On May 19 in Dallas, Automation Anywhere did something the RPA category has needed for three years and the agent platform category has never seen before: it got Cisco, NVIDIA, Okta, and OpenAI to co-sign a single enterprise agent stack. The result, EnterpriseClaw, is in preview now with GA "later in 2026" — and it lands in a market where Gartner expects 40% of enterprise apps to ship task-specific AI agents by year-end, up from less than 5% a year ago. The same analyst firm warns that over 40% of agentic AI projects will be canceled by the end of 2027 — mostly because of escalating cost, unclear ROI, and weak controls.
That tension is the entire reason EnterpriseClaw exists. The RPA market is on track to expand from $35.27 billion in 2026 to $247.34 billion by 2035, but only if the bots can survive the transition into something larger — autonomous, identity-bearing, reasoning agents that act across cloud, desktop, and on-prem systems without leaking regulated data. CIOs are being asked to bet seven- and eight-figure automation budgets on which platform can pull that off. EnterpriseClaw is the first time a deterministic RPA leader (Automation Anywhere), a hyperscale security vendor (Cisco), a foundation-model accelerator (NVIDIA), an identity standard (Okta), and the highest-profile frontier-model lab (OpenAI) have all stood behind the same answer.
What Changed: One Stack, Five Vendors, Zero Solo Acts
EnterpriseClaw is not a re-skinned RPA platform. It is an orchestration and governance layer designed specifically to run "claw-style" AI agents — autonomous, multi-step, decision-making — inside the messy reality of enterprise systems. The May 19 press release names four hard problems and five vendors solving them in concert:
- Process awareness comes from Automation Anywhere's own Process Reasoning Engine and Contextual Intelligence Graph — the same plumbing that, per Automation Anywhere's 70-deployment dataset published in April, produces 80%+ auto-resolution on IT tickets and 50% ITSM cost reductions inside Fortune 1000 customers.
- Runtime and on-prem inference comes from NVIDIA OpenShell and NIM microservices running Nemotron open-weight models. NVIDIA AI Enterprise — the production license required to ship those NIMs into a data center — runs $4,500/GPU/year, a number CFOs need on the BOM before anyone signs.
- Identity and authorization comes from Okta, whose CPO Ely Kahn put the design principle on record: "AI Agents must have first-class identities" with "rigorous authentication and least-privilege authorization." Okta's own agentic enterprise data shows 88% of organizations have already had a suspected or confirmed AI agent security incident — and only 22% treat agents as independent, identity-bearing entities.
- Runtime security comes from Cisco AI Defense and the open-source DefenseClaw framework Cisco announced at RSAC 2026. DefenseClaw is being engineered to use NVIDIA OpenShell as its sandbox, meaning the same runtime that boots a Nemotron NIM also enforces tool-call inspection, prompt-injection blocking, and memory-poisoning detection.
- Reasoning comes from OpenAI's GPT-5.5 and other frontier models, configurable per workflow. OpenAI's Jason Lee, head of enterprise technology, framed the win bluntly: "EnterpriseClaw enables organizations to bring frontier AI capabilities into agentic systems for regulated and mission-critical processes."
Automation Anywhere CEO Mihir Shukla summarized the bet in one line: "For AI to have a transformational impact on business, it needs to do work where the work actually happens." Translated for CIOs: the agent has to run inside the SAP instance, the Citrix VDI, the airgapped claims-processing network, and the on-prem ERP — not in a sandboxed cloud demo. Cisco's announcement-day stock reaction was a modest -2.94% versus a +0.72% historical average on similar AI partnership news — a signal investors are still pricing the bundling complexity, not the strategic logic.
Why This Matters: Two Audiences, One Stack
For CIOs, CTOs, and CISOs
The EnterpriseClaw architecture closes three governance gaps that have been killing agent pilots. First, identity — most agent frameworks today still authenticate as a service account or a copy of a human user's token, an approach Okta's research correlates with the 88% incident rate cited above. Treating agents as first-class identities means every action is signed, scoped, certified, and reviewable, with revocation that works at agent-granularity instead of user-granularity. Second, runtime protection — Cisco AI Defense is already inspecting MCP traffic in real time at Global 2000 customers, blocking unauthorized tool usage, harmful action chains, and memory-poisoning attempts. With DefenseClaw integration, those policies travel with the agent into on-prem deployments. Third, hybrid placement — Nemotron NIMs let regulated workloads (healthcare, claims, finance, defense) run frontier-class reasoning locally without sending payloads to a public API.
