A private AI deployment is not proved by a polished executive demo. It is proved when the organisation can name the authorised systems, data fields, identities, actions, approval boundaries, audit record and rollback path behind one useful workflow. VERTU VPS is positioned as an AI-native enterprise interface above authorised business systems, with deeper native experience available on selected VERTU devices including AlphaFold depending on availability and configuration. That makes pilot design more important than feature count.
The right first pilot is deliberately narrow: one executive decision, a controlled data set, a small user group and a measurable human outcome. It should show what VPS can see, what it may prepare, what requires confirmation and what it cannot do. The 12 questions below convert ‘private deployment’ from marketing shorthand into an evidence package an executive, security team and system owner can review.
> The short answer: Do not begin with enterprise-wide access. Begin with one authorised workflow, named owners, least privilege, human confirmation, immutable audit evidence and a tested kill switch. Scale only after the pilot proves value and control on the exact configured environment.
What is confirmed, and what is not
The current VERTU product knowledge base, revision 136, describes VPS as an authorised overlay above existing ERP, CRM, finance, approval, inventory, sales, reporting and messaging systems for eligible clients, depending on deployment. It explicitly says VPS does not replace existing systems or human judgement. Hermes Agent can extend into authorised workflows through VPS, with significant actions requiring user approval.
Capability and availability vary by client environment, service agreement, region, device, integration and software version. The official live VPS, Hermes Agent and AlphaFold pages control changing commercial facts. This article is an implementation-question framework, not a promise that every named integration or workflow is available by default.
authorised-system pilot readiness matrix
| Check | Evidence or current signal | Decision use |
|---|---|---|
| 1. What decision is the pilot improving? | One executive decision has an owner, frequency, current delay and measurable outcome | Scope anchor |
| 2. Which systems are authorised? | Every ERP, CRM, finance, approval, inventory or messaging source is named | Explicit allow-list |
| 3. Which data fields are necessary? | A field-level minimum dataset supports the decision | Data minimisation |
| 4. Who can see and ask? | Pilot users, roles, device requirements and authentication are defined | Identity boundary |
| 5. What can VPS prepare? | Summaries, briefs, recommendations and drafts are listed | Read or prepare lane |
| 6. What can it action? | Every write, approval, message or system change has a risk level | Action allow-list |
| 7. How is consent and confirmation shown? | The user sees target, payload, consequence and reversibility before significant action | Human control |
| 8. What is logged? | Identity, timestamp, source, prompt or request reference, retrieved evidence, output, approval and result are traceable | Audit requirement |
| 9. How are errors contained? | Stale data, permission error, hallucinated field, connector outage and conflicting source have defined states | Fail visibly |
| 10. How is access revoked? | A kill switch disables connector, workflow, user, device session and stored memory as appropriate | Rollback gate |
| 11. What proves business value? | Baseline and post-pilot measures cover time, decision quality, exception capture and user burden | Outcome evidence |
| 12. What earns the next scope? | A written scale rule requires control and value thresholds | No automatic expansion |
A strong score in one row cannot cancel a hard failure in identity, safety, permission or recovery. Date every fact that can change and keep market, account, device or object configuration attached to the evidence.
1. What decision is the pilot improving?
Choose a recurring decision such as approval triage, cash exception review or sales risk briefing. Describe the current path and the human responsible. ‘Use AI across the business’ is not a pilot. A strong outcome might be fewer systems opened, faster review of exceptions or better evidence attached to an approval—without reducing required control.
Decision use: Scope anchor. Evidence to keep: One executive decision has an owner, frequency, current delay and measurable outcome.
2. Which systems are authorised?
Record system, tenant, environment, owner and integration method. Do not connect production simply because a demo connector exists. Use a sandbox or controlled dataset where possible. VERTU VPS is positioned above authorised systems; unauthorised discovery or broad credential reuse would contradict the deployment boundary.
Decision use: Explicit allow-list. Evidence to keep: Every ERP, CRM, finance, approval, inventory or messaging source is named.
3. Which data fields are necessary?
Map each field to the outcome and exclude sensitive columns that do not change the decision. Mask or tokenise where appropriate. Define retention and whether prompts, summaries, embeddings, logs or temporary files persist. ‘Read access to CRM’ is too broad; a pilot should state the exact objects and fields.
Decision use: Data minimisation. Evidence to keep: A field-level minimum dataset supports the decision.
4. Who can see and ask?
Separate executive, delegate, system owner, administrator and support roles. Use managed identities, strong authentication and revocation. If AlphaFold or another selected device provides a deeper native experience, confirm actual configuration and eligibility. Device ownership alone must not grant broad enterprise data access.
Decision use: Identity boundary. Evidence to keep: Pilot users, roles, device requirements and authentication are defined.
5. What can VPS prepare?
Define non-destructive outputs separately from actions. An executive brief can surface exceptions and context; it should show source and time so the reader can verify. Business analysis remains support for human judgement, not legal, financial, tax, accounting or investment advice. State confidence and unknowns rather than presenting a single generated narrative as fact.
