, formerly known as Forge, was founded by Markie Wagner, a Stanford-trained machine learning engineer who previously worked at Google and Waymo. Wagner built Poetic around a deceptively simple thesis: autonomous AI agents—the kind that reason, improvise, and consume hundreds of thousands of tokens per session—are the wrong architecture for high-stakes enterprise processes where a single error triggers regulatory penalties. Instead, Poetic built a proprietary programming language that converts natural-language process definitions into deterministic, near-tokenless execution. The result is software that, in Wagner's words, "learns like AI but runs like code."
What Changed: The $50M Round and Why OpenAI Invested
The funding round itself signals a strategic inflection. OpenAI's participation is particularly telling—the company that built the most powerful probabilistic AI models on Earth is investing in a startup that deliberately avoids using them for execution. Leigh Marie Braswell, partner at Kleiner Perkins, stated that Poetic "executes complex processes with accuracy that exceeds what human teams can deliver."
The traction justifies the valuation premium. Poetic reached an eight-figure annualized revenue run rate in 2025 with just four employees—an efficiency metric that dwarfs typical enterprise SaaS benchmarks. Named customers include SoFi, where the platform achieved 99% quality in fraud investigations within five weeks; AIG, where it hit 99% accuracy on labor-intensive insurance operations; and Chime. The company reports a 100% pilot-to-production conversion rate—a metric that stands in stark contrast to the industry average where only 8% of enterprise AI projects deliver meaningful ROI.
One Fortune 500 financial services customer reported $200 million in annual fraud detection savings through the platform. The system also self-repairs when underlying software changes, eliminating the maintenance burden that consumes 60–75% of traditional RPA budgets.
Why This Matters
For CTOs and CIOs: The Deterministic Architecture Shift
The enterprise automation market is fracturing along a fundamental architectural fault line: deterministic versus probabilistic execution. Traditional AI agents—built on large language models—are probabilistic by nature. The same input can produce different outputs across runs. For customer service chatbots, that variability is acceptable. For compliance investigations, fraud adjudication, and insurance underwriting, it is not.
The math is unforgiving. Lusser's Law dictates that chaining three AI components at 90% reliability each reduces end-to-end reliability to approximately 73%. A fourth component drops it below 66%. In financial services, where bias is detected post-deployment in 31% of production models and regulatory concerns emerge an average of 3.2 months after deployment, compound unreliability is an existential risk.
Poetic sidesteps this entirely. By encoding expert workflows into deterministic execution paths—rather than asking an LLM to reason through each step—the platform delivers 100% process adherence. No token consumption variability. No hallucination risk. No prompt injection surface. For CIOs evaluating automation platforms under SR 26-2, the Federal Reserve's revised model risk management guidance issued April 2026, deterministic execution is not a feature preference—it is a compliance requirement.
For CFOs: The Token Cost Crisis Has a Solution
The economics of AI agents are broken for high-volume enterprise processes. According to EY's analysis of agentic AI token costs, a single agentic workflow can consume hundreds of thousands of tokens per session, inflating the cost of what was once a $0.04 chat interaction to $1.20 per orchestration—a 30x increase. At enterprise scale, this creates what EY calls a seven-layer cost structure: tokens, licenses, infrastructure, governance overhead, organizational change, failure recovery, and potential regulatory taxes.
Poetic's near-tokenless execution model collapses that cost structure. By compiling natural-language process definitions into deterministic code rather than running them through an LLM at inference time, the platform eliminates the variable token costs that make traditional AI agents 50–500x more expensive per transaction than RPA. For a CFO managing a $35 billion RPA portfolio—the current global market size—Poetic represents a third path: smarter than RPA, cheaper than agents.
Market Context: The Three-Way Automation War
The enterprise automation landscape in 2026 has split into three competing paradigms, each with distinct cost profiles, accuracy guarantees, and governance models.
Legacy RPA vendors (UiPath at 35.8% market share, Automation Anywhere at 9.7%, SS&C Blue Prism at 8.7%) offer deterministic execution but struggle with unstructured data and complex decision logic. Their total cost of ownership is punishing: for every $1 spent on licensing, enterprises spend $3.41–$4.00 on consulting and maintenance. Maintenance alone eats 60–75% of total automation budgets.
AI-native agent platforms (Salesforce Agentforce, ServiceNow AI Agents, LangChain Enterprise) offer cognitive flexibility but at token-driven variable costs of $2–$5 per agent action plus platform licenses. IDC projects total agent platform spend will hit $143 billion by 2027, but 79% of organizations have not yet adopted AI/LLM workflows at enterprise scale, largely due to governance readiness gaps.
