The SEI/Accenture practice rubric is fully public, the assessment instrument is not, and "nobody scores the agent" no longer survives PwC's and KPMG's public pages
The question
"Is the Accenture / Carnegie Mellon SEI AI Adoption Maturity Model practice-level rubric public or partner-gated, and does any Big-4 gated assessment material score authority at the AGENT level?"
Context: two follow-ups merged from [[2026-09-26-big4-agentic-readiness-frameworks-primary-verify]]. The first decides whether an external benchmark exists for Org Map's scoring rubric. The second tests whether per-agent authority is still a clean differentiator for phData's Organizational Intelligence (OI) work, since the September absence finding covered public landing pages only.
What we already know (from the vault)
- The September brief established the model's eight dimensions and two goals from Accenture's newsroom, and left open whether the practice statements behind them were public. It also reworded the differentiation claim to "none of them scores authority at the agent level." See [[2026-09-26-big4-agentic-readiness-frameworks-primary-verify]].
- The November 4 brief found the model is a free SEI-published PDF with no questionnaire or tool offered for download, and an "early adopter program" reachable by contact form. It hit its fetch cap before opening the PDF. See [[2026-11-04-accenture-ai-navigator-refinery-coupling]].
- The July landscape brief is where the autonomy axis was first recommended as a client-visible scoring dimension. See [[2026-07-15-agentic-assessment-framework-competitive-landscape]].
- Thoughtworks' Agentic Scope of Authority Framework is the closest per-agent authority instrument previously logged. Its nine blocks are unpublished, its tool is unreachable, and its published content is a binary checklist, not an ordinal rubric. See [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]].
- Accenture's AI Refinery is a runtime with no diagnostic attached, so it is not where an Accenture assessment instrument would live. See [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]].
What the web says
Retrieved 2026-11-11. The SEI and KPMG documents were downloaded and text-extracted, so the quotes and keyword counts below come from the full text, not from a summary.
- The practice-level rubric is public, in full, with no gate. "The AI Adoption Maturity Model v1.0" downloads as a 124-page PDF from SEI with no form or login. Sections 4.4 to 4.11 give every dimension's capability areas, each with numbered goals, numbered practice statements, example artifacts, and the maturity level that capability area is pinned to. There are 25 capability areas. My count from the extracted text is roughly 64 goals and 244 practice statements; that is a token count, not a figure SEI states. The PDF metadata shows it was regenerated 2026-08-13 and modified 2026-08-26, later than the 2026-06-30 library date, so the file has been revised at least once since launch. (SEI, AI Adoption Maturity Model v1.0 PDF, primary)
- The scoring logic is also public. Three maturity indicators (Accountability, Planning, Resourcing) each carry a statement for all five levels. The default method is lowest-wins: "a dimension is rated to its lowest ranking capability areas, and a capability area is rated to its lowest ranking maturity indicator, even if all the goals and practice areas are satisfied." SEI calls this "a harsh scoring mechanism." Each capability area is tied to one level by a fixed sentence, for example all goals and practices of Human-in-the-Lead Automation are required "for this dimension to be at Aligned AI Level Maturity (Level 3)." (same PDF, sections 4.13 and 5)
- What is gated is the instrument and the data, and the gate is an engagement. Goal ratings come from "normalization of assessor evaluations, cross-validation with organizational unit survey data, and verification through supporting artifacts." None of that is published: no survey, no assessor guide, no appraisal method document, no benchmark dataset. The word "appraisal" appears zero times and "assessment tool" zero times in the PDF. The only public route in is SEI's library page, which invites readers to "share feedback or get involved by joining our early adopter program" through a contact form at
/forms/ai-maturity-model-feedback/. Accenture's launch release separately describes "an assessment tool that enables structured implementation and benchmarking of client outcomes across industries," which has no public surface. I did not submit the form. (SEI library page, primary) - The model does touch agents, but the unit it rates is the organizational unit. The word "authority" appears zero times in 124 pages and "delegat-" once, in narrative. Agent-specific practices exist: "Identify and define the allowable entitlements for AI models, agents, and systems"; "Apply identity and access management principles and practices to non-human entities"; "Segment access and apply principles of the separation of duties to AI agents based on their role and identity"; "Implement continuous monitoring for agentic and autonomous solutions"; and "Define critical and irreversible decision points" that need human oversight. Each asks whether the organization has the practice. None rates how much authority a given agent holds, and there is no autonomy scale. The AI Model and Agent Security capability area is pinned at Level 2, so agent entitlements are treated as a baseline hygiene item. (same PDF, sections 4.6.3, 4.9.1, 4.10.2, 4.10.3)
