The Answers That Came Back: The Scorecard

Companion page to the Utility Voice article The Answers That Came Back, October 2026. Every numbered question from The Questions We Need Answered: The Full List, scored against what was documented at SAP for Utilities 2026 and in public sources through October 22, 2026.

Eleven answered. Twenty-seven partial. Twenty-eight with no answer.

How it is scored. Answered means a dated, documented answer to the question as asked: a conference slide, a filing, a published page. Partial means a documented answer to part of the question, or tooling that makes an answer possible without anyone having given it. No answer means nothing that addresses it. A session that discussed a topic does not count, and neither do third-party estimates of SAP’s prices.

Questions are paraphrased where the original was long; the full wording is on the question list. Article numbers there match the series index on the home page.

A. The Regulator Test

3 answered · 2 partial · 0 no answer. Three of five answered, by one utility, for AI in IT delivery and support. Nobody answered where an agent sits in the organization or how to shut it down.

  • 1. Who authorized it to exist? — Answered. OG&E: every AI tool has a clear business purpose and an accountable owner (conference slides). New York’s PSC (Case 26-M-0552) now requires an inventory of every AI use case, due November 16. Scope: IT delivery and support AI.
  • 2. What rules govern what it should be doing? — Answered. OG&E: governed by a named corporate cyber security policy. Nexus Water: guiding principles, “no-fly zones” and decision rights (conference slides).
  • 3. Who is watching what it actually does? — Answered. OG&E: people review important results; tools are logged and monitored. PG&E: a “Supervise” stage tracks quality, failure and cost in production (conference slides).
  • 4. Where does it sit in your organization? — Partial. No utility said where an agent sits. PPL’s Utility Canvas places capabilities in value streams and domains, not agents.
  • 5. How would you shut it down? — Partial. SAP describes Agent Gateway as the runtime policy enforcement point, but no published way to suspend an agent, quarantine its outputs or revoke its access. Nobody described a kill switch.

B. Inside your own utility

5 answered · 7 partial · 12 no answer. Five answered, more than SAP fully answered in its own room. The answers govern AI already in use under human approval; they stop where an agent would act on its own.

  • 6. If not SAP’s governance framework, then what? — Answered. Four utility frameworks on conference slides: Nexus Water (no-fly zones, decision rights, approvals around financial, customer and policy impacts); PG&E (register, certify, supervise, adjudicate); OG&E (accountable owner, named policy, least access); SRP (“SRP makes the final decision — every time”).
  • 7. How do we operationalize agentic AI with true write access to production, in a way that satisfies security, compliance, internal audit and the regulator? — Partial. Every utility answer was “not yet, and deliberately so”: human-owned approvals at PG&E, Nexus and OG&E. Nobody described granting an agent write access.
  • 8. When an agent acts on production data, who is accountable? — Answered. A named human owner: OG&E, PG&E and Nexus Water all said so. In every production example, a person approves the result.
  • 9. Would we have let a new employee do this on day one, unsupervised? — No answer. Nothing.
  • 10. Who approved capability thirty-one? — Answered. PG&E: “No named SME, no plugin.” Owner and risk tier at intake; findings return to intake; “No silent behaviour drift.”
  • 11. Who is accountable for AI-generated code accumulating in our internal tooling? — Answered. PG&E: “Zero SDLC gates removed, 100% of approvals human-owned, 100% of reviews human-in-the-loop.” OG&E uses its own AI to check its integrator’s specifications, code and test scripts.
  • 12. What exactly is the agent reading, and can we vouch for it? — Partial. The preconditions were named (SAP: the standard process you implemented is the ontology; Avista: business logic lived in reports and scripts). No utility said it can vouch.
  • 13. Is the data our agents read governed, current and traceable for the regulator? — Partial. Central Hudson’s 79,000-bill comparison shows traceability for a migration (Agents on the Slides), not for an agent.
  • 14. Does this data mean what the AI thinks it means? — No answer. Nothing direct.
  • 15. What to federate, what to replicate, and what to leave alone? — Answered. Avista: Business Data Cloud for the SAP domains, Databricks for GIS, outage and ADMS data, data products shared without re-replication; “retire, don’t replicate”; “complementary, not competing.”
  • 16. What customer does a customer-facing agent actually see — real, current, consented? — No answer. SAP’s Utilities Customer Self-Service Agent is available; no utility was named as running it.
  • 17. Could a customer-facing agent see one accurate, consented view of each ratepayer, and know where its authority stops? — No answer. Nothing from a utility.
  • 18. Which product owns which job, and which one owns the truth? — Partial. Avista’s domain split answers it for analytics. For master data, see question 45.
  • 19. Which of our integrations sit on the wrong side of the published-API line? — No answer. Nothing from a utility.
  • 20. Which of our AI experiments are driving in an endorsed lane? — No answer. No utility addressed it. The question got sharper: PG&E’s heaviest AI-assisted S/4HANA work runs on a hyperscaler’s coding agent and SAP-specific MCP server, not on Joule.
  • 21. What happens to our limits when agents multiply the traffic? — No answer. Nothing.
  • 22. Have you run the three-pile sort on your integration landscape? — No answer. Nothing.
  • 23. Have we modeled 2027 API consumption against the documented limits? — No answer. Nothing.
  • 24. Can a utility predict and control what an autonomous SAP process will cost? — Partial. SAP published the unit prices (Standard action 0.02 AI Units, Advanced 0.05), a 12-month expiry and near-real-time visibility (SAP commercial guide, Oct 2026 edition). The mapping from a process to a number of actions is still missing (question 30).
  • 25. How will we treat metered AI consumption in our rate case? — No answer. Nothing. New York asked for an inventory and policies, not cost.
  • 26. Have we drawn the line between the read tier and the act tier in writing? — Partial. Closest written lines: Nexus Water’s no-fly zones and approvals; PG&E’s human-owned approvals; SAP’s exception-resolution agent starts “upon user request.” Joule authorization reported in the room as all-or-nothing, so the platform cannot yet enforce the line.
  • 27. Are we paying generative prices for deterministic work? — Partial. OG&E keeps billing and financial close on deterministic, monitored job chains with AI beside them, not inside. Nobody priced the split.
  • 28. Has anyone put Joule Work Desktop through a security architecture review? — No answer. No utility reported one.
  • 29. Have we classified any BTP AI pilots against Annex III? — No answer. Nothing from a utility. SAP AI Agent Hub records carry an EU AI Act class (tooling only).

