VMD AGENTIC
From operational signal
to governed decision.
An intelligence and orchestration layer being built above the systems a company already runs — designed to turn operational events into analysis, options, approved decisions and verified outcomes.
Why it exists
The data is already there. The decision is not.
Enterprise systems record what happened with great precision. What they rarely carry is the work that turns a record into a decision — and that work still falls to people crossing systems, reports and context by hand:
- Noticing that something changed, among everything that did not.
- Establishing why it matters, and to which objective.
- Working out what it is likely to cost if nothing is done.
- Assembling the alternatives, and what separates them.
- Deciding, with the evidence attached rather than remembered.
- Checking afterwards whether the decision produced what it promised.
Closing that gap is not the same as removing the people in it. Every part of this platform is built so that a person can see what it concluded and why, and material decisions are designed to stay with the people accountable for them.
Where it sits
Your systems remain. The layer above them is what changes.
VMD AGENTIC is not a replacement for the systems that run a business. Those systems hold operational truth, and they are where transactions belong. The platform is designed as a layer above them:
Systems of record
ERP, CRM, production and warehouse systems, finance systems, databases and documents. They continue to run the processes and to hold what actually happened. Nothing here asks them to move.
VMD AGENTIC
The layer that reads across them, works out what an event means against the objectives a business has set, and keeps the evidence for every conclusion it reaches.
Decision and governed action
What the organisation decides, recorded with what it was based on — and, where execution is later permitted, a bounded action followed by a check that it worked.
Designed to work across enterprise systems and data is the accurate description today. No production connector to any business system is built, and no specific vendor system is supported. The platform currently reads its own tenant data; how it reaches an external system is architecture, not a shipped integration.
The operating model
From event to outcome.
One sequence describes how the platform is meant to work, end to end. It is the product’s spine, and stating it in full is also the only honest way to show how much of it exists:
Observe
Take in the operational events and measurements that matter, with their provenance intact, so that everything downstream can be traced back to something real.
Analyse
Establish what happened, what is related to it, and whether the difference is material against a stated objective — rather than against a threshold buried in a report.
Predict
Where the data and a defensible method exist, state what is expected next as a falsifiable claim with its uncertainty attached, never as a confident number.
Recommend
Set out the available responses and what separates them, including the economic consequence of each.
Approve
Put any material or risky action behind an explicit human decision and an explicit policy, so that nothing consequential happens on the platform’s own authority.
Execute
Carry out only actions that are permitted, bounded and reversible in the ways the policy requires — and only in domains where execution has been enabled at all.
Verify
Check afterwards whether the action was actually taken and whether the expected outcome followed, and record the answer either way.
This is the operating model the platform is being built toward, not a description of what runs today. Observation and analysis are built, and prediction is built in the energy domain. Recommendation, approval, execution and the verification of an executed action are design: there is no approval workflow, no execution path and no connection to a business system in the platform today, and no stage will be enabled ahead of the evidence, permissions and verification that justify it.
What is built
The foundation is real, and deliberately unglamorous.
VMD AGENTIC is in development, and what exists today is the layer everything else has to be safe to stand on. Each of the following is implemented rather than specified:
Tenancy and isolation
Multiple organisations on one platform, separated in the database rather than in application code. Row-level policies deny by default, reads are gated on membership and writes on a named permission — so a query that forgets its tenant filter still returns nothing.
Identity, roles and administration
Authentication, a permission model, system roles, atomic tenant provisioning and membership administration, with the screens to operate them.
Audit and decision trace
Consequential actions are recorded against the tenant and identity that performed them and are readable back in the product. Separately, every figure the platform shows can be opened to what produced it — the inputs, the rule applied and the evidence underneath.
Objectives and materiality
What an organisation is trying to move is a versioned record with its own provenance, and whether a difference deserves attention is decided by materiality logic against that objective.
Operational attention
A standing answer to whether anything needs a person right now, assembled only from figures that can prove themselves — server-rendered, with no simulated activity and no placeholder charts.
Evidence discipline
Every number carries its type — measured, simulated, generated, forecast or unknown — beside the number itself, and the absence of a marker never means measured. Stable editorial description and a finding derived from live data are kept visibly distinct.
What is equally true: no agent runs on this platform today, and there is no model in any of the paths above. Every figure the product shows is computed from stored data by a deterministic rule, and the platform performs no action in any external system. That is the position the rest is being built out from, and it is a deliberate one.
Decision intelligence
What separates this from an assistant.
