PivotSolvo

Governance & security


Governance as architecture, not as policy.

A policy document says what people should do. Architecture decides what the software will do when the deadline is tomorrow and everybody is tired. We put the rules in the second place.

What constrains the design


Two frameworks, not one.

Circular 230 governs what a practitioner may do in federal tax practice, including the diligence and competence obligations that follow you when you use a tool you did not build. COSO's guidance on internal control over generative AI governs whether the control environment around that tool is any good.

PivotSolvo is designed against both. Circular 230 alone would let you build something with no control environment at all; COSO alone would let you build a beautifully controlled system that a practitioner still cannot use. We work from the IRS Office of Professional Responsibility's introductory guidelines on responsible AI use in federal tax practice, alongside the AICPA Code and SSARS, which already cover AI-assisted work without needing an AI-specific standard.

The PivotSolvo AI Governance & Ethical Use Charter is a written document that applies to every firm on the platform. Nine sections — approved and prohibited uses, protecting sensitive data, verification, human sign-off on high-risk work, no loopholes, disclosure of material AI involvement, bias and fairness, slowing down under pressure, and escalation — plus accountability, training, a COSO mapping and a Circular 230 mapping table. It is reviewed whenever a module is added, and at minimum annually.

Available to founding firms on request.

The hard rules


Things the platform will not do, at any subscription level.

These are not settings. There is no toggle, no override, and no "just this once because it's the 15th".

No automatic filing or transmission

Nothing is filed or transmitted to a government authority automatically. Ever.

No auto-approval above materiality

Nothing above the firm's own materiality threshold is approved without a person, and the threshold belongs to the firm.

AI never moves client funds

No payment, transfer or disbursement is initiated by an agent under any circumstances.

No unreviewed client communication

Substantive communication reaches a client only after a person has read it. Low-stake automated reminders — a missing bank statement — are the deliberate exception.

No batch approval of write-backs

Write-backs, filings and notice responses are approved one at a time. There is no "approve all" button, by design, and adding one has been explicitly ruled out.

No skipping under deadline pressure

The review step cannot be bypassed because the week is bad. The platform does not offer a faster path that removes a person.

Containing the agents


What happens when an agent misbehaves.

Assume one will. The question a firm should ask a vendor is not whether an agent can go wrong, but what the software does when it does.

  • Least privilege per agent. Each agent gets exactly the access its job needs and nothing more.
  • No autonomous irreversible actions. Every consequential action is a proposal that waits for a person.
  • Full action logging. Every read, proposal, approval and write, append-only.
  • Input validation on anything a client sends. Uploaded documents and inbound mail are treated as untrusted input, not as instructions.
  • Rate limits and anomaly thresholds. Unusual volume or behavior auto-pauses the agent rather than continuing at speed.
  • Tenant and agent execution isolation. Separation at the network level, not only in the query.
  • A manual kill switch. One agent or the whole platform, from a button in the app.
  • Adversarial testing. Agents are tested against deliberately hostile inputs in an isolated environment.
An append-only audit trail entry showing snapshot, proposal, approval, write and verification
Every entry is defensible because the whole chain is on one screen.

Security


Where we actually stand.

We would rather tell you this plainly than let you find out during due diligence.

What is true today

  • A separate database per subscribing firm, with row-level security inside it
  • Encryption in transit and at rest
  • Credentials held in a managed key vault, never exposed to an agent or an export
  • Authenticated endpoints only; no unauthenticated function endpoints
  • An append-only audit log of every action on the platform
  • AI processing runs inside our own managed cloud environment, with zero data retention on the model provider side
  • Subscribing firms interact through the application only — no database access, no source access
  • Restricted-practice agents will not run unless a CPA, EA or attorney is verified at your firm, checked against the licensing source and rechecked every two years
  • Connections to your other services are authorized by your firm, with credentials in a key vault set up for your firm alone
  • Retention schedules by record type, with disposal queued for a person to approve

What is not true yet

We do not have a SOC 2 Type II report. None exists today. Strictly, SOC 2 is an attestation report rather than a certification, and nobody is "SOC 2 certified" — but however it is phrased, we do not have one, and any vendor implying otherwise about a platform this young is worth a second look.

It is on the roadmap, and dated: a Type II examination, Security and Confidentiality scope, with the observation window planned to open at the end of 2026 so the report lands well ahead of the 2027 season. We skipped Type I deliberately — a point-in-time letter would not tell you anything useful. Processing Integrity is deferred to the second year.

Until that report exists, what we can offer is this page, the charter, and a conversation with the CPA who built the thing.

Your clients, your data


Disclosure, ownership and getting out.

We think material AI involvement should be disclosed

Not buried, and not left to the client to infer. PivotSolvo supports disclosure through the engagement letter, a cover letter with completed returns, and a footer on financial statements — the places a client actually reads. How a firm words it is the firm's call; that it should be said is ours.

The data is the firm's

A full export, whenever you want it, without a support ticket or a notice period. A platform that is hard to leave is a platform that has stopped trying to be worth staying on.

Client confidentiality cuts both ways

We hold ourselves to the same confidentiality duty our subscribing firms owe their clients. Your client data is not used to demonstrate the product to anyone else — when we need to show the platform working, we show a fabricated client with planted issues, which is exactly what we do in our own investor and partner demonstrations.

Where judgment stays

Professional judgment and review cannot be removed from this work, and we do not claim to have removed them. PivotSolvo's north star is AI as the operating system of a firm rather than a tool bolted to the side of it — but the Circular 230 caveat sits in our own charter alongside that sentence, in writing, on purpose.

If a control on this page ever gets in the way of a release, the control wins. That is the whole point of writing it down before we needed it.

Early access

Ask us the hard questions first.

Founding firms get the charter, the architecture and an honest answer about what is built and what is not. Write to us.