Overview
Governance overview
Portfolio, obligations, and control status as of batch 2026-07 — covering internal agents, vendor-platform agents, and every team in scope. Six items need attention before the 27 Jul cycle.
8 +4
Agents governed
8 internal · 4 vendor platforms
23/23
Use cases tiered & owned
100% coverage
6
Needs attention
1 sign-off · 5 client inputs
0
Open AI incidents
1 near-miss YTD · closed
43
Decisions on record
append-only · 0 edits
Dimension status
Weight = regulatory pressure + data exposure + gap · recalculated per batch
| Dimension | Pillar | Weight | Maturity | Open items | Next review |
|---|---|---|---|---|---|
| D6 · Data Authorization | Foundations | 9 · Critical | 1.0 / 4 | 1 rule review due | 27 Jul |
| D2 · Exposure Tiering | Outward-facing risk | 9 · Critical | 1.5 / 4 | 0 | 27 Jul |
| D10 · Audit & Assurance | Technical assurance | 9 · Critical | 1.0 / 4 | 2 evidence gaps · T4 | 27 Jul |
| D8 · Incidents | Outward-facing risk | 8 · Critical | 1.0 / 4 | runbook in draft | Sep 26 |
| D1 · Decision Rights & Portfolio | Strategy | 7 · Elevated | 2.0 / 4 | 1 sign-off · 2 assessments | 27 Jul |
| D5 · Vendor Exposure | Economics | 6 · Elevated | 1.5 / 4 | 4 embedded-AI flags | Q3 scan |
| D7 · Cost & Value Activated | Economics | 6 · Elevated | 1.5 / 4 | 3 metered lines on watch | Monthly |
| D9 · Workforce Capability | Strategy | 5 · Standard | 3.0 / 4 | leadership session Aug | Aug 26 |
| D11 · Brand & Trust | Outward-facing risk | 5 · Standard | 1.0 / 4 | profile in draft | Q4 26 |
| D3 · Lifecycle Health | Foundations | 4 · Standard | 2.0 / 4 | 5 graduations blocked | 27 Jul |
| D4 · Evaluation | Technical assurance | Deferred · DEC-0022 | — | live-state preview ready | Q1 27 |
| D12 · Agent Safety | Technical assurance | Deferred · DEC-0022 | — | 2 platform-agent flags live | Q1 27 |
Tasks due before next batch
27 Jul 2026
| Payroll calendars — scope decision on Pay Period rulePilot-client audit finding · awaiting sponsor & COO-office sign-off | Sign-off | D1 |
| Rate conversion — system-ready rate exemplarPending from client · graduation-to-skill trigger | Client input | D3 |
| JScript write-audit — object-model export + corpusPending from client · prompt → skill graduation | Client input | D3 |
| Inforce reconcile — tier/plan crosswalk + match keysPending from client | Client input | D3 |
| Verify BAA eligible-services listMonthly re-verification · rule set v1.2 | Scheduled | D6 |
Recent decisions
append-only ledger · newest first
| DEC-0043 | Agent composition standard adoptedUse cases cluster by workflow · skill · data affinity · 11 Jul | Approval |
| DEC-0042 | Cost & Value activated ahead of scheduleToken economics adopted as programme module · 10 Jul | Activation |
| DEC-0041 | Add-in surfaces set to Never for member dataPending BAA coverage change · 30 Jun | Rule |
Workspace · Agent registry
Agent registry
Every agent in scope for the organisation — 8 internal agents composed from the 23 Phase 1 use cases, and the vendor-platform agents arriving inside existing contracts. Internal agents cluster on workflow affinity · skill affinity · data affinity. Open any agent to see its composition.
