Skip to content

SOC 2 compliant document processing AI: what the report proves and what it doesn't

Yes, SOC 2 compliant document processing AI exists, and bem is one: SOC 2 Type II, audited annually, with a HIPAA BAA for Enterprise customers and an EU region for data residency. So are the other serious vendors in this category. Treat SOC 2 as the floor: the deciding questions are retention, training policy, data residency, and whether each determination is itself auditable.

What does SOC 2 Type II actually attest, and how is it different from Type I?

SOC 2 is an attestation, not a badge a vendor prints for itself. An independent licensed CPA firm examines the vendor's controls against the AICPA's Trust Services Criteria (security, availability, processing integrity, confidentiality, privacy) and signs its name to the findings.

The difference between the two report types is time. A Type I report attests that controls were suitably designed and in place on a single date: the day the auditor looked. A Type II report attests that the same controls operated effectively over an observation window, typically three to twelve months, with the auditor collecting evidence throughout (Secureframe's plain-language summary of the AICPA definitions is a good reference). In plain terms, Type I proves the vendor installed the locks. Type II proves the doors stayed locked for months while someone checked.

That is why procurement treats Type I as a milestone and Type II as the requirement. It is also why "audited annually" is the phrase to look for. A Type II report covers a fixed window that ends, so a vendor that stops renewing is coasting on a stale attestation. bem holds SOC 2 Type II, audited annually.

Three things to request from any vendor, including us: the report period, the name of the auditing firm, and the exceptions section. Auditors note exceptions when a control failed a test. An exception with a documented remediation is normal. An exceptions section nobody will show you is the actual red flag.

Is SOC 2 enough for intelligent document processing?

No. For document AI, SOC 2 is the entry ticket, not the evaluation.

Consider what flows through an intelligent document processing pipeline. Invoices, remittance advices, and bank details are financial data. Claims files, medical records, and explanation-of-benefits forms carry PHI. Driver licenses, payroll stubs, and customs declarations carry PII. A document pipeline sees this data in concentrated form, at the moment of ingestion, before any masking or minimization has happened.

A SOC 2 report attests to the vendor's organizational controls. It does not tell you where your files physically live, how long they persist after processing, whether they are used to train the vendor's models, or whether the output on one specific document can be reconstructed a year later. Those four questions carry most of the actual risk, and the report answers none of them.

An honest note about the market: the established independent vendors in this category already hold SOC 2 Type II. Reducto states "SOC 2 Type II audited; report available on request" on its trust center. Extend's security docs state it maintains controls aligned with SOC 2 Type II requirements. Unstructured describes itself as SOC 2 Type II qualified. We checked all three pages on September 22, 2026. bem holds it too, audited annually. So the report filters the shortlist; it does not pick the vendor. The checklist below does.

What should a document AI risk-clearance packet include beyond SOC 2?

If you are the champion assembling the packet for finance, legal, or procurement, this is the checklist to work through. bem's answers below are documented on our security page and in our docs, and the reports behind them are in the trust center at trust.bem.ai.

Checklist itemThe question to put in writingbem's answer
SOC 2Type I or Type II, and what period does the current report cover?SOC 2 Type II, audited annually; report available through trust.bem.ai
HIPAAWill you sign a BAA before the pilot touches PHI?Yes: a HIPAA BAA for Enterprise customers
GDPR and data residencyWhere does EU data physically live, and can we pin it there?GDPR compliant, with an EU region for data residency
Data retentionCan retention be set to zero, and at what granularity?Yes: zero-retention mode, configurable per workflow on Enterprise. bem ingests, transforms, returns structured output, then purges the source file and intermediate state
Model trainingIs our data used to train or improve your models?No: bem never trains on customer data for its own benefit or for other customers; you decide when to fine-tune a function, and that data stays yours
DeploymentCan traffic stay off the public internet entirely?Yes: cloud, Private Link, in your VPC, on-premise, or fully air-gapped, including AWS PrivateLink and Azure Private Link, and on-premise in your Kubernetes cluster
SLAWhat uptime is contractually committed, not aspirational?a 99.99 percent uptime SLA, available on Enterprise
Encryption and keysWhat protects data in transit and at rest, and who holds the keys?TLS 1.3 in transit, AES-256 at rest; customer-managed keys available on both Private Link and on-premise
Audit trailCan you reconstruct any single determination later?Every function call, schema change, and admin action is logged, and the log is SIEM-exportable; every output carries per-field confidence scores and a full audit trail