The architectural significance is that EnterpriseClaw is the first commercial stack to ship all three together with a single accountable vendor for the support contract. ServiceNow's Project Arc control tower attempts the orchestration piece; Microsoft's Agent 365 SKU attempts identity and observability; Google's Gemini Enterprise platform attempts the model and gateway layer. EnterpriseClaw is the first attempt to ship the full vertical bundle with named accountability at every layer.
For CFOs, CEOs, and Boards
The financial story is more direct. The average enterprise is sitting on a decade of RPA depreciation — hundreds or thousands of brittle bots that break every time an SAP screen moves, with licensing tied to bot count rather than business outcome. Forrester's 2026 prediction is blunter: adaptive process orchestration — the new category that combines deterministic RPA with non-deterministic AI agents and human oversight — is the next maturity step, and vendors that don't pivot will be displaced.
EnterpriseClaw is Automation Anywhere's pivot. For CFOs, that means three concrete budget questions surface this quarter: (1) what percentage of the existing RPA bot fleet can be retired by an agent that reasons across systems instead of clicking each screen, (2) what does the migration from per-bot licensing to per-outcome consumption cost over 24 months, and (3) what is the residual cost of running Nemotron NIMs on owned GPUs versus calling GPT-5.5 over the public API for the same workflow. The next section turns those questions into numbers.
Market Context: The Compressed Agent Platform Landscape
The agent platform market is consolidating faster than any enterprise software category in the last decade. Gartner's agentic AI hype cycle places the category at the peak of inflated expectations, with the long-term ceiling at approximately 30% of enterprise application software revenue by 2035, surpassing $450 billion, up from 2% in 2025. The short-term reality is that 42% of enterprises already have agents in production and 72% are running production and pilots combined, per Onereach.ai's mid-2026 survey.
Inside that consolidation, four buyer archetypes have emerged. Microsoft-centric organizations gravitate to Copilot Studio for its native fit with Teams, SharePoint, and Outlook, and Microsoft's per-user licensing model. CRM-centric organizations are standardizing on Salesforce Agentforce, especially after the Coworker launch on May 21 extended the agent into Slack, Teams, ChatGPT, and mobile. RPA-heavy organizations have historically defaulted to UiPath, the #1 Gartner Magic Quadrant RPA vendor for six consecutive years. Regulated, hybrid-infrastructure organizations — banks, insurers, defense, healthcare — have been waiting for a stack like EnterpriseClaw that can run identical agents in cloud and on-prem without architectural compromise.
Forrester's Adaptive Process Orchestration Wave, scheduled for late 2026, will be the first formal head-to-head benchmark. Decisions, ProcessMaker, and 33 other vendors are already named in the landscape report. EnterpriseClaw is the only one shipping with a five-vendor co-sign on day one.
Framework #1: When to Choose EnterpriseClaw vs Copilot Studio vs UiPath vs Agentforce
The decision matrix below is built from vendor pricing disclosures, Gartner adoption data, and the deployment patterns documented in Forrester's 2026 automation research. CIOs should score their environment against each row before any pilot contract is signed.