C. Only SAP can answer

3 answered · 13 partial · 12 no answer. SAP answered about platform, inventory and dates. It answered almost nothing about cost per process or the operational boundary.

  • 30. How many actions will a given business process generate, in our configuration? — No answer. Not published by SAP. Third-party estimates exist; none scored.
  • 31. What precisely is one action, in each agent implementation we license? — Partial. SAP’s generic definition: “a single step executed by an AI agent as it completes a task.” Per-implementation definitions, and which actions are Advanced, are not published (SAP commercial guide, Oct 2026 edition).
  • 32. How do retries, branches and tool invocations map onto actions? — No answer. Nothing.
  • 33. Can consumption be attributed to a named business process or cost centre? — No answer. Nothing.
  • 34. Can we impose a hard ceiling, and what happens at it? — Answered. No automatic cap. Customers “can proactively deactivate agents or features directly in SAP for Me”; otherwise overage is billed “under excess-use terms” (SAP commercial guide, Oct 2026 edition). A switch thrown by hand is a control, not a cap.
  • 35. Can SAP return a pre-execution estimate, or a bound, before an intent is dispatched? — No answer. Nothing new.
  • 36. Will there be volume tiers in 2027? — Partial. SAP now has two tiers by action type (Standard, Advanced), not by volume. Volume tiers for 2027 are unpublished.
  • 37. Where does the overage penalty land in the new model? — Partial. Overage is billed at the AI Unit rate under excess-use terms. The effect of the earlier 30-point discount reduction is not addressed.
  • 38. What does Joule Studio runtime cost after the promotion ends? — Partial. Production runtime is free through the end of 2026, then billed on compute, storage, memory, executions and tokens. Rates not published.
  • 39. Is the architecture we intend to run on the endorsed list? — Partial. SAP’s API policy requires “SAP-endorsed” pathways for large-scale replication and for AI that executes sequences of API calls, but does not list which architectures are endorsed.
  • 40. Is remediation of integrations on non-published APIs inside our RISE scope, or ours? — No answer. Nothing.
  • 41. Is your standard cited in the Official Journal? — Partial. EN 18286, the AI quality-management standard, was cited on October 14 (Decision (EU) 2026/2274). Whether SAP’s claims rest on it is unstated.
  • 42. What is your plan for EN 18286, and what happens to your position the day it is cited? — No answer. The day came on October 14. No SAP statement found.
  • 43. What does Verification Seal verification test, and who grants, revokes and audits it? — No answer. Nothing.
  • 44. Does agent inventory extend to tool-level dependencies, or only to agents? — Partial. SAP AI Agent Hub inventories SAP, custom and third-party agents plus MCP servers and LLMs; a LeanIX AI governance assistant is dated Q4 2026. Tool-level dependency mapping is not stated.
  • 45. MDG or Reltio: which one owns business partner, premise, device and connection object? — Partial. SAP calls them complementary, and Reltio’s matching is “coming very soon.” The IS-U ownership question is unanswered.
  • 46. Which agent actions belong on Integration Suite and which on n8n — and where does each stop at a switching order? — Partial. Both appear as SAP build tools. Neither has a stated stopping point. n8n inside Joule Studio is not yet generally available.
  • 47. When does a tabular foundation model (TabPFN / RPT) become something a utility can license for a rate case? — Partial. Available in SAP AI Core since September, with no GA status, pricing or metric stated.
  • 48. What is generally available today? — Answered. Dated on SAP’s utilities roadmap slides: self-service and case-management agents now; billing-document and asset-health agents Q4 2026; exception-resolution, BPEM and tariff agents Q2 2027; DER agents at ideation. Detail in Agents on the Slides.
  • 49. How many agents are live against SAP’s 400-agent year-end target? — Answered. About 300 agents and 40 assistants, “quality over quantity” (Q3 earnings call, per published transcript). No utility named.
  • 50. Where does agent authority stop at the OT boundary — SCADA, ADMS, OMS? — No answer. Asked in seven of ten articles. Still no SAP answer. FERC staff now tells entities to consider the business need for AI tools inside a CIP-networked environment (FERC staff report, Oct 13, 2026).
  • 51. Will SAP publish a NERC CIP mapping for the agent governance stack? — Partial. FERC staff mapped AI tools to CIP-004, -005, -007, -011 and -013 (FERC staff report, Oct 13, 2026). SAP’s own mapping is still missing.
  • 52. What is the agent identity lifecycle: provisioning, rotation, certification, deprovisioning? — Partial. SAP Cloud Identity Services gives agents their own identities, provisioned at registration. Rotation and deprovisioning are not stated.
  • 53. Will there be SAP-aware SOAR actions: suspend an agent, quarantine its outputs, revoke its seal? — No answer. Nothing.
  • 54. Are agent actions distinguishable from human actions at every hop? — No answer. Not answered. FERC staff names the risk: AI tools may “mimic user behavior.”
  • 55. Will there be a utilities lane in the AI Golden Path? — Partial. Not in the Golden Path, but SAP’s utilities slides now show “Industry AI / FDE Agents” built by forward-deployed engineers, with a ten-week proof of concept that becomes product (Agents on the Slides).
  • 56. Will there be right-sized BDC entry tiers for co-ops and municipals? — No answer. Nothing.
  • 57. Will security for the autonomous enterprise stay out of the upsell? — No answer. Nothing.

