Why Zovos

AI that you can put in the record.

Most governance, risk, and compliance teams already use AI. They use it in a chat window beside the real work, where a 600-row control matrix or a 90-page policy has to be uploaded before every question and nothing the assistant writes is cited, governed or saved with the evidence. That makes today’s AI workflow painful, expensive and inaccurate. Zovos puts the agents inside the system of record, cites every answer to the rule paragraph, and keeps all of it under your review and approval.
Beside the work

AI beside the work loses the detail.

A general-purpose assistant knows only what you paste or upload into one conversation. This is what happens to a control matrix or a policy on the way in.

01

The file becomes tokens.

Microsoft researchers measured one 576-row, 23-column spreadsheet at 61,240 tokens in a plain encoding, and found that merged cells and multi-table sheets do not survive the conversion well.1 A control matrix with owners, frequencies and test steps, plus the policy and the rule, fills much of a context window before anyone asks a question. NIST SP 800-53 alone lists 1,189 controls and enhancements across 492 pages.9

02

Long inputs lose the middle.

In a 2023 study, a model answered 75.8% of questions correctly when the relevant document came first among twenty and 53.8% when it sat in the middle. That was worse than the 56.1% it scored with no documents at all.2 A 2025 study found that 11 of 13 long-context models fell below half of their short-input accuracy at 32,000 tokens when the question and the answer did not share the same words.3 A rule and the policy that implements it rarely do.

03

Big uploads are searched, not read.

When project files outgrow the context window, chat assistants switch to searching them and read only the passages they judge relevant.5 Long PDFs can be processed as text only, which flattens the test-of-controls tables in an audit or SOC report.6 Nothing in the answer shows which rows or pages were never read, so “no gap found” looks exactly like “the row was never retrieved.”

04

The answer goes nowhere.

The reply lives in a chat. It does not update the control, the issue log or the workpaper. The next conversation starts from zero, so the same matrix and the same policies are uploaded again. A 2025 study of 18 models found that performance consistently degrades as that input grows.4

61,240
tokens to encode one 576-row spreadsheet, before a single question is asked.1
11 of 13
long-context models fell below half of their short-input accuracy at 32,000 tokens.3
17% to 33%
of answers from purpose-built legal research tools with retrieval were hallucinated.12
$20
per million output tokens at list price for a current top-tier model, plus $4 per million tokens of input it reads.7
01 · The problem

Today’s AI workflow is painful.

Every conversation starts by uploading the control matrix and the policies again, and every answer has to be copied back into the record by hand.

Copy, paste, repeat.

Staff paste rule text and policy excerpts into a general chatbot, then copy the answer back into a spreadsheet, a Word file or a ticket. Nothing the assistant wrote is tied to the control, finding or evidence it was about.

Context starts from zero.

A general assistant does not know your charter, your regulator, your risk appetite or last cycle’s findings. Every new conversation has to rebuild that context and upload the same files again.

Exam material cannot go there.

Federal banking rules bar disclosing confidential supervisory information, such as a report of examination, without the agency’s permission.10 The material that matters most in an exam stays out of a consumer chat tool, so the hardest work gets no AI help.

02 · The problem

Today’s AI workflow is expensive.

Large uploads fill large context windows, and every token is metered again on each follow-up question, before anyone has checked the answer.

Big files, big windows.

A control matrix, a few policies and a vendor SOC report can run to hundreds of thousands of tokens, and every one is metered. A current top-tier model lists at $4 per million input tokens and $20 per million output tokens, so long answers drafted from long uploads are the most expensive work a model does.7

Every follow-up pays again.

Model APIs are stateless, so each new question sends the whole conversation back, uploads included, and the model reads it again.8 Caching can discount the repeats but does not make them free. Ten follow-ups on one large upload means paying to read it ten times, whether that lands on an invoice or as a usage limit your team hits mid-task.

Rework is the hidden bill.

An answer with no source trail has to be checked against the rule and re-keyed into the system of record. The hours saved on the first draft are spent again on verification, on top of the AI add-on you already pay for.

03 · The problem

Today’s AI workflow is inaccurate.

Long uploads lose the middle, and nothing in the answer shows which rows or pages were never read.

Fluent is not the same as correct.

Stanford researchers found that 2023 chat models hallucinated on 58% to 88% of verifiable questions about court cases.11 Purpose-built legal research tools with retrieval still did so 17% to 33% of the time.12 Regulations carry the same citations, amendments and cross-references.

Trained on a snapshot.

On April 17, 2026, SR 26-2 replaced SR 11-7, and the OCC’s companion bulletin placed generative and agentic AI outside its scope.13,14 A model trained before that date can still describe SR 11-7 as current guidance.

Review happens twice.

Without a source trail, a second person re-reads the rule to verify the output. The AI adds a step instead of removing one, and a chat transcript is not evidence an examiner can rely on.

Benchmarks

No benchmark measures this work.

Public finance benchmarks measure analyst work such as reading filings. A 2026 review found that public leaderboards do not capture the demands of financial-services work.19

What public benchmarks test

FinanceBench
Questions about public company filings. A 2023 model with retrieval answered 81% of them incorrectly or refused them.15
FinBen
Thirty-six datasets across forecasting, credit scoring, fraud and other analyst tasks.16
Finance Agent
Five hundred thirty-seven entry-level financial analyst tasks. The best published score was 64.37% in June 2026.17
RIRAG
Question answering over one regulator’s rulebook, scored on whether answers miss or contradict obligations.18

What we found no public benchmark for

  • Testing a control against its evidence.
  • Scoring a risk and control self-assessment.
  • Mapping a rule change across a control library.
  • Drafting a response to an exam request letter.
  • Reviewing a SOC report for exceptions and user entity controls.
  • Validating a finding and its remediation before closure.