- KPMG publicly prescribes a per-agent risk level and per-agent authority limits. Its seven-page paper "AI governance for the agentic AI era" (PDF dated 2025-07-24) lists ten control considerations. The first is "Assess agent risk: Complete an impact assessment and determine the risk level of the agent." The third is to "define clear scope constraints on what an agent can and cannot do, including limits on data access, tool usage, and decision-making authority." It names no levels, no scale and no tool, and its TACO typology (Taskers, Automators, Collaborators, Orchestrators) classifies agents by function without assigning a risk tier to each type. (KPMG PDF, primary)
- PwC publicly recommends tiering agents by autonomy. Its page "Responsible adoption of AI agents in the enterprise" (released 2025-04-10, last modified 2025-12-19) says: "Incorporate agents' degree of autonomy and the potential impact of their decisions into risk tiering and prioritization structures. This will enable greater levels of governance for AI agents that have more autonomy or are used in more critical spaces while fast tracking lower risk agents." It also asks which agent attributes "would trigger inclusion into a centralized AI inventory," including "the degree of autonomy under which they perform." Recommendations only; no rubric, no levels, no assessment product described. (PwC, primary, via Playwright after a 403)
- Search-summary only, not opened, do not quote. A PwC Japan report is summarized as using five autonomy levels, L1 to L5, mapped against functional risk domains. A trade-press write-up summarizes Deloitte guidance as "tiered autonomy" that moves from view-only, to act-with-approval, to act automatically. Nothing surfaced for EY beyond an opinion blog. No public page from any of the four describes a gated tool that scores delegated authority per agent. (PwC Japan, AI News on Deloitte, both unverified)
Convergences and contradictions
- Contradiction with the vault's current wording. The September sentence, that none of these frameworks scores authority at the agent level, is still true of the three frameworks that brief checked. It is not true of the firms. KPMG and PwC both publish agent-level guidance that makes the individual agent the unit of risk assessment, and PwC names degree of autonomy as a tiering input. An informed listener can rebut the broad version with two public URLs.
- Convergence on the gap that remains. Across SEI, KPMG and PwC, nothing public gives an ordinal scale for agent authority with level definitions and a way to rate one agent against it. The agent is named as the unit; the measuring stick is not published. That matches the Thoughtworks finding, where the published content was also a checklist with the instrument withheld.
- The gated half is unobserved, not negative. I found no description of a gated Big-4 instrument that scores per-agent authority, and I could not look inside any gate. Given that two firms publicly tell clients to tier agents by autonomy, it would be unsurprising if their engagement tooling does it. That is an inference and should be read as one.
Synthesis for RDCO
First half, answered cleanly: public rubric, gated instrument. The practice statements are not behind a partner portal. Anyone can read all 25 capability areas, their goals and practices, the three maturity indicators and the lowest-wins rating rule. So the sharpest external benchmark for Org Map's scoring rubric exists and is free, and the backlog note's other branch ("if gated, OI's rubric is unobserved by competitors too") does not apply to SEI's side. What SEI and Accenture hold back is the part that turns a rubric into a repeatable assessment: the survey, the assessor normalization method and the cross-client benchmark. That is the same split Thoughtworks uses, and it is a reasonable template for what OI should publish and what it should keep.
Three things in the SEI structure are worth comparing Org Map against directly. One, the lowest-wins rule, which SEI itself calls harsh; it prevents a strong dimension from averaging away a weak one, and it is a deliberate design choice OI should either adopt or be able to explain not adopting. Two, the three maturity indicators are about institutionalization (who is accountable, is there a plan, is it resourced), scored separately from whether practices are performed; this is how the model distinguishes a team that can do something from an organization that reliably does. Three, each capability area is pinned to a single level, so the model encodes an opinion about sequence. Agent entitlements sit at Level 2 and human-in-the-lead oversight at Level 3, which says SEI regards both as early, foundational work and not as advanced maturity. These are Ray's picks for comparison, not founder-stated priorities.
Second half: the differentiator needs narrowing a second time. The wording has now moved twice. July said the Big-4 were pre-agentic. September said they score the enterprise and none scores the agent. Tonight's evidence supports only this: public Big-4 guidance already treats the individual agent as a unit of risk assessment, and PwC ties governance intensity to the agent's degree of autonomy, but no firm publishes an ordinal authority scale or a per-agent rating instrument. The defensible claim for OI is therefore about the instrument and its output, meaning a defined scale, a rating per agent, and a result that routes work, and not about being first to notice that agents need individual authority limits. That idea is in a 2025 KPMG paper and a 2025 PwC page.