D. Peers, for each other

0 answered · 5 partial · 4 no answer. In presentations, partly. In numbers, not yet.

  • 58. Has anyone negotiated a hard cap or alert threshold on AI consumption, and what did it cost elsewhere? — No answer. Nothing.
  • 59. Has anyone built a capability registry that survived a real backlog? — Partial. PG&E built the design; OG&E routes 80+ projects through governed intake. Neither shows it surviving a backlog yet. Avista: ~1,500 applications, “almost none of it documented in processes.”
  • 60. Has anyone taken the accountability question to internal audit? — No answer. Nothing.
  • 61. Has anyone asked a vendor whether its standard is cited in the Official Journal? — No answer. Nothing.
  • 62. Is anyone treating a CRA declaration as CIP-013 evidence? — Partial. No utility reported it. FERC staff now places AI vendor risk inside the CIP-013 supply-chain plan, using NIST’s AI Risk Management Framework.
  • 63. Has anyone inventoried the AI experiments touching SAP data, or modeled agent-era API consumption? — Partial. OG&E tracks its AI estate and the value it delivers. New York utilities must inventory theirs by November 16.
  • 64. What did your account team say about which master data product you are meant to be on? — No answer. Nothing from a peer.
  • 65. Has your commission asked what AI costs, or how it is recovered? — Partial. New York asked for inventory, policies and risks — not cost. California’s SB 1011 (human approval for high-risk utility AI) was moved to the inactive file in May; it is not law.
  • 66. If you have numbers, bring them — compare before renewal season. — Partial. The venues exist: the IBM AI Lighthouse Council (seven utility members), SAP’s customer-influence program, the EA network. No numbers compared yet.

The sixteen new questions

San Antonio and SAP’s October commercial guide raised sixteen new questions, from whose utility ontology SAP is building to what makes an agent action Advanced. They are on the full list, marked (San Antonio), and are not scored here. They carry into the next round.

Correct the record

If you answered one of these questions in San Antonio, or anywhere in public, and it is not reflected here, tell me on Utility.Community and I will correct the score. Anonymous numbers are welcome in room D.