Deterministic AI compilers like Poetic represent an emerging third category: systems that use AI to understand processes but execute them without AI at runtime. This hybrid approach—learn probabilistically, run deterministically—addresses the core paradox of enterprise automation: you need intelligence to understand the process, but you need determinism to run it at scale.
The regulatory environment is accelerating this shift. SR 26-2, issued jointly by the Federal Reserve, OCC, and FDIC in April 2026, establishes risk-based oversight for quantitative models. The EU AI Act's high-risk requirements and the US Treasury's framework both effectively mandate deterministic behavior for AI in financial services decision-making. For enterprises in regulated industries, the governance burden of probabilistic agents—auditability, explainability, reproducibility—is becoming a dealbreaker.
Framework #1: Automation Architecture ROI Calculator
Use this calculator to estimate the three-year total cost of ownership across automation approaches for a compliance or fraud investigation workflow processing 500 cases per day.
Scenario A: Traditional RPA (UiPath/Automation Anywhere)
| Cost Component | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Platform licensing | $180,000 | $180,000 | $180,000 |
| Implementation consulting | $450,000 | $75,000 | $75,000 |
| Bot maintenance (60–75% of budget) | $270,000 | $350,000 | $400,000 |
| Exception handling (human labor) | $600,000 | $550,000 | $500,000 |
| Annual Total | $1,500,000 | $1,155,000 | $1,155,000 |
| 3-Year TCO | $3,810,000 | ||
| Accuracy: 85–92% on structured data |
Scenario B: AI Agent Platform (Agentforce/ServiceNow)
| Cost Component | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Platform license | $250,000 | $250,000 | $250,000 |
| Token/action costs ($2–5 × 500/day × 250 days) | $500,000 | $625,000 | $750,000 |
| Governance & compliance overhead | $200,000 | $150,000 | $125,000 |
| Failure recovery (hallucination remediation) | $300,000 | $200,000 | $150,000 |
| Annual Total | $1,250,000 | $1,225,000 | $1,275,000 |
| 3-Year TCO | $3,750,000 | ||
| Accuracy: 88–94% with guardrails |
Scenario C: Deterministic AI Compiler (Poetic-class)
| Cost Component | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Platform licensing | $300,000 | $300,000 | $300,000 |
| Implementation (5-week deployment) | $150,000 | $0 | $0 |
| Runtime costs (near-zero tokens) | $25,000 | $30,000 | $35,000 |
| Maintenance (self-repairing) | $50,000 | $50,000 | $50,000 |
| Annual Total | $525,000 | $380,000 | $385,000 |
| 3-Year TCO | $1,290,000 | ||
| Accuracy: 99%+ from week 5 |
Bottom line: Deterministic AI compilers deliver 66% lower 3-year TCO than RPA and 65% lower than AI agent platforms—while achieving 99% accuracy versus 85–94% for alternatives. The ROI advantage compounds in regulated industries where each accuracy point below 99% carries compliance penalty risk.
Framework #2: Enterprise Automation Architecture Decision Matrix
Not every process belongs on the same platform. Use this matrix to map your automation portfolio across the three paradigms.
When to Choose Each Approach
| Decision Criteria | Traditional RPA | AI Agent Platform | Deterministic AI Compiler |
|---|---|---|---|
| Best for | High-volume, structured, rule-based tasks | Unstructured data, judgment calls, creative tasks | High-stakes, complex, compliance-critical workflows |
| Accuracy requirement | 85–95% acceptable | 88–95% with human review | 99%+ mandatory |
| Data type | Structured (forms, databases, APIs) | Unstructured (emails, documents, conversations) | Mixed (structured rules + unstructured inputs) |
| Regulatory exposure | Low–Medium | Medium–High (auditability challenges) | High (audit-ready by design) |
| Volume | >10,000 transactions/day | 50–5,000 transactions/day | 100–5,000 high-value transactions/day |
| Token cost sensitivity | N/A (no tokens) | High (variable, unpredictable) | Low (near-zero tokens at runtime) |
| Change frequency | Monthly (brittle to UI changes) | Continuous (model updates) | Self-repairing on software changes |
| Time to production | 3–6 months | 2–4 months | 5 weeks (per Poetic benchmarks) |
| Governance burden | Low (deterministic audit trail) | High (explainability, reproducibility) | Low (deterministic + AI-learned) |
| Example use cases | Invoice processing, data entry, report generation | Customer service, content generation, research | Fraud investigation, underwriting, compliance review |
Portfolio Allocation Guide
For a typical Fortune 500 financial institution with 200+ automated processes:
- 60% Traditional RPA: High-volume, low-complexity tasks (invoice processing, account reconciliation, data migration). These are working. Don't fix them.
