Dipp AI Research · Volume I
The Enterprise Superintelligence Report
Published 10 August 2026. A field study of what enterprises actually deployed, what went wrong, and the architecture the evidence points to.
37 pages · 191 citations · Free to read
Figure · Orcher™
Executive summary — headline findings
FIG. P03-01
78%
of enterprises run AI agents in production
Vol. I, p.3
28%
can trace an agent action end to end
Vol. I, p.6
54%
reported an agent-related incident
Vol. I, p.2
38%
can name an accountable owner
up from 7% — Vol. I, p.2
Contents
- 01
The trust gap
Production adoption at 78% against 28% traceability, and the incidents that followed.
- 02
Oversight theater
Why Human-in-the-Loop fails under volume, bias, timeouts, and missing authority.
- 03
Human-in-the-Role
Binding action to the authority of a named professional, enforced at execution.
- 04
The control plane
Seven components across two layers, and the failure mode each one answers.
- 05
Protocols and fabric
MCP, A2A, ACP, and why interoperability is not authorization.
- 06
Economics
The Big Model Fallacy, a 4,500× price spread, and verified execution as the unit of value.
- 07
Standards and outlook
Regulatory regimes, benchmark gaps, and where accountability settles next.
Headline findings
What the 2026 data showed
78%
of enterprises run agents in production
p.2
54%
reported an agent-related incident
p.2
34.8%
of AI traffic carries sensitive data
p.24
4,500×
price spread across models and providers
p.25
Chapter 04 — the control plane
Seven components, cited passage by passage.
Each component below carries the passage of Vol. I that specifies it. The interactive control-plane map on the Orcher page links directly to these anchors.
- 01
Directive Interface
Vol. I, p.2
“Seventy-eight percent of enterprises run agents in production, and fifty-four percent reported an agent-related incident, yet only thirty-eight percent can name an accountable owner for an agent decision. Accountability cannot be reconstructed from a prompt log after the fact; it has to be recorded as the first act of execution.”
- 02
Role Identity Fabric
Vol. I, p.6, p.11, p.22
“Non-human identities now outnumber human identities 144:1, up from 92:1, while ninety-two percent of organizations have no visibility into agent identities and eighty-six percent cannot enforce policy against them. A registry entry records that an agent exists; it does not constitute authorization to act.”
- 03
Logic Scrubber
Vol. I, p.6–p.7
“When AI advice was wrong, unaided human accuracy of eighty-two percent fell to forty-five and a half percent. In the analogous security-operations case, 2,992 alerts per day yielded a thirty-seven percent investigation rate, and gateway approvals auto-proceeded after a thirty-second timeout. Verification against a system of record does not fatigue.”
- 04
Immutable Audit Ledger
Vol. I, p.5, p.8
“The July 2026 registry-proxy incident produced more than 17,000 attacker actions across a five-day detection gap. Five percent of CISOs are confident they could contain unauthorized agent behaviour, while forty-seven percent have already observed it. Telemetry answers what happened to the system; evidence answers who authorized it.”
- 05
Data Control Gateway
Vol. I, p.24
“The share of enterprise AI traffic carrying sensitive data rose from 10.7 percent to 34.8 percent, against 410 million recorded DLP violations and retention exposure of up to seven years under the policies surveyed. A vendor default that changed last quarter is not a control.”
- 06
Cost Governance
Vol. I, p.9, p.25
“Across more than 400 models and 70 providers the price spread reaches 4,500×, and sixty-eight percent of enterprise AI programmes run over budget. The Big Model Fallacy is the assumption that the frontier tier is the correct default; stakes, not habit, should select the model.”
- 07
Observability
Vol. I, p.6–p.7
“Only twenty-eight percent of organizations can trace an agent action end to end, and roughly sixty percent cannot terminate an agent mid-execution. Single-provider dashboards cannot show the cross-provider path an enterprise directive actually takes — precisely where accountability disappears.”
Select a component to reveal its Vol. I excerpt.
Fig. CP-01 · The seven-component control plane
Selected figures
Straight from the report
Figure · Orcher™
Case file — model registry proxy compromise, 21 July 2026
FIG. P05-01
Timeline
T0
Zero-day in registry proxy
A trusted supply-chain path is compromised.