An assistant answers the question it was asked. A decision-intelligence platform is built around the sequence a decision actually needs, and is accountable for each step of it:
- What happened, and is it solid enough to act on?
- Why it matters — to which objective, and by how much.
- What is likely to follow if the situation is left alone.
- Which responses exist, and what separates them.
- What each one is worth, and what it would cost to be wrong.
- What should be approved, by whom, and on what evidence.
- What actually happened afterwards.
In the energy domain the platform reaches the middle of this sequence today: it establishes what happened, prices the exposure, and evaluates alternatives against a declared objective. It stops deliberately short of naming one of them as the recommendation — ranking alternatives, and the approval and execution that would follow, are not built.
Governance
Autonomy is earned, not assumed.
The word agentic is used across the industry to mean software that acts on its own. That is not the direction this platform takes. Autonomy here is something a capability earns, stage by stage, when evidence, policy, permissions and verification justify it — and the order it earns it in is fixed: read-only first, then recommendation, then human approval, then bounded execution, then verification.
Tenant isolation
Enforced in the database, denying by default, on every table.
Least privilege
Reads require membership, writes require a specific named permission, and privileged database functions are hardened rather than trusted.
Evidence before assertion
The platform may not call something a risk without a risk model, a priority without materiality logic, or a correlation a cause. Editorial description and derived finding never swap places.
Read-only by default
The platform performs no write into any business system. Execution is not switched off by configuration; it does not exist yet.
Approval for material action
Anything consequential is designed to require an explicit human decision under an explicit policy, rather than an agent’s own judgement.
Bounded tools and permissions
When execution is built, an action will be possible only where it has been permitted, within limits declared in advance — not wherever a model can reach.
Verification after the fact
A business-changing action is not complete when it is issued. It is complete when the platform has checked what followed.
The first four are enforced in the platform today. The last three describe how execution will be permitted once it is built; they are design commitments, and naming them here is not a claim that they are running. Untrusted content — documents, messages, free text from a database, and any future model output — is treated as data and never as instruction, and that boundary is enforced outside the model by design.
Domain intelligence
One platform, specialised by domain.
The platform holds what is generic — tenancy, permissions, objectives, evidence, trace and audit. A domain brings its own data model, its own calculations and its own evidence, and publishes into the generic model rather than bending it. The architecture is designed so that specialised intelligence can be developed around areas such as:
- Executive and organisational performance
- Operations
- Procurement and inventory
- Production
- Finance
- Maintenance
- Sales
One domain is built: energy. The rest name where the architecture is designed to extend — they are not products, there is no page behind them, and no date is attached to them. HELIX is a different thing again, and listed under its own heading in the portfolio: a standalone enterprise operational-intelligence application VMD configures around one organisation’s own systems — not a domain of this platform, and not built on it.
The first domain
VMD Energy Intelligence.
Energy is the first domain built on this platform, and the reason much of the foundation above has the shape it does. It applies the same approach — measured evidence, economic consequence, and a trace from every result back to the intervals that produced it — to commercial and industrial energy. It is the concrete case of what the platform is for, in one operational domain.
In development
What is being built next.
Named as work in progress rather than as capability, and listed apart from the foundation so that the two cannot be confused:
- The agent runtime, and the orchestration that coordinates work across specialised agents
- An approval surface where material actions are presented, decided and recorded
- Bounded execution, in a contained environment, for actions that have been approved
- Verification of an executed action against the outcome it was expected to produce
- A connector abstraction, so a domain can read from an enterprise system through a declared contract
- Broader decision workflows across more than one domain
No release date is attached to any of this, and nothing above may be read as available. Some of it will change shape as the foundation proves what it can carry.
Boundaries
What VMD AGENTIC is not.
Stated plainly, because each of these is a real product somebody sells, and none of them is this:
- Not a replacement for an ERP, a CRM or any other system of record. Those systems keep running the business, and this platform is built on the assumption that they stay.
- Not an autonomous AI workforce. Nothing here acts on its own authority, and the roadmap does not end with software that does.
- Not a chatbot placed over a database.
- Not a dashboard. A dashboard reports a figure; this is built to carry the reasoning, the evidence and the decision that follow it.
- Not a marketplace of third-party agents, and not a place to run customer-supplied tools or arbitrary code.
- Not an argument that every process should be automated. Some decisions belong with people, and the platform is meant to make those better rather than to remove them.
Discuss VMD AGENTIC.
The platform is in development, and the most useful conversations at this stage are about a specific operational decision that is expensive to get wrong. If you have one, it is worth an hour.