All · 12BenAdmin · 5ACA · 3Vendor platforms · 4T4 · 3Awaiting input · 4
BenAdmin agents · 5
composed from 13 team use cases · live build · Jul 2026
Case Setup AgentTurns client & carrier documents into system-ready configuration
4 casesT3Projects + files1 sign-off · 2 inputsDec 26
Affinity basisworkflow — implementation build · skill — extract → transform → propose · data — implementation workbooks & carrier documents
| Use case | Source | Vehicle | Boundary | Tier |
|---|---|---|---|---|
| Carrier tab fill | CSM-6 | Prompt + skill | Fills only mapped fields · output feeds apply-fills tool | T3 |
| Form library update | SS-3 | Prompt · draft | Propose-don't-execute · marked-up template + worksheet | T3 |
| Payroll calendars | SS-6 | Prompt v2 | Pay Period rule awaiting scope sign-off · closes 4 audit findings | T3 |
| Rate conversion | SCA-1 | Prompt | Transcribe-don't-invent · math shown per line | T3 |
Data Integrity AgentFinds what is missing, mismatched, or broken in enrolment data
3 casesT4Projects · paste-inCrosswalk pendingSep 26
Affinity basisworkflow — reconciliation & data QA · skill — compare & flag-don't-fix · data — census, coverage, carrier feeds
| Use case | Source | Vehicle | Boundary | Tier |
|---|---|---|---|---|
| Workbook gap check | CSM-1 | Prompt + skill | Flags gaps only · never fills | T2 |
| Inforce reconcile | EDI-3 | Prompt + skill | Paste-in ≤400/600 rows · above caps → skill route (DEC-0029) | T4 |
| Feed-error explain | EDI-4 | Prompt + golden set | Reads Error Library every run · explains, never edits feeds | T3 |
Config QA AgentTests changes, triages tickets, writes and audits platform scripts
3 casesT3Chat · ProjectsCorpus pendingJan 27
Affinity basisworkflow — change QA · skill — test, diagnose, audit · data — change tickets & platform scripts
| Use case | Source | Vehicle | Boundary | Tier |
|---|---|---|---|---|
| Test scenarios | FSQ-1 | Prompt + skill | Scenario scaffold + past-change exemplars | T2 |
| Ticket triage | FSQ-5 | Prompt + rule library | Routes by root cause · never adjudicates | T3 |
| JScript write-audit | SCA-3 | Prompt → skill | Object-model reference inline · graduates on client corpus | T2 |
Vendor & platform agents · 4 platforms
agents arriving inside existing contracts · same registry, same rules
| Platform & agents | Function | Access today | Governance status | Next |
|---|---|---|---|---|
| HubSpot BreezeCustomer · Prospecting · Data agents · Assistant on every seat | Sales & Marketing | Assistant live on all seats · agents credit-metered | Intake pending | Q3 intake |
| monday.com agentsNative agents shipped to all customers · Agent Accounts | PMO · case boards | Live in tenant, no setup required — boards hold client casework | Live in tenant | Toggle audit · Jul |
| Microsoft 365 CopilotCopilot chat + agents in Teams / SharePoint | Org-wide productivity | Features auto-enabled Q2 · per-seat licence not purchased | Inventoried | Q3 scan |
| Zoom AI CompanionMeeting summaries · drafting | Meetings org-wide | Default-on with paid seats · client calls in scope | Inventoried | Q3 scan |
Product & engineering AI enters the registry at Q3 intake — governance covers the whole organisation, not one workstream
Foundations · D6 · Data Authorization · Weight 9 · Critical
Data Authorization
The most consequential table in the workspace: which AI surface may touch which data class. PHI rows are decided by verified BAA coverage — never by convenience. One near-miss YTD was caught by exactly this discipline.
Coverage is not uniform across the vendor's own products. Chat and Projects sit under the executed BAA; the Excel/PowerPoint add-ins and the desktop agent product do not (today). The matrix below is why that distinction is enforced as a rule, not a habit. Monthly verification due 27 Jul.
Authorization matrix · rule set v1.2
R-101 – R-120 · verified 13 Jul
| Surface | PHI / member data | Client-confidential | Internal | Public | Rule |
|---|---|---|---|---|---|
| Claude chat BAA workspace | Min-necessary | Allowed | Allowed | Allowed | R-101 |
| Claude Projects BAA workspace | Min-necessary | Allowed | Allowed | Allowed | R-102 |
| Excel / PowerPoint add-ins not on eligible-services list today | Never | Never | Allowed | Allowed | R-104 |
| Desktop agent (Cowork-class) not covered today · re-review on vendor change | Never | Never | Case-by-case | Allowed | R-105 |
| Connectors (M365 · Monday · Zoom) Work Orchestration pilot scope | Min-necessary | Read scope | Allowed | Allowed | R-112 |
| AI Gym environment | Never | Anonymised only | Allowed | Allowed | R-110 |
| Personal AI accounts | Never | Never | Never | Allowed | R-120 |
Coverage verification history
monthly + on vendor change
| Date | Check | Result | Action taken |
|---|---|---|---|
| 13 Jul 26 | Eligible-services list vs live surfaces | Match | None required |
| 27 Jun 26 | Training material surface references | Mismatch found | Slide corrected pre-delivery · logged NM-001 · DEC-0038 gate added |
| 31 May 26 | Baseline — BAA scope vs planned surfaces | Match | Rule set v1.2 issued |
Workspace · Decision log
Decision log
Append-only record of approvals, rules, corrections, activations and deferrals — grouped by what each decision governs. Entries cannot be edited after commit; supersessions link forward.