Decision use: Read or prepare lane. Evidence to keep: Summaries, briefs, recommendations and drafts are listed.
6. What can it action?
Classify actions from read-only to draft, user-confirmed execution and multi-level approval. The knowledge base describes L1–L5 risk control and human-in-the-loop boundaries in supported deployments. Do not assume those controls are configured identically for every client. For the pilot, name the exact action and the required confirmation.
Decision use: Action allow-list. Evidence to keep: Every write, approval, message or system change has a risk level.
7. How is consent and confirmation shown?
A generic ‘approve’ button is not sufficient for high-consequence operations. Show what system will change, which records, which recipients and whether the action can be reversed. Hermes Agent is described as working with user approval. The interface and audit record should make that approval meaningful rather than ceremonial.
Decision use: Human control. Evidence to keep: The user sees target, payload, consequence and reversibility before significant action.
8. What is logged?
Logs should allow a reviewer to reconstruct why the workflow ran and what happened without unnecessarily copying private source content. Protect the audit trail from ordinary user editing. Define who can read it, how long it is retained and how it links to existing enterprise records. A screenshot of a successful demo is not an audit system.
Decision use: Audit requirement. Evidence to keep: Identity, timestamp, source, prompt or request reference, retrieved evidence, output, approval and result are traceable.
9. How are errors contained?
The pilot should return ‘source unavailable’, ‘not authorised’, ‘data not mature’ or ‘human review required’ instead of inventing a complete answer. Test known failures. A business interface becomes dangerous when it quietly replaces missing evidence with plausible prose. Route unresolved items to a named person and preserve the reason.
Decision use: Fail visibly. Evidence to keep: Stale data, permission error, hallucinated field, connector outage and conflicting source have defined states.
10. How is access revoked?
Test revocation before launch. The Hermes Agent experience is positioned with reviewable and revocable permissions, integrations, sessions and memories where supported. The client must verify the controls in its configuration. A pilot is not controlled if deprovisioning requires an improvised support escalation.
Decision use: Rollback gate. Evidence to keep: A kill switch disables connector, workflow, user, device session and stored memory as appropriate.
11. What proves business value?
Count systems opened, minutes to decision, exceptions missed, rework and audit completeness. Include false positives and cases where the system correctly refused. Do not use number of prompts as the primary success metric. Compare the same workflow before and after with a small sample large enough to expose variation.
Decision use: Outcome evidence. Evidence to keep: Baseline and post-pilot measures cover time, decision quality, exception capture and user burden.
12. What earns the next scope?
Define minimum pass rates for source grounding, permission enforcement, audit completeness, user confirmation and outcome improvement. Name vetoes such as unauthorised data, untraceable action or repeated source failure. Expansion should add one system, user group or action class at a time, with revisioned evidence and rollback.
Decision use: No automatic expansion. Evidence to keep: A written scale rule requires control and value thresholds.
Put the framework into a real decision
A useful pilot might cover a daily executive approval brief. VPS reads a limited set of authorised approval records and finance context, highlights exceptions, links each observation to source data and prepares a prioritised brief. The executive can open the evidence, reject the summary or confirm a permitted action. A write to the approval system requires a separate confirmation showing record and consequence. Every step enters the audit trail.
Before launch, the team injects four failures: missing finance source, stale approval status, user without permission and a conflicting system value. The interface must stop or label uncertainty, not smooth it over. After two weeks, compare decision time, reopened approvals, missed exceptions and user corrections with the baseline. Only a pilot that proves both usefulness and control should add another workflow.
A decision record another person can audit
Create a dated record containing the exact object, device, system or route; the source URL and access time; the person responsible for verification; and the condition that would change the recommendation. Preserve earlier observations instead of overwriting them when facts move. For product, policy and service claims, record the market and configuration. For private identifiers, store a redacted public copy and a controlled private copy.
End with one of four outcomes: proceed, wait, request evidence or decline. State the reason in one sentence, list the unresolved risk and name the next review date. This makes later performance review more honest: an editor can see what was knowable at publication rather than judging the decision with hindsight.
Action checklist
Name one decision and owner.
Allow-list exact systems and fields.
Define user and device eligibility.
Separate read, prepare and action lanes.
Require meaningful confirmation.
Create immutable audit lineage.
Test five failure modes.
Prove revocation and rollback.
Measure an executive outcome.
Set vetoes and a staged scale rule.
Recheck live product availability.
Preserve the exact configuration revision.
Related VERTU reading
These links cover adjacent decisions. They do not replace the current primary source or the exact configuration under review.
Sources and verification
Sources were accessed on 21 July 2026. Product availability, software features, laws, official alerts, policies, prices and service terms can change. Recheck the live source before acting. Legal, financial, insurance, health, travel-security and conservation references are general information rather than individual professional advice.
Final view
Private deployment should mean explicit authority, limited data, visible control and reproducible evidence. A VERTU VPS pilot earns trust by improving one real executive decision while making permissions, sources, confirmation, failure and rollback inspectable. Begin narrow, test hostile and missing-data cases, and scale only from a verified configuration—not from the ambition of the demo.