| Decision Dimension | EnterpriseClaw | Microsoft Copilot Studio | UiPath Agent Builder | Salesforce Agentforce |
|---|---|---|---|---|
| Best-fit buyer | Regulated, hybrid, multi-system | M365-centric, cloud-first | RPA-heavy, legacy ERP | CRM-centric, sales/service |
| On-prem inference | Yes (Nemotron NIM) | Limited (Azure tenancy) | Partial (UiPath Orchestrator) | Cloud-only |
| Foundation model choice | GPT-5.5 + Nemotron multi-model | Azure OpenAI (locked) | OpenAI + Claude options | Atlas + multi-model |
| Identity model | Okta first-class agent identity | Entra ID + agent extensions | UiPath Orchestrator service accts | Salesforce platform identity |
| Runtime security | Cisco AI Defense + DefenseClaw | Microsoft Defender + Purview | UiPath Guardrails | Salesforce Shield |
| Pricing visibility | Preview, not public | ~$30/user/month + $200/25k credits | Tier-based, contact sales | Flex Credits, 20-120 per action |
| Typical payback (500+ FTE) | 6–9 months (projected) | 3–5 months | 6–12 months | 5–8 months |
| Worst fit when… | Pure cloud, single-app footprint | Heterogeneous, non-Microsoft | Cloud-native, no legacy debt | Outside CRM workflows |
Choose EnterpriseClaw if at least three of the following are true: (1) you have a $5M+ annual RPA spend you want to migrate, (2) regulated data must stay on-prem, (3) you have already standardized on Okta for workforce identity, (4) Cisco is already your runtime security vendor, (5) you want optionality between Nemotron open models and OpenAI frontier models without re-platforming. Choose Copilot Studio if more than 80% of your workforce already lives in Microsoft 365 and your agent workloads are document-, email-, and meeting-centric. Choose UiPath Agent Builder if your dominant constraint is connecting AI reasoning to legacy mainframe and screen-scraped ERP systems with a deep RPA bot inventory you cannot retire yet. Choose Agentforce if the primary work surface is CRM, customer service, or sales — and especially if you have already committed Flex Credits.
The matrix is not exclusive. Most Global 2000 firms will end up running two of these platforms; Forrester's 2026 prediction explicitly warns against single-vendor agent strategies because of model-lock risk.
Framework #2: The 5-Phase RPA-to-Agent Migration Timeline
Migration sequencing is where 40% of agent projects die, per Gartner. The phased plan below is calibrated for an enterprise with 500+ existing RPA bots and a 12-month budget cycle.
Phase 1 (Weeks 1–4): Bot Inventory and Identity Foundation. Audit every active RPA bot by business owner, transaction volume, and break frequency. The 30% of bots that break weekly are the highest-ROI candidates for retirement. In parallel, deploy Okta for AI Agents (GA since April 30, 2026) to discover shadow agents, register human owners, and define agent role profiles. Success criterion: a tagged inventory and a non-human-identity standard signed off by IAM, security, and compliance leadership.
Phase 2 (Weeks 5–10): Single-System Pilot with Cisco AI Defense. Pick one high-volume, high-leverage workflow — claims investigation, vendor invoice reconciliation, ITSM ticket triage — and rebuild it as an EnterpriseClaw agent. Cisco AI Defense Explorer Edition should red-team the workflow before any production traffic touches it, with success measured against the Global 2000 baseline Cisco published in their Q1 report. Success criterion: documented prompt-injection resistance, blocked unauthorized tool calls, and a benchmark ROI on the pilot workflow.
Phase 3 (Weeks 11–20): Cross-System Expansion with Nemotron NIMs. Add at least two more business systems to the pilot agent and migrate inference to on-prem Nemotron NIMs for regulated workloads. The $4,500/GPU/year NVIDIA AI Enterprise license must be modeled against the GPT-5.5 API spend it replaces — break-even typically lands between 8 and 12 million tokens per day per GPU, depending on the prompt mix. Success criterion: identical agent behavior across cloud and on-prem deployments, with audit trail completeness above 99%.
Phase 4 (Months 6–9): Bot Retirement Wave. For every successful agent workflow, decommission the corresponding RPA bots. Cohort the retirement waves — first the bots with >20% break rates, then the bots with the highest per-transaction maintenance hours. Track three KPIs: cost saved per retired bot, transactions reassigned to agents, and incident rate change. Success criterion: 25%+ of legacy bot fleet retired or absorbed.
Phase 5 (Months 10–12): Governance Productization. Move from per-pilot governance to a control-tower model. Adopt either a ServiceNow- or Microsoft-style universal control plane on top of EnterpriseClaw, plus the agent observability layer Cisco and NVIDIA are shipping under DefenseClaw + OpenShell. Success criterion: a single dashboard reporting agent identity, cost, error rate, and policy violations — and a documented kill-switch process tested at scale.
