Platform
The AI proposes. Your policy decides.
Thirty-plus agents are embedded across the application, doing the day-to-day work beside you. Every one of them runs under the same policy, inside your permissions, and logs every token it spends — because in a system whose output has to survive an audit, what matters is not what a model can do but what it is allowed to do, and whether you can prove what it did.
The financial problem
Automation you cannot explain is automation you cannot deploy.
- An accrual adjusted by a model, with no record of why, is an audit finding waiting to happen.
- Ungoverned assistants leak data across roles — the analyst sees what only finance should.
- Time saved is worthless if every output has to be re-checked manually anyway.
- The value is in the volume work: matching, mapping and anomaly detection at a scale humans cannot cover.
The technical problem
Scope, policy and trace are the three controls that matter.
- Scope: the agent sees exactly what the signed-in user may already see — no service account with god rights.
- Policy: what agents may do is configured at product, company and role level, not hard-coded.
- Guardrails: requests and responses are both checked; anything flagged routes to a managed review queue.
- Trace: every interaction and every action is logged immutably and is explainable after the fact.
How it is governed
The RevUpra Agent, and every specialist under it.
Agents are not a collection of separate tools that each need their own controls. One agent stands in front of all of them, and every other agent answers to it — which is what makes a single policy enforceable across the whole estate.
The RevUpra Agent
One front door
Every request enters through the RevUpra Agent. It decides which specialist should handle the work, passes the caller's identity along with it, and owns the answer that comes back. Nothing reaches a model without going through it first.
Always-on guardrails
Checked in both directions
A request filter and a response filter run on every single interaction — the inbound prompt before a model sees it, the outbound answer before a person does. They are agents too, which means they are configured, versioned and audited like the rest.
Skill agents
Narrow on purpose
Contract extraction, field mapping, deal efficiency, cash matching, billing reconciliation, market scout, app builder and the rest. Each is scoped to one job, so its permissions can be scoped to one job too — a general-purpose agent has to be trusted with everything.
Policy
Three levels, merged key by key
Policy is configuration, not code. Each level narrows the one above it, and they merge field by field — so a role can tighten a single threshold without restating a policy it otherwise agrees with.
-
01
Product
The seeded baseline every company starts from — what agents may read, what they may propose, and what they may never do. Shipped by us, visible to you.
-
02
Company
Your controls environment. Turn a capability off entirely, lower a threshold, or require review where we would have allowed a direct action.
-
03
Role
The narrowest level. A role-scoped policy applies only to users holding that role, so an analyst and a finance approver can face different limits on the same agent.
An agent with no applicable policy does not get a default of "allowed". It gets the product baseline, which is the most restrictive of the three.
Access
Scoped by role. Restricted by attribute.
An agent is not a service account. It runs as the person who asked.
RBAC — what the role may do
The agent inherits the signed-in user's entitlements exactly. If a person cannot open the settlement screen, no amount of asking an agent nicely will produce settlement data. There is no elevated identity behind the assistant, because an identity with god rights is the single thing that makes every other control cosmetic.
ABAC — which rows they may see
Role decides the capability; attributes decide the data. Region, business unit, vendor, customer and programme scope all narrow what the same agent returns to two people holding the same role — so a regional analyst's assistant answers about their region and nobody else's.
Audit
Every token is accounted for
Not a sample. Not an aggregate. Every interaction, with what it cost and what it touched.
Token spend, per call
Tokens in and tokens out are recorded against the agent, the user and the page it was asked from — so spend is attributable, not a monthly surprise.
Flagged, not buried
Anything a guardrail catches is raised to system administrators with the prompt, the response and the rule that fired.
Reviewable after the fact
An administrator can audit any interaction — what was asked, what was answered, which agent, under which policy version.
Immutable trace
The record is written as it happens and is not editable afterwards, which is the only form of evidence an auditor will accept.
Bring your own agent
Connect any agent. Including the one you already run.
The provider is configuration. Point the RevUpra Agent at our default, at your own enterprise agent behind your own contract and data-residency terms, or at whichever model your security team has already approved. The governance layer does not move: whatever is connected still enters through the RevUpra Agent, still passes both checks, still runs inside the caller's permissions, and still logs every token.
That is the point of governing the seam rather than the model — you can change your mind about the model without renegotiating your controls.
What it does
Agents doing the volume work
The tasks worth automating are the ones that are too numerous for people and too rule-bound to be interesting.
In-app assistant
Answers questions about your own agreements, accruals and claims, scoped to your permissions.
Contract capture
Reads an executed agreement and proposes the structured terms for a human to confirm.
Field mapping
Proposes integration mappings for a new partner feed from the shape of the data.
Anomaly detection
Flags accrual, claim and price movements that do not fit the pattern, with the evidence attached.
Deal analysis
Summarises what a proposed structure would have cost against your own historical volume.
Review queue
Everything flagged, in one place, with the proposal, the reasoning and the accept or reject decision recorded.
The line we draw
There is a real distinction between AI that answers and AI that acts. An assistant scoped to what you may already see is low-risk and immediately useful. An agent that adjusts an accrual is a different thing entirely, and needs an explicit policy, a review gate and an immutable record of what it did and on what basis.
We build both, and we keep the boundary between them configurable — because where you want it drawn depends on your controls environment, not on ours.
Ask the assistant on this site.
If it is switched on, the chat button in the corner is the same governed-assistant pattern, pointed at public content instead of your ledger. Ask it something hard.