How it works
Between the agent's intent and the real action
Gatewise is an action control layer for AI agents. It evaluates proposed business actions against deterministic policies and trusted evidence before execution. Here is the full path a single action takes.
Action request
refund.create
Refund 349 EUR to Acme Corp
waiting for decision
Evidence and policy check
- Invoice statuspaid
- Customer verifiedyes
- Existing refundnone
- Automatic refund limit100 EUR
- Manager approval threshold100 - 500 EUR
Result
evaluating...
The flow
Three steps, one deterministic result
- 01
Intercept the action
The agent does not call the business tool directly. It sends the proposed action to Gatewise first: the action type, its parameters and the context it has.
- Action type, for example refund.create or crm.owner.change
- Parameters such as amount, record id, recipient
- Agent identity and the user or workflow it acts for
- 02
Check the evidence
Gatewise reads the system data the rule depends on, then evaluates the policy for that action type. Nothing here is generated by a model.
- Trusted business data such as invoice status or contract dates
- Deterministic rules and monetary or percentage thresholds
- Duplicate and existing task checks
- Approval requirements and the responsible role
- 03
Return a decision
Gatewise returns one outcome with a reason and the evidence it used. The agent continues only when the policy permits it.
- ALLOW: execute now
- BLOCK: stop, with the rule that blocked it
- REQUIRE APPROVAL: hold until a named approver decides
- NO ACTION: nothing needs to happen
- INSUFFICIENT EVIDENCE: required data is missing or stale
Outcomes
Five decisions, no in between
A small, fixed set of outcomes keeps agent behaviour predictable and keeps integration simple.
- ALLOW
Evidence and policy conditions are met. The agent may execute the action.
- BLOCK
A policy rule forbids this action. The agent cannot execute it.
- REQUIRE APPROVAL
The action is valid but above an autonomous threshold. A named human has to approve it.
- NO ACTION
The action is unnecessary. A duplicate or an existing task already covers it.
- INSUFFICIENT EVIDENCE
Required system data is missing or stale, so the decision cannot be made safely.
Rollout
Shadow mode first, enforcement second
Run Gatewise without blocking anything. It evaluates agent actions against your policies and shows what would have been allowed, blocked or escalated. Enforcement is then enabled per action type.
- actions evaluated
- 1,284
- would have been blocked
- 42
- required approval
- 11
- duplicate actions detected
- 7
Interface example. These numbers are not Gatewise customer results.
Architecture
Where Gatewise sits
Designed to sit between agent runtimes and the tools they use. Framework agnostic by design.
- proposes
Agent / copilot
Any agent runtime or framework. Gatewise does not care which one.
- intercepts
Gatewise API / SDK / gateway
Designed to sit between agent runtimes and the tools they use.
- decides
Policy and evidence engine
Deterministic rules, thresholds, duplicate checks and required evidence.
- executes
Business systems
CRM, support, payments, billing and internal tools.
Result
An operational record of what agents tried to do
Each decision is stored with its reason, the policy that produced it and the evidence it used.
- 09:41Refund 79 EURsupport-agent-v3ALLOW
- 09:43Discount 25 percentReason: Exceeds 10 percent autonomous discount limitsales-copilotBLOCK
- 09:44Renewal follow-upReason: Existing task already assignedcs-agentNO ACTION
- 09:47Refund 640 EURReason: Finance approval requiredsupport-agent-v3REQUIRE APPROVAL
- 09:51Delete customer recordReason: Deletion restricted to humansops-agentBLOCK
Interface example. Not Gatewise customer data.
Give your agents freedom without giving them a blank cheque.
Tell us what your agents are allowed to do today. We will review whether Gatewise fits your agent workflow.
Works with your agents and existing business tools