Append-only ledger. 43 entries · 0 retroactive edits · integrity check passed at batch 2026-07.
AllApprovalsRulesCorrectionsActivationsDeferrals
| ID | Date | Decision | Type | Decided by | Links |
|---|---|---|---|---|---|
| Programme & tiering | |||||
| DEC-0043 | 11 Jul | Agent composition standard adoptedUse cases cluster into agents on workflow · skill · data affinity | Approval | Steering committee | Registry |
| DEC-0042 | 10 Jul | Cost & Value (D7) activated ahead of scheduleToken economics adopted as programme module · partially supersedes DEC-0022 | Activation | SC · CFO endorse | D7 |
| DEC-0035 | 19 Jun | Five-tier exposure model v1.0 adopted; all 23 use cases classified | Approval | Steering committee | D2 |
| DEC-0022 | 02 Jun | D4, D7, D12 deferred to Q1 27 reviewD7 portion superseded by DEC-0042 · deferral on record by design | Deferral | Steering committee | Scorecard v1 |
| Surfaces & data authorisation | |||||
| DEC-0041 | 30 Jun | Excel/PowerPoint add-ins → Never for member dataPending BAA coverage change · re-review on vendor change or quarterly | Rule | Steering committee | R-104 · D6 |
| DEC-0038 | 27 Jun | Training material corrected pre-delivery; surface gate added"Work with the real file" guidance restricted to covered surfaces | Correction | Compliance lead | NM-001 |
| BenAdmin agents | |||||
| DEC-0033 | 16 Jun | Work Orchestration Agent approved as pilot with connector scope limitsM365 + Monday + Zoom read scope · min-necessary on covered surfaces | Approval · cond | SC + Compliance | A-2026-016 |
| DEC-0029 | 09 Jun | Data Integrity paste-in caps set — ≤400 carrier / ≤600 census rowsAbove caps → skill route with file handling | Rule | Custodian + EDI lead | Inforce reconcile |
| Vendors & platforms | |||||
| DEC-0040 | 29 Jun | Quarterly embedded-AI vendor scan establishedPlatform agents added to scan scope Jul 26 | Cadence | Custodian | D5 · D12 |
Showing 9 of 43 · grouped by governed object · newest first within groups
Outward-facing risk · D2 · Exposure Tiering · Weight 9 · Critical
Exposure Tiering
Five tiers, adopted as v1.0 (DEC-0035), assigned by a scored five-question protocol — deterministic by design: same inputs, same tier, every time. The empty T5 row is the point: nothing here warrants prohibition-level exposure, and the model proves that rather than asserting it.
23 +8
Classified
use cases + composed agents
4
T4 use cases
quarterly re-tier · full packs
0
T5
prohibition list empty
Tier definitions · v1.0
| Tier | Meaning | Count | Controls |
|---|---|---|---|
| T1 | Internal · no client or member data | 0 | AUP only |
| T2 | Internal work product · anonymised or internal data | 6 | Tower-owner approval |
| T3 | Client-facing output · human reviews before send | 13 | Named reviewer · SC approval |
| T4 | Regulated-adjacent or connector-scope work | 4 | Joint approval · full evidence pack · quarterly re-tier |
| T5 | Prohibited — autonomous regulated decisions, PHI on uncovered surfaces | 0 | No path — bright line |
Scored protocol · 5 questions
0–14 → tier
| Question | Scores |
|---|---|
| Q1 · Highest data class touched | 0–3 |
| Q2 · Audience of the output | 0–3 |
| Q3 · Regulatory adjacency | 0–3 |
| Q4 · Error blast radius | 0–3 |
| Q5 · Autonomy of action | 0–2 |
0–2 → T1 · 3–5 → T2 · 6–8 → T3 · 9–11 → T4 · 12–14 or any bright-line hit → T5 · two assessors, independent scores
Technical assurance · D10 · Audit & Assurance · Weight 9 · Critical
Audit & Assurance
"A human reviewed it" must be provable. Every T3+/T4 workflow carries a four-part evidence pack — configuration, boundary rules, review record, outcome sample — assembled on demand for a client audit, a DOL inquiry, or an OCR question.
