Smarsh AI Constitution

A letter to our customers

As AI takes on more responsibility in compliance, you deserve clarity about how it works, what you control, and who is accountable.

This abridged version of Smarsh’s AI constitution highlights 12 principles across four areas:

  • The boundaries we uphold. Protecting compliance and market integrity.
  • How we build. Standards for fairness, explainability, and transparency.
  • What you control. Options to configure, validate, and override AI.
  • How we hold ourselves accountable. Responsibilities to regulators, customers, and the market.

These principles guide how we develop and deploy AI, with governance built in and autonomy set by you. As capabilities and regulations evolve, we update the constitution and explain why.

Clients can request the complete constitution from their account representative; prospective clients can contact our sales team.

Principle 1: Protecting compliance and market integrity

We develop AI Products designed to support your compliance program and the markets you operate within.

Some risks are too serious to justify in exchange for a potential benefit. Our responsibility to the integrity of the compliance system guides how our AI surfaces, scores, and reports risk, including when commercial pressure or a firm’s immediate interests point elsewhere. These standards extend to autonomous agents that close alerts or generate compliance outputs.

For example:

  • We do not unilaterally suppress, alter, or delay a genuine risk finding. Threshold configurations are customer decisions, and Smarsh designs these controls to be transparent, auditable, and applied with customer authorization.

Principle 2: Your data stays yours

Your firm and your clients own your communications data, and any use for AI training depends on your explicit consent.

Smarsh is a data custodian and uses your data only to deliver services.

These standards cover foundation models and other AI we develop, including proprietary compliance intelligence models, and third-party AI embedded in our products.

For example:

  • Model training: When we develop or update any AI model including domain-adapted compliance intelligence models, we train on synthetic or consented data, never on your archived communications.

Principle 3: Security and privacy built in

We design AI to protect your data and agentic systems to resist threats.

Smarsh AI development begins with a security and privacy review. Models operate in isolated environments, following data minimization principles that limit access to what each task requires.

As AI carries out multistep workflows, connects to external systems, and acts on your behalf, its attack surface expands. We design safeguards against these risks and excessive agency, supported by industry-standard security certifications and independent third-party audits.

For example:

  • Agents cannot call external systems a customer has not specifically approved.

Principle 4: Fairness in decisions

Our AI is designed and tested to minimize disparate impact across employee demographic groups.

At Smarsh, fairness is a baseline standard. We test models for bias before deployment and monitor fairness post deployment. When we detect bias, we address it, with processes designed to preserve, not degrade, any model fairness properties.

For example:

  • Before Smarsh deploys a new supervision model, we test the model across demographic groups for disproportionate flagging of employees from different backgrounds.

Principle 5: Explainable and defensible

We design Smarsh AI to be understandable.

Explainability helps you defend AI-assisted outcomes through visibility into model training, threshold settings, performance characteristics, and the indicators behind the outcome.

For example:

  • A complete audit trail for each AI-assisted decision shows how it was made and who reviewed it, supporting regulatory examinations.

Principle 6: Clear about confidence and limitations

We communicate AI confidence, limitations, and failures so you can apply informed judgment.

At Smarsh, confidence indicators reflect model certainty rather than a target alert volume, helping reviewers weigh findings proportionately. A 60% confidence finding warrants different consideration than a 95% finding.

We disclose known limitations, edge cases, and areas of lower performance, and explain what changes as models evolve.

For example:

  • AI outputs, including risk alerts, include confidence indicators, with uncertain communication patterns flagged as lower confidence rather than assigned a default risk score.

Principle 7: Governed autonomy and control

You choose how autonomous Smarsh AI operates in your environment, with visibility into its access, actions, and changes.

Smarsh AI is designed to support three levels of autonomy: assistive AI surfaces findings for human review, augmentative AI recommends actions and drafts outputs with human in the loop, and autonomous AI acts on behalf of the human where applicable. Your compliance team configures each workflow and can change autonomy levels as your confidence and regulatory context evolve.

For example:

  • Complete action logs cover each autonomy level, with human and agent actions recorded and available for regulator review.

Principle 8: Model risk management ready

We design AI to support your model risk team’s validation, documentation, and review requirements.

Financial institutions have formal obligations to validate, document, and challenge models, including those supplied by vendors. At Smarsh, we support that work through documentation, independent validation, and direct assistance.

For example:

  • We provide comprehensive documentation for each AI model in production.

Principle 9: One standard for AI we build or incorporate into our Products

We apply the same governance standards to AI we develop and third-party components in our products.

At Smarsh, AI components are evaluated before deployment for data protection, explainability, fairness, security, and auditability, regardless of origin. Third-party providers undergo vetting, contractual requirements, and ongoing review by our AI Governance Committee.

For example:

  • Provider contracts prohibit retaining or using your data for their own training, under the same terms applied to Smarsh-developed models

Principle 10: Named accountability and clear escalation

Named people at Smarsh are responsible for AI Governance, with a clear path to raise concerns.

Our AI Governance Committee brings together product, engineering, legal, security, and compliance leadership to review deployments against this constitution. Customers, regulators, and employees have a defined channel to raise concerns.

For example:

  • Our AI Governance Committee reviews and approves AI deployments and product changes before release.

Principle 11: Built for regulatory reality

At Smarsh, we design our AI to address regulatory requirements and share updates to help you respond to change. SEC Rule 17a-4, FINRA Rule 4511, MiFID II, FCA requirements, CFTC standards, and EU AI Act requirements for AI systems inform product design from the outset.

Our dedicated regulatory intelligence function monitors guidance, enforcement actions, and emerging standards across these frameworks.

For example:

  • We track AI-specific developments, including the EU AI Act, SEC AI guidance, and equivalent frameworks.

Principle 12: Sustainable compute

AI is energy intensive. Smarsh is conscious of AI consumption and the environmental impact of AI workloads across our platform. We seek to optimize model efficiency, preferring right-sized models where quality is equivalent, and we prefer third-party AI providers with credible energy commitments for AI workloads.

For example:

  • When two AI approaches deliver equivalent quality, we use the more efficient one.

Keeping this constitution current

We review this constitution periodically and update it when material changes in our AI practices or regulations warrant it.

Questions, concerns, and formal submissions can be directed to your account executive, with responses provided within seven business days.

This document describes Smarsh’s current approach to AI governance and provides neither legal nor regulatory advice. It creates or modifies no contractual obligations and does not supersede customer agreements.