Enterprise Agentic AI: 27% Ready, Frameworks to Assess
Research on 177 deployments debunks vendor hype—only 27% of processes suit full agentic automation. PASF scores suitability; PADE blueprints step-level designs with 9 patterns.
Vendor Hype Masks Harsh Realities in Agentic Deployments
Big Tech pitches autonomous AI agents as job-killers and efficiency saviors, with Salesforce claiming 45,000 Agentforce deployments, Microsoft 400,000 custom agents, and ServiceNow 8,500 embeddings. But verified data from 47 deployments shows vendor efficiency gains of 42% drop to 21% independently, time savings from 68% to 31%. Only 30% of production AI projects hit ROI, under 20% show EBIT impact. Failures stem from data quality (34%), governance (28%), scope creep (22%), not model errors (16%). Short-term metrics ignore maintenance, oversight costs. "Fear is the product. The apocalypse is the pitch deck." Marco van Hurne calls this a structural gap driven by vendor incentives.
PASF: Eight Questions Triage Processes into Automation Zones
Before building, score processes with PASF (Process Automation Suitability Framework) across eight weighted dimensions predicting success from 177 deployments (2022-2026, 136 sources). Top weights: structurability (20%, formal steps/reversibility; <4 score = <15% success) and risk profile (20%, financial/legal/physical/reputational harm).
Questions:
- Structurable? (Steps, inputs/outputs, logic specifiable; reversible?)
- Risk profile?
- Data clean/accessible? (Rule-bound vs. tacit judgment)
- Frequency?
- Exception frequency?
- Reversibility.
- Rule-boundedness.
- Stakeholder impact (error sensitivity).
Composite PASS (0-10) sorts into zones: I (7+; 27%, automate now), II (5.5-6.9; 17%, pilot 3-6 months), III (4-5.4; 21%, human-in-loop), IV (<4; 12%, no-go 12-24 months). 23% unclassifiable. IT/SSC leads (7.8 PASS, 63% Zone I: structured, API-clean, rollback-friendly). Customer service (6.9, 41% Zone I, 21% III: exceptions like angry customers). Finance (6.4, even spread). Healthcare prior auth (Zone III: high risk/low reversibility, 40% steps automated with review). Legal drafting like Harvey AI (5.1, Zone III: costly errors).
Successes like Klarna (7.6 PASS: 2.3M convos, 700 FTE equiv.), Lemonade Jim (7.8: 30% no-human claims), PagerDuty (7.9: 60% MTTR cut) were Zone I pre-agent due to structure, data, escalation—not agent magic. "They are all Zone I because the process was Zone I before the agent arrived, but not because the agent transformed a chaotic process into something governable."
PADE: Step-Level Blueprints Pick from 9 Agentic Patterns
For suitable processes, PADE (Process Automation Design Engine) decomposes into steps, assigning paradigms: AI assistant (human decides), agentic (autonomous multi-step/tools), browser/computer-use (RPA-like, no API). Then, for agentic steps, selects from 9 patterns based on complexity, error tolerance, horizon:
| Pattern | Use Case |
|---|---|
| ReAct | Default moderate complexity (reason-act loop) |
| Plan-and-Execute | Long-horizon |
| Orchestrator-Subagent | Cross-system coord |
| Critic-Actor | Low error tolerance, iterative check |
| Reflexion | Learn from failures |
| Memory-Augmented | Context/long-state |
| Multi-Agent Debate | High-stakes consensus |
| Single-Tool Agent | Bounded single-system |
| Hierarchical Planning | Multi-level high-tool count |
Outputs blueprint: automation type, pattern, governance, human pull-in triggers. Tested as free app (register at EigenVector site). Lowest accuracy in Zone III ambiguities (e.g., clinical auth: assistant vs. agentic debate). Pairs with OCG (Ontological Compliance Gateway) neuro-symbolic arch for Zone III safety.
"Processes with structurability scores below 4 have a deployment success rate of less than 15% according to the data, regardless of how well they score on everything else."
Sector Patterns Reveal Where Agents Thrive
IT/SSC dominates due to APIs, definitions, rollbacks. Customer service handles volume but exceptions drag to Zone III. Finance varies (invoice match Zone I, underwriting IV). Zone I wins share traits: pre-existing structure enables catchable errors. Vendors overclaim by cherry-picking.
Paper: "From Suitability to Blueprint..." (TechRxiv/EigenVector). OCG paper separate. Slide deck TL;DR available.
Key Takeaways
- Score processes with PASF's 8 dimensions before pilots; prioritize structurability/risk (40% weight).
- Expect Zone I (27%) only; pilot II, human-loop III, skip IV.
- Verified gains half vendor claims—factor in oversight/maintenance.
- Use PADE for blueprints: match 9 patterns to step traits, not one-size agents.
- Target IT/SSC first (7.8 PASS); avoid judgment-heavy like underwriting.
- Build governance early: data (34% failures), not just models.
- Test PADE app on your processes; iterate on Zone III with OCG-like determinism.
- Ignore job-apocalypse hype; agents augment Zone I structures.
- Measure long-term: 3-6mo vendor metrics miss entropy.
- Reassess IV processes in 12-24mo as tech matures.