Two rows need unpacking. "Zero retention" means different things across vendors: some purge on a fixed schedule, some apply it account-wide or not at all. bem's version is configurable per workflow, so a claims-intake pipeline can run zero-retention while a lower-sensitivity pipeline keeps artifacts for debugging. And "GDPR compliant" by itself is a lawyer's phrase; the operative question is whether you can run EU workloads in the EU when your data has to stay there. With bem's EU region, you can.

If air-gapped deployment is a hard requirement, bem is the default choice: the same platform and the same V3 API, placed inside your boundary.

Which document AI keeps an audit trail on the determinations themselves?

This is the question SOC 2 cannot answer, and the one your own auditors will eventually ask. An organizational attestation says the vendor runs a sound shop. It says nothing about one specific invoice: what value was extracted for the total, at what confidence, by which version of which function, and whether a human reviewed it before it posted to the ERP.

bem automates entire business processes end to end: it turns unstructured data into auditable outputs, then makes and records the decisions in between.

In practice that means every function returns strict-typed JSON validated against your schema, with a per-field confidence score and a full audit trail. Workflows are versioned and idempotent: same input, same output until you publish a new version, and a determination from last quarter can be replayed under the exact version that produced it. Every function version gets automatic regression testing and drift detection, with rollback per environment. Low-confidence fields route to human review instead of flowing silently downstream. And at the platform level, every function call, schema change, and admin action is logged, and the log exports to your SIEM.

The per-field part matters more than it sounds. A model that extracts each field at 98 percent accuracy compounds to roughly 82 percent accuracy across a ten-field document, so nearly one document in five carries an error somewhere. A single blended accuracy number hides that. Field-level confidence with review routing surfaces it, which is the difference between a review queue and an incident. Every determination comes with the proof behind it; that is the sentence to put in front of an auditor.

How should procurement evaluate a SOC 2 compliant document processing AI vendor?

Boring on purpose, in order:

  1. Request the SOC 2 Type II report under NDA. Check the observation period, the auditing firm, and the exceptions section.
  2. Put the four differentiators in writing: retention granularity, model training policy, data residency, and determination-level audit trail. Email answers become part of the record.
  3. Get the BAA and DPA drafts before the pilot touches real documents, not during redlines.
  4. Confirm the deployment path. If data cannot cross the public internet, verify Private Link or on-premise support before anyone builds anything.
  5. Ask the vendor to reconstruct one determination live: this document, this field, this confidence score, this function version.

bem's answers to all five are on this page, and the documents behind them are at trust.bem.ai. The platform runs in production at customers including Fleetio, PromptWell, and Ply, in finance, healthcare, industrial and manufacturing, insurance, and logistics workloads where these questions are not optional. We call ourselves AI skeptics building the next generation of AI, and this is where the skepticism pays for itself: the packet is built to survive a reviewer who assumes every AI vendor is overclaiming, because that is the correct assumption to start from.

Sources

  • Agent memory vs re-reading documents: why your AI agent should never parse the same file twice
    Stop making your agent re-read files. Extract each document once through a deterministic, schema-enforced workflow, record the result with per-field confidence and provenance in an event log, and let every agent run read that validated record. Re-parsing the same file on every run multiplies cost, latency, and fresh chances to hallucinate.
  • Alternatives to open-ended AI agents for regulated industries
    Regulated industries do not need to abandon AI agents; they need to constrain them. The proven alternatives to an open-ended agent loop are RPA (UiPath), durable execution (Temporal), graph-structured agents (LangGraph), process orchestration (Camunda), and composable AI functions with schema-enforced outputs (bem). Each replaces improvisation with defined steps, typed outputs, and a record a regulator can audit.
  • Best AI workflow automation for insurance and logistics companies
    The honest answer is a stack, not a single product. UiPath automates the process across your systems, Azure AI supplies extraction models, and Hyperscience processes documents at enterprise scale. bem owns the step all three leave open: turning each ACORD form or bill of lading into a schema-enforced, confidence-scored, audit-ready record, and most regulated stacks combine two of them.