2/4
T4 packs complete
2 gaps in flight
2
Audit packs delivered YTD
sponsor questionnaire · broker RFP
< 1 day
Pack assembly time
target for any examiner ask
Evidence completeness — T4 and highest-volume T3
4 components per pack
| Workflow · agent | Config | Boundary rules | Review record | Outcome sample | Pack |
|---|---|---|---|---|---|
| File triage & fix-it note Filing Data Agent · T4 | Held | Held | Held | Held | 4/4 |
| Intake snapshot Filing Data Agent · T4 | Held | Held | Held | Held | 4/4 |
| Inforce reconcile Data Integrity Agent · T4 | Held | Held | Held | Gap | 3/4 |
| FLOW orchestration Work Orchestration Agent · T4 pilot | Held | Held | Gap | Held | 3/4 |
Gaps assigned with owners · close-by 27 Jul · OCR dry-run pack in draft as the Q3 exercise
Technical assurance · D12 · Agent Safety · Deferred · watch flags live
Agent Safety
No internal agent has write scope today — so full agent-safety controls are deferred honestly rather than performed. What is not deferred: the watchlist. Vendor platforms are shipping agents into the tenant right now, and named internal triggers would activate this dimension the day autonomy arrives.
Watchlist — live flags
reviewed each batch
| Flag | What's happening | Why it matters here | Status |
|---|---|---|---|
| monday.com native agents | Platform shipped agents to all customers, no setup required · Agent Accounts let external agents authenticate into tenants | Boards hold live client casework — an agent acting on them would be an ungoverned write actor | Live in tenant |
| HubSpot Breeze agent lineup | Customer, Prospecting and Data agents GA · Assistant on every seat · outcome-based pricing | Autonomous client-facing replies drafted from CRM data — brand and advice-bounds exposure | Intake pending |
| Recap & board write connector | Caseload Agent's natural next step is writing recaps onto boards directly | First internal write scope — crosses the autonomy line | Trigger armed |
| FLOW write-back | Work Orchestration pilot could graduate from read + draft to task write-back | Same line, wider surface (M365 · Monday · Zoom) | Trigger armed |
Activation switches on: autonomy fields in the registry · action logging · kill-switch standard · autonomy budget per agent
Workspace · Obligations
Obligations register
The regulatory surface Selerix's AI programme answers to — federal frameworks in force, state AI statutes, and what is on watch. Each obligation maps to the controls that satisfy it. Prepared with programme counsel input; not legal advice.
