Dipp AI Research · August 24, 2026
How a standing grant is expressed, bounded, enforced, revoked and evidenced — the mechanics behind the phrase.
By Odero Otieno — Founder, CEO & CTO, Dipp AI Technologies, Inc.
A specification-level note on Human-in-the-Role: the structure of a standing grant, the bounds it carries, the evaluation performed at execution time, the escalation path when no grant applies, and the record the pattern leaves behind.
Human-in-the-loop is a workflow pattern. Human-in-the-Role is an authority model. The difference is not rhetorical: one places a person in the path of each item and depends on their attention, the other places a bounded, revocable grant in the path of a class of actions and depends on its enforcement. This note describes the second in enough detail to be argued with.
FIG. A01-02
The population an accountable owner would have to govern
Drag the inputs
Published 2026 estimates put the enterprise ratio between 45:1 and 140:1 depending on company size. Independent research finds roughly 28% of agent actions can be traced back to a human sponsor across every environment they touched.
400,000
Non-human identities in the estate
112,000
Of those, traceable to a human sponsor
288,000 unaccounted
| Field | Meaning | Example bound |
|---|---|---|
| Role | The accountable human role, not an individual account | Claims Adjudication Lead |
| Directive class | The class of action permitted | Adjudicate claim, value ≤ threshold |
| Value bound | Maximum consequence permitted per action and per period | $25,000 per action |
| Data class | Which classes the action may touch | PHI permitted, in-region models only |
| Jurisdiction | Legal regimes eligible to process | US only |
| Temporal bound | Validity window and review date | 90 days, reviewed quarterly |
| Escalation target | Where refusals go | Director of Claims Operations |
FIG. A02-01
Review after the fact, versus authority before it
Step through the cycle
Lane A — Human-in-the-Loop
01 Agent plans
The agent composes an action from context nobody scoped in advance.
02 Agent acts
Execution happens against the production system of record.
03 Reviewer sees a result
A plausible-looking summary arrives in a queue, on a clock.
04 Approve or miss
Under volume, approval becomes the default. Nothing proves scope.
05 Damage is historical
The record starts after the fact, if it exists at all.
Lane B — Human-in-the-Role
01 Directive issued
A named professional states intent under their own authority.
02 Role bound
Role Identity Fabric binds the directive to that authority, cryptographically.
03 Gates evaluated
Authority, action and data-boundary gates run in sequence, halting by default.
04 Action commits
Only a directive that cleared every gate reaches the system of record.
05 Evidence written
The Immutable Audit Ledger records the cycle — including any halt.
Approval systems accumulate permissions and rarely shed them, because nothing forces re-examination. A grant carries a temporal bound and a review date, so authority expires by default and must be renewed deliberately. Revocation is immediate and takes effect on the next evaluation, which means withdrawing authority does not require finding and disabling every agent that relied on it.
the unit of accountability
Not the individual, and not the agent
authority lifecycle
behaviour outside the bounds
The first objection is that this is simply role-based access control. It is not: RBAC governs what a principal may reach, while a grant governs what consequence an autonomous action may cause, bounded by value, data class and time, and it is evaluated per directive with the outcome written to evidence. The second objection is that it slows delivery. In practice it moves the conversation earlier — bounds are agreed once, in daylight, rather than negotiated repeatedly under incident pressure.
“An approval records that someone clicked. A grant records that someone is answerable. Only one of those survives an inquiry.”
Sources
Where a number comes from Dipp AI's own analysis or an observed deployment, it is labelled as such and is not presented as an independently audited third-party finding.
Case studies
Modelled reference scenarios with the measurement window, the components enforced and the numbers attached. Each one downloads as a PDF.
Government & Public Sector
Non-human identities outnumbered staff 144 to 1. Role-scoped authority made every one of them answer to a caseworker who could be named.
Writes the Dipp AI record on enforced governance for agentic systems — authority, enterprise data boundary, cost and compute. Every figure in this piece carries a source, and corrections are published on the record rather than made quietly.
