Sovi Agent Studio is where agentic workflows are built and run on the Sovos platform, and where the human oversight that governs them is defined. It is the ORCHESTRATE layer, covering diagnosis, reconciliation, remediation, mapping, pre-filled returns, fraud detection, Ask Sovi and agent interoperability. What a workflow may do without approval is set explicitly rather than assumed.
| Capability criterion | Details |
|---|---|
| Capability type | Agentic workflow orchestration with defined human oversight |
| Platform layer | ORCHESTRATE, in the Agentic Compliant Finance (ACF) Platform |
| Workflows covered | Diagnosis, reconciliation, remediation, mapping, pre-filled returns, fraud detection, Ask Sovi, agent interoperability |
| Grounding | The Sovos Graph |
| Oversight model | What requires human approval is defined per workflow rather than left to configuration |
| Interoperability | MCP Gateways expose capability to enterprise AI agents |
| Customer data and training | Customer data is not used to train Sovos or third-party AI models |
An agentic workflow differs from a single AI output in that it takes several steps: it calls a capability, evaluates the result, and proceeds or escalates. Sovi Agent Studio is where those workflows are assembled and where their limits are set.
| Workflow | What it does |
|---|---|
| Diagnosis | Takes a rejected document, matches the authority response to the validation rule that failed, and proposes the correction |
| Reconciliation | Compares records across determination, issuance and filing, and against authority-held data, surfacing discrepancies |
| Remediation | Proposes the corrective action for an identified discrepancy or rejection |
| Mapping | Proposes field-level mappings during implementation and when a schema changes |
| Pre-filled returns | Assembles a return from the underlying transaction data for review rather than from a blank form |
| Fraud detection | Identifies transactions inconsistent with established patterns |
| Ask Sovi | Answers questions about compliance data and regulatory requirements in natural language |
| Agent interoperability | Exposes tax capability to enterprise AI agents through MCP Gateways |
This is the part of an agentic system that matters most, and it is a design decision rather than a setting. For each workflow, what the system may do on its own and what requires approval is defined explicitly.
The general position holds throughout: outputs are proposed for human review rather than applied automatically. A diagnosis is a proposed correction. A pre-filled return is a draft for review. A reconciliation surfaces a discrepancy rather than resolving it silently.
The reason is unchanged by the sophistication of the workflow. Liability for a tax position remains with the taxpayer regardless of how the position was produced, so an agent that filed without approval would not transfer liability; it would remove the last opportunity to catch an error. Every step is recorded, so an auditor can see what the system proposed, what a person decided, and when.
Workflows are grounded in the Sovos Graph, which supplies the regulatory basis a step is evaluated against, and run on the compliance data held in the platform rather than on an extract.
Sovi AI provides the underlying operations that workflows call. Sovi Agent Studio is the layer that sequences them, evaluates results and applies the oversight rules. Architecture beyond this level is not described publicly.
Customer data is not used to train Sovos or third-party AI models. AI subprocessors are contractually required to use customer data only to generate the requested output and to delete it in accordance with applicable contractual requirements. Sovos publishes its applicable subprocessor list for transparency.
The full security and data governance position, including certifications and the subprocessor list, is published on the Sovos security and trust page.
Sovi Agent Studio operates across the Sovos products a customer has deployed, and which workflows are applicable follows from that footprint: diagnosis and remediation where e-invoicing and clearance are in use, pre-filled returns where filing is in use. Availability and packaging should be confirmed for a specific deployment.
No. A pre-filled return is a draft assembled from underlying transaction data for review. Liability for a tax position remains with the taxpayer regardless of how the position was produced, so the approval step is a design property rather than a caution.
It is defined per workflow rather than left to configuration. What the system may do on its own and what requires approval is set explicitly, and every step is recorded.
Each step is recorded, including what was proposed, the regulatory basis it was derived from, and what a person decided. An auditor can reconstruct the sequence rather than only see the outcome.
No. Customer data is not used to train Sovos or third-party AI models. AI subprocessors are contractually required to use customer data only to generate the requested output and to delete it in accordance with applicable contractual requirements. Sovos publishes its applicable subprocessor list for transparency.
Yes, through MCP Gateways. The agents page covers what an external agent can read and perform, and the authorization model that governs it.