What stays unknown, and how much weight to put on it. Whether a gated Deloitte, PwC, EY or KPMG tool already produces a per-agent authority score could not be determined without an account or an engagement, which the run rules exclude. The prudent planning assumption is that at least one does or soon will, because the public guidance points straight at it. If OI's positioning depends on that not being true, it is resting on something unverified. Positioning on the coupling of the rating to delivery, which the vault has found no competitor doing outside Thoughtworks, is the sturdier ground.
Why this is in the vault
It closes both open follow-ups from [[2026-09-26-big4-agentic-readiness-frameworks-primary-verify]]: it gives Org Map's rubric a named, free external comparator with a public rating rule, and it replaces the "none scores authority at the agent level" sentence used in OI competitive material with a narrower claim that the KPMG and PwC public pages cannot knock down.
Open follow-ups
- How do Org Map's dimensions and practices map onto SEI's 25 capability areas, and where does each side have coverage the other lacks? The rubric is now fully readable, so this is a document comparison, and it would show whether OI's scoring has gaps a buyer who has seen the SEI model would notice.
- Does the PwC Japan report define its L1 to L5 autonomy levels with criteria, and does Deloitte's primary guidance define its autonomy tiers? Both are search-summary only tonight. If either publishes level definitions, a public ordinal autonomy scale from a Big-4 firm already exists and the remaining gap narrows again.
- Which non-Big-4 public autonomy scales are becoming the reference standard? The search surfaced a Cloud Security Alliance six-level autonomy framework, Singapore IMDA's Model AI Governance Framework for Agentic AI, and a NIST agent standards initiative. If buyers anchor on one of these, OI's scale should map to it instead of standing alone.
- Has SEI published or announced an assessment method, assessor qualification, or licensing scheme for the model since v1.0? CMMI's commercial weight came from its appraisal method and licensed appraisers. If SEI repeats that, the instrument stops being Accenture-adjacent and becomes a market standard.
Related
- [[2026-09-26-big4-agentic-readiness-frameworks-primary-verify]] - parent brief; its two open follow-ups are answered here and its agent-level sentence is narrowed
- [[2026-11-04-accenture-ai-navigator-refinery-coupling]] - identified the public PDF and the early adopter contact form; this brief reads the PDF it could not reach
- [[2026-07-15-agentic-assessment-framework-competitive-landscape]] - origin of the autonomy-axis recommendation that this brief constrains
- [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]] - same publish-the-thinking, gate-the-instrument pattern, and the other per-agent authority framework on file
- [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]] - establishes that Accenture's runtime carries no diagnostic
Sources
Vault:
- [[2026-09-26-big4-agentic-readiness-frameworks-primary-verify]]
- [[2026-11-04-accenture-ai-navigator-refinery-coupling]]
- [[2026-07-15-agentic-assessment-framework-competitive-landscape]]
- [[2026-09-27-thoughtworks-agentic-scope-of-authority-nine-blocks]]
- [[2026-09-11-accenture-ai-refinery-runtime-vs-methodology]]
Web (primary, retrieved 2026-11-11):
- SEI, "The AI Adoption Maturity Model v1.0," 124-page PDF, no gate, copyright 2026 Carnegie Mellon University, document marking DM26-0590; full text extracted and keyword-searched: https://www.sei.cmu.edu/documents/6534/The_AI_Adoption_Maturity_Model.pdf
- SEI Digital Library page for the model; one PDF download, early adopter program via contact form, no tool or survey offered: https://www.sei.cmu.edu/library/ai-adoption-maturity-model/
- KPMG, "AI governance for the agentic AI era," 7-page PDF dated 2025-07-24; ten control considerations including per-agent risk level and decision-making authority limits: https://kpmg.com/kpmg-us/content/dam/kpmg/pdf/2025/ai-governance-for-agentic-ai-era.pdf
- PwC, "Responsible adoption of AI agents in the enterprise," released 2025-04-10, modified 2025-12-19; autonomy as a risk-tiering input (WebFetch returned 403, read via Playwright): https://www.pwc.com/us/en/tech-effect/ai-analytics/responsible-ai-agents.html
Web (search-summary only, not opened, do NOT quote):
- PwC Japan, cyber risks of AI agents, reported L1 to L5 autonomy levels: https://www.pwc.com/jp/en/knowledge/thoughtleadership/cyber-risk-of-ai-agents.html
- AI News write-up of Deloitte agentic guidance, reported tiered autonomy: https://www.artificialintelligence-news.com/news/deloitte-agentic-ai-guidelines-published/
- KPMG Australia TACO framework page: https://kpmg.com/au/en/services/ai-services/taco-framework.html
Not determinable from outside the gate:
- Contents of the SEI early adopter program, Accenture's assessment tool and its benchmark data
- Whether any Deloitte, PwC, EY or KPMG engagement tool produces a per-agent authority score