- 15% AI Agent Platforms: Customer-facing workflows requiring natural language understanding, creative problem-solving, or real-time personalization. Accept the token costs for the cognitive value.
- 25% Deterministic AI Compilers: Compliance investigations, fraud adjudication, underwriting decisions, regulatory reporting—any process where accuracy below 99% carries financial or legal risk. This is Poetic's sweet spot and the category's highest growth vector.
Pre-Migration Checklist (Before Moving a Process)
- Accuracy audit: What is the current error rate? If >2%, the process is a candidate for deterministic AI.
- Regulatory mapping: Does the process fall under SR 26-2, EU AI Act high-risk, or SOX controls? If yes, deterministic execution is likely required.
- Token cost projection: For AI agent alternatives, calculate cost at current volume × $2–5 per action × 250 business days. If annual token costs exceed $500K, evaluate deterministic alternatives.
- Maintenance burden: If current RPA maintenance exceeds 50% of total automation budget for this process, the self-repairing capability alone justifies migration.
- Pilot scope: Start with one process, target 5-week deployment, measure accuracy at week 2 and week 5.
Case Study: SoFi's Fraud Investigation Transformation
SoFi Technologies, the digital financial services company with over 10 million members, deployed Poetic to automate fraud investigation workflows that had resisted automation for years. The results validate the deterministic AI thesis at production scale.
The Problem: SoFi's fraud investigation process required analysts to navigate multiple internal systems, cross-reference transaction data, apply complex decisioning rules, and document findings—a multi-hour process per case that was too nuanced for traditional RPA bots and too high-stakes for probabilistic AI agents where a wrong decision could mean regulatory penalties or customer harm.
The Deployment: Poetic ingested SoFi's procedures, training materials, and expert feedback, then encoded the fraud investigation workflow into its deterministic execution engine. The platform went from pilot to production in five weeks—roughly one-sixth of the typical 3–6 month RPA deployment timeline.
The Results:
- 99%+ accuracy in fraud decisioning from week five
- 100% process adherence in compliance investigations
- Zero hallucination incidents—a metric that probabilistic AI agents cannot guarantee
- Self-repairing capability maintained accuracy when SoFi's underlying systems changed
The Lesson: The constraint was never intelligence—SoFi's fraud analysts were already excellent. The constraint was encoding that intelligence into software that ran with the same consistency. Poetic's approach of "learning like AI, running like code" solved the encoding problem without introducing the reliability risks of autonomous agents. For the 82% of financial services AI projects that fail, the SoFi case suggests the failure mode is not insufficient AI capability but excessive AI autonomy in contexts that demand determinism.
What to Do About It
For CIOs and CTOs: Technical Next Steps
Audit your automation portfolio for processes where accuracy below 99% carries regulatory or financial risk. These are your deterministic AI candidates. Start with compliance investigations, fraud adjudication, and underwriting—the processes where traditional RPA is too rigid and AI agents are too unpredictable. Evaluate whether your current governance framework can support SR 26-2 requirements for model risk management, and consider how deterministic AI compilers reduce that burden by eliminating the probabilistic components that require explainability documentation.
For CFOs: Financial Next Steps
Run the ROI calculator above with your actual numbers. Replace the illustrative figures with your real licensing costs, headcount allocated to exception handling, and compliance penalty exposure. Pay particular attention to the token cost trajectory for AI agent platforms—EY's research shows these costs compound unpredictably as workflows grow in complexity. Request a pilot scope that delivers measurable accuracy and cost metrics within 6 weeks, not 6 months. The 100% pilot-to-production conversion rate Poetic reports suggests that if the pilot works, production deployment is not the bottleneck—organizational readiness is.
For Business Leaders: Strategic Next Steps
Reframe the automation conversation from "AI agents vs. RPA" to a three-paradigm portfolio. The enterprises that will capture the most value from automation in 2026–2028 will be those that match the right architecture to the right process class. 91% of financial services firms already agree that automation improves compliance—but only 8.6% have achieved enterprise-wide orchestration. The gap is not technology. It is architecture—knowing which paradigm fits which workflow, and having the governance framework to run all three.