T1
Sandbox escape
Execution leaves its intended boundary.
T2
17,000+ actions
Automated activity at machine speed, none of it attributable to a role.
T3
5-day detection gap
Nothing in the path was hashed, scoped or observed cross-provider.
Mapped in Vol. I to two components: Data Control Gateway (egress) and Observability (cross-provider trace).
Figure · Orcher™
Automation bias — measured accuracy under wrong machine advice
FIG. P06-02
| Condition | Reader group | Accuracy | Reading |
|---|---|---|---|
| Correct AI suggestion | Radiologists | ~80% | Machine agreement raises confidence and accuracy together. |
| Incorrect AI suggestion | Inexperienced readers | <20% | Advice overrides independent judgement almost entirely. |
| Incorrect AI suggestion | Experienced readers | 45.5% | Experience mitigates but does not remove the effect. |
| Unaided baseline | All readers | 82% | The reviewer performs better with no machine advice at all. |
Figure · Orcher™
Incident-to-component mapping
FIG. P08-01
| Incident | What actually failed | Orcher component that answers it |
|---|---|---|
| Registry proxy compromise | Uncontrolled egress and no cross-provider trace | Data Control Gateway + Observability |
| Destructive production delete | No verification against the system of record before commit | Logic Scrubber |
| Ignored stop instruction | Authority was never bound to a role that could revoke it | Role Identity Fabric |
| Covert crypto-mining workload | No tamper-evident record of what executed and why | Immutable Audit Ledger |
Figure · Orcher™
Two letters, one regulation — July 2026
FIG. P10-01
July 2026
JUL 13
“We Must Act Now”
200+ economists and 16 Nobel laureates call for binding limits on autonomous systems.
JUL 15
China AI Agent Law in force
Three-tier decision authorization: what an agent may decide is set by law, not by configuration.
JUL 24
Open Weights letter
The opposing case for open model release and distributed capability.
Alongside: Illinois third-party audit requirements, the EU AI Omnibus, and roughly 100 measures across 38 US states.
Figure · Orcher™
Vendor training-policy comparison
FIG. P24-01
| Provider class | Enterprise default | The caveat that matters |
|---|---|---|
| Frontier API providers | No training on API inputs | Consumer and free tiers differ; retention windows are set by policy and by court order. |
| Hyperscaler AI services | No training on customer data | Scope depends on the specific service and region selected. |
| SaaS applications with embedded AI | Training on by default in several products | Roughly 300,000 organizations affected by a single 2026 policy switch, with retention up to seven years. |
| Open-weight self-hosted | No provider training path | Control moves to the operator, along with the obligation to prove it. |
Orcher Component Schematic v1 — the directive execution path
Fig. ORC-19
Layer 1 · Accountability — sequential, halt-by-default
L0 · Entry
Directive
A professional states intent in plain language, inside a declared role.
C1 · Layer 1 · Gate 1
Directive Interface
Intent is parsed into a scoped, bounded task with explicit limits.
C2 · Layer 1 · Gate 2
Role Identity Fabric
The task is bound to accountable identity and its entitlements.
C3 · Layer 1 · Gate 3
Logic Scrubber
Policy and systems of record are checked before anything commits.
C4 · Layer 1 · Gate 4
Immutable Audit Ledger
The decision path is hashed into evidence that survives challenge.
OUT · Exit
Committed action + Dipp Intelligence
The action executes under verified authority; the record stays with the firm.
Layer 2 · Control — continuous, applied to every stage above
C5 · Layer 2
Data Control Gateway
Data boundary, redaction and no-training guarantees enforced in path.
C6 · Layer 2
Cost Governance
Economic ceilings bound autonomy before spend runs away.
C7 · Layer 2
Observability
One view across every model, framework and agent in flight.
Download
Take Vol. I with you.#
The complete 37-page PDF, every citation intact. Access is gated: tell us who you are and where to send it, and the download unlocks as soon as the form is submitted — with a copy of the link emailed to you.
Gated download · one form, then the file
Read it, cite it, challenge it.
Vol. I is free to read and intended to be argued with. If you have a counter-citation, send it.