The five phases are sequential because skipping the identity foundation in Phase 1 is the single highest predictor of project failure in Gartner's cancellation data. The migration is not technically hard; it is organizationally hard. Treat it like a multi-year ERP rollout, not a tactical automation refresh.
Case Study: The Claims-Investigation Reference Pattern
The launch press release cites one named use case in detail — claims investigation — and it is worth dissecting because it is exactly the pattern most regulated enterprises will replicate. The scenario: a customer files a claim that requires data from a desktop application (the agent intake form), an on-premises system (the policy management database), internal documents (medical records or incident reports), and cloud services (third-party fraud-scoring APIs). In the legacy RPA world this took three to five separate bots glued together by a human caseworker, and any UI change broke the chain.
In the EnterpriseClaw pattern, a single agent — authenticated via Okta with claim-handler-scoped permissions, reasoning via either a local Nemotron NIM (for HIPAA payloads) or GPT-5.5 (for fraud signals), inspected at every tool call by Cisco AI Defense — gathers all four data sources, applies the policy engine, and routes the case to a human reviewer with a recommendation and full audit trail. Mihir Shukla's framing applies here directly: the agent does the work where the work happens, not in a separate workflow tool. The sensitive payload never leaves the regulated network. The audit log is signed end-to-end.
The reference outcome from the analogous Automation Anywhere 70-deployment dataset suggests a 50%+ reduction in handle time and a 30–50% reduction in licensing cost. EnterpriseClaw's value lift over the deterministic baseline is the additional context-spanning reasoning — the agent does not just resolve tickets that look like prior tickets, it can investigate cases it has never seen by reasoning across documents and systems. Lessons that translate to other industries: pick a workflow where the bottleneck is data-stitching, not decision-making; design the agent identity model before the agent itself; instrument cost per resolved case before scaling beyond pilot.
What to Do About It
For CIOs and CTOs (next 30 days):
- Request an EnterpriseClaw preview demo and require a hands-on red-team session with Cisco AI Defense Explorer Edition before any LOI.
- Inventory your existing Okta tenancy for agent-identity readiness — discovery, registration, certification workflow, and least-privilege scoping.
- Validate whether your current GPU inventory supports a Nemotron NIM pilot, or whether you need to budget for new hardware in the 2027 cycle. Cross-reference SUSE and other sovereign-AI on-prem patterns for deployment topology.
For CFOs (next 60 days):
- Build the Framework #1 decision matrix with your actual workforce composition, RPA spend, and identity-system commitments. Score every cell.
- Model the bot-retirement cash flow over 24 months. Use 30% bot breakage as the starting attrition curve and validate against your help-desk ticket data.
- Negotiate EnterpriseClaw preview pricing against UiPath Standard, Copilot Studio credits, and Agentforce Flex Credits in parallel — the platforms know they are being benchmarked. The current preview window before late-2026 GA is the best leverage point of the cycle.
For Business Leaders (next 90 days):
- Pick one cross-system, regulated-data workflow that has resisted automation for at least two years. That is your EnterpriseClaw pilot candidate, not the easy wins.
- Designate an agent owner per business function — not a project manager, an accountable owner with budget and headcount authority. Gartner's cancellation data correlates strongly with the absence of this role.
- Treat the migration as a 12-month roadmap with quarterly board reviews. The 40% project-cancellation rate is overwhelmingly a governance failure, not a technology failure.
The strategic frame is simple. The RPA category as it existed from 2014 to 2024 — deterministic, screen-bound, per-bot-licensed — is finishing its lifecycle. The next $200 billion of automation spend will flow to whichever platform can run autonomous, identity-bearing, security-bounded agents across hybrid infrastructure with frontier-grade reasoning. On May 19, Automation Anywhere, Cisco, NVIDIA, Okta, and OpenAI placed their joint bet. The CIOs who run the decision matrix this quarter, before the GA pricing locks in late 2026, will own the migration economics. The ones who wait will pay GA list prices and inherit a brittle bot fleet they should have started retiring six months ago.