All · 10Federal · 6State · 4In force · 7On watch · 3
| Obligation | Status | What it demands of AI use | Mapped controls | Review |
|---|---|---|---|---|
| Federal | ||||
| HIPAA Privacy & Security Rules45 CFR 160/164 · PHI at the core of the platform | In force | Minimum-necessary PHI on covered surfaces only · BAA before any disclosure to an AI vendor | D6 · R-101–120 | Monthly |
| HIPAA Breach Notification45 CFR 164.400+ · 60-day clock | In force | AI data-handling failures assessed as potential breaches · evidence & timeline discipline | D8 · severity matrix | Per event |
| HIPAA Security Rule NPRMProposed rewrite · asset inventories, mandatory risk analysis | Proposed | If finalised: written inventory of systems touching ePHI — the AI registry becomes compliance evidence | Registry · D6 | Quarterly |
| ERISA fiduciary duties + DOL FAB 2026-01Process-not-results standard for plan administration | In force | Documented, prudent process where AI touches plan administration · human accountability preserved | D1 matrix · D10 packs | Quarterly |
| DOL cybersecurity guidanceExtends to all plan service providers | In force | AI vendors inside the security programme · client diligence answers ready | D5 book · D11 profile | Semi-annual |
| IRS ACA reporting §6055/60561094/1095 accuracy · penalty exposure per return | In force | AI-assisted filing work needs accuracy evidence and human sign-off before submission | Filing Data Agent · D10 | Filing season |
| State | ||||
| Texas TRAIGA (HB 149)Home state · effective 1 Jan 26 | In force | Prohibited-use lines · NIST AI RMF alignment is a statutory safe harbour — the Canon maps to RMF | Scorecard · policies | Quarterly |
| Illinois HB 3773AI in employment decisions · client employers affected | In force | No AI-driven adverse employment inference in Selerix workflows · client-facing answer prepared | D2 tiering · D11 | Semi-annual |
| Colorado SB 26-189Replaced repealed SB24-205 · obligations 1 Jan 27 | Jan 27 | Consequential-decision framework · registry + tiering position early compliance | Registry · D2 | Q4 26 |
| State AI patchworkMulti-state client base · new bills each session | Watch | Quarterly horizon scan folded into batch review | D5/D11 cadence | Quarterly |
10 obligations · 7 in force · 3 on watch · mapped controls current at batch 2026-07
Strategy · D1 · Decision Rights & Portfolio · Weight 7 · Elevated
Decision Rights & Portfolio
Who approves what, and the org-wide AI portfolio those decisions govern. Every use case has a named approver and a tier-driven approval path — the ERISA "prudent process" answer.
100%
Portfolio with named approver
no orphaned use cases
3
Pending decisions
oldest 9 days
4.2 days
Median time-to-decision
target ≤ 7
5
Functions in portfolio
Eng & Product intake Q3
Approval queue · 3
the assessment sits behind each pending decision
| Decision | Waiting on | Holder | Age |
|---|---|---|---|
| Payroll calendars — Pay Period scope ruleCase Setup Agent · pilot-client audit finding · option (b) recommended with named-reviewer check on every derived row | Sign-off | Sponsor & COO office | 9 days |
| New use case — Broker RFP supportScored 6 → Tier 3 proposed · conditions drafted: named reviewer, house voice pack, no commercial commitments from a draft | Compliance endorse | Compliance lead | 5 days |
| Engineering & Product AI intakeCode assistants + in-product AI · same intake door, same protocol · nothing scored until intake | Scheduled | Product leadership | Q3 |
Portfolio by function
org-wide
| Function | Agents | Use cases | Highest tier | State |
|---|---|---|---|---|
| BenAdmin operations | 5 | 13 | T4 | Live build · 4 awaiting inputs |
| ACA services | 3 | 10 | T4 | Gym live · promotion re-tiers |
| Sales & Marketing | — | 1 proposed | TBD | Broker RFP in queue · HubSpot intake pending |
| PMO & workplace | — | — | TBD | monday / M365 / Zoom platform agents inventoried |
| Engineering & Product | — | — | TBD | Q3 intake scheduled |
Decision-rights matrix
v1.0 · charter at sign-off
| Decision | Decides | Consulted | Informed |
|---|---|---|---|
| New use case · T1–T2 | Tower owner | Custodian | SC digest |
| New use case · T3 | Steering committee | Compliance | Sponsor & COO office |
| New use case · T4 | SC + Compliance jointly | Security | Executive team |
| Surface / data-class change | SC on custodian brief | Security officer | All users · AUP note |
| Exception to a Never rule | No path — bright line | — | — |
Outward-facing risk · D8 · Incidents · Weight 8 · Critical
Incidents
Zero open incidents — and a discipline built so the first real one is handled inside HIPAA's 60-day clock, not discovered against it. Near-misses are captured deliberately: the programme's one save so far produced a permanent gate.
0
Open incidents
YTD
1
Near-miss captured
caught pre-delivery → DEC-0038
3
Tabletops scheduled
W3 safety practices
Severity matrix
runbook v0.4
| Sev | Definition | Clock |
|---|---|---|
| S1 | Confirmed PHI on an uncovered surface / disclosure | Breach analysis immediately · 60-day notification clock governs |
| S2 | PHI-adjacent handling failure, contained before disclosure | 24-hr assessment · compliance determination |
| S3 | Erroneous AI output reached a client | 48-hr correction + client note |
| S4 | Boundary violation, internal only | Batch review |
| NM | Near-miss — caught before impact | Logged · pattern review quarterly |
Register
complete YTD
| ID | Event | Outcome |
|---|---|---|
| NM-001 | 27 Jun · Workshop slide suggested "work with the real file" on a surface outside BAA coverage · caught in pre-delivery review | Corrected before any exposure · surface gate added to all training material (DEC-0038) · closed |
Tabletop scenarios — scheduled, hypothetical
no real incidents behind these
| Exercise | Scenario | What it rehearses | Sev |
|---|---|---|---|
| TT-01 | A CSM pastes an un-anonymised client census into a chat while rushing a renewal — a colleague notices before send | Containment · who calls compliance · evidence capture in the first hour | S2 |
| TT-02 | An unreviewed AI recap with a stale renewal date is emailed to a client, who forwards it to their broker | Correction protocol · client note in house voice · named-reviewer gap analysis | S3 |
| TT-03 | An 834 segment containing member identifiers is pasted into a carrier portal's embedded AI chat to "decode an error" | Breach analysis · 60-day clock mechanics · vendor-surface boundary refresher | S1 |
Each exercise ends with a timed debrief against the runbook · findings feed runbook v0.5
Economics · D7 · Cost & Value · Activated · DEC-0042
Cost & Value
Activated ahead of the original deferral — because AI cost stopped being one licence line. The operating principle: cost follows autonomy. Humans-in-chat cost a flat seat; the moment work graduates to skills, agents and platform toggles, cost becomes metered.
Illustrative figures. Dollar amounts on this page are programme estimates at list pricing for teaching and planning — not invoices. Token economics is now a taught module in the AI Gym.
~$2.1K/mo
Estimated AI spend today
seats ≈ 90% of it
73%
Seat utilisation · 30 seats
9 daily · 13 weekly · 8 light
3
Metered lines on watch
tokens · credits · per-seat creep
Aug
Value evidence lands
blind-eval + time-capture readout
Cost lines — every source, one table
reviewed monthly
| Line | Model | Today · est. | What moves it | Watch |
|---|---|---|---|---|
| Claude seats · 30 | Fixed per-seat | ~$1.9K/mo | Headcount only — usage inside a seat is free at the margin | Stable |
| API tokens at skill graduation | Metered per run | ~$120/mo | Each prompt → skill graduation moves runs from seat to meter | Watch |
| M365 Copilot | $30/user/mo if licensed | $0 today | 30 seats would add ~$900/mo — governance position before licensing, not after | Creep risk |
| HubSpot Breeze credits | Credit-metered · outcome-priced | $0 today | Customer Agent bills per resolution once enabled — cost scales with autonomy | Watch |
| monday.com agents | Included today · pricing evolving | $0 today | Vendor pricing for agent actions under review industry-wide | Monitor |
| Zoom AI Companion | Bundled with paid seats | $0 marginal | No direct line — exposure is data, not dollars | Stable |
Unit economics — the teaching card
AI Gym token module
Inforce reconcile, one run: ~90K input tokens + ~8K output ≈ $0.31/run at API list. At 150 reconciles a month ≈ $47/mo — replacing hours of side-by-side comparison per run. Fix-it note, one run: ~25K in + ~3K out ≈ $0.10/run.
The lesson the cohorts learn: token cost is almost never the constraint — review capacity is. The economics justify far more automation than the named-reviewer gates currently allow, which is exactly why the gates, not the budget, are the control.
The lesson the cohorts learn: token cost is almost never the constraint — review capacity is. The economics justify far more automation than the named-reviewer gates currently allow, which is exactly why the gates, not the budget, are the control.
Usage distribution — 30 seats
Jul batch
| Band | Seats | Profile |
|---|---|---|
| Daily | 9 | ACA analysts + EDI — Gym habits carried into live work |
| Weekly | 13 | CSM + Shared Services — workflow-triggered use |
| Light | 8 | Adjacent roles — coaching targets for the W6 champion push |
Utilisation is the leading value indicator this early — the Aug readout pairs it with blind-eval quality scores and time capture