With no outside measure of accuracy, the control is the review. Every Zovos agent output stays a proposal that a named person accepts, with the exact inputs it was shown pinned to the run.

Real work

Where the chat way breaks.

Five tasks a compliance, risk or audit team runs every cycle, done the chat way and the Zovos way.

Compliance

A rule change against the control matrix

Today
You paste the amended paragraph, upload a 600-row control matrix and ask which controls it touches.
Where it breaks
The matrix alone runs to tens of thousands of tokens, the affected rows sit mid-file, and the rule and the control rarely share wording.1,2,3 A short list of affected controls proves nothing about the rows left off it, so the team re-checks every row.
In Zovos
The Regulatory Monitor flags the changed paragraph against the indexed version. Gap Analysis and Policy Mapping propose the gaps and policy changes against your own library, each with its citation, for a person to accept. Regulatory change
BSA/AML

A policy redline after an amendment

Today
You upload a 90-page BSA/AML policy and the amended rule and ask for a redline.
Where it breaks
Obligations get missed or contradicted, which is exactly what regulatory question-answering research scores.18 The redline lives in a chat, outside the policy’s approval history.
In Zovos
When a paragraph moves, the Drafter and Policy Mapping agents propose the redline inside the policy, linked back to the change. A person approves it. Policies
Third-party risk

A vendor SOC 2 review

Today
You upload the vendor’s SOC 2 Type II report and ask for the exceptions and the complementary user entity controls.
Where it breaks
Long PDFs can be processed as text only, which flattens the test-of-controls tables.6 The user entity controls have to be matched to your own controls, which the chat cannot see.
In Zovos
The SOC Reviewer agent summarizes the report and its exceptions on the vendor record for human review, and the platform alerts you before the report expires. Third-party risk
Exam management

The first-day exam letter

Today
You paste the request letter and last cycle’s findings and ask for a response plan.
Where it breaks
Exam correspondence and findings can be confidential supervisory information that may not be shared without the agency’s permission.10 The chat cannot see where the evidence lives either.
In Zovos
The first-day letter starts from request-list templates tied to the frameworks under exam, and examiners work in a scoped, time-boxed room. Exam management
Internal audit

An audit workpaper

Today
You paste the test results and ask for the workpaper write-up.
Where it breaks
No public benchmark we found measures how well a model tests evidence against control attributes. The write-up is not tied to the sample, the evidence or the sign-off, so the auditor re-performs the work to document it.
In Zovos
The IA Workpaper Drafter writes from the evidence linked to the engagement. A claim that names a source it was not given is dropped before a person sees the draft. Internal audit
What Zovos does instead

Agents in the system, people in charge.

Records, not uploads
Your control matrix, risk register, vendor list and policies import from CSV or Excel into real records, with a dry-run report on every row before anything is written. Agents work on those records, alongside 15 integrations in 6 categories.
Integrations
The right paragraph, not the whole file
Agents retrieve the numbered paragraphs that match the task and drop weak matches rather than padding the answer. Every citation is versioned, across 2,339 citations in 57 frameworks.
Regulatory change
Proof of what it read
Every agent run records each input it was shown with a SHA-256 fingerprint. Zovos has 15 agent task types, and every output stays a proposal until a person accepts it.
AI agents
Governed data
Models run on AWS Bedrock, which does not store prompts or completions or use them for training, and Zovos does not train on customer data. Each tenant has its own schema and encryption key, and the audit trail is append-only.
Security
How it compares

One system, not a patchwork.

Zovos compared with spreadsheets and a legacy GRC suite
CapabilityZovosSpreadsheetsLegacy GRC suite
Regulatory monitoringAn agent reads Federal Register and eCFR primary sourcesManual reading and re-keyingContent feed, often an add-on module
Citation versioningEvery citation is versioned, and rule text is kept per amendment dateManual copy and pasteVaries by content module
Evidence-gated closeNo finding closes without remediation evidenceNot enforcedConfigurable workflow
Examiner accessScoped, time-boxed grants that are read-only apart from a closed set of fieldwork actionsEmailed copiesRole configuration per deployment
Internal-audit independence wallAuditees see only their own requests, and auditors cannot approve what they may auditFolder permissionsSeparate audit module
Agent human-gating15 agent task types, and output stays a proposal until a person accepts itNot applicableAdd-on module, where offered
Connectors15 integrations in 6 categoriesManual import and exportIntegration projects

Zovos cells reflect shipped product behaviour. Competitor columns describe tool categories rather than specific vendors.

Sources

Where the numbers come from.

Every outside figure on this page links here. We read each source on October 23, 2026. The studies tested the models available when they were run.

  1. Microsoft Research, “SpreadsheetLLM: Encoding Spreadsheets for Large Language Models,” 2024. arxiv.org
  2. Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” Transactions of the ACL, 2023. arxiv.org
  3. Modarressi et al., “NoLiMa: Long-Context Evaluation Beyond Literal Matching,” ICML 2025. arxiv.org
  4. Chroma, “Context Rot: How Increasing Input Tokens Impacts LLM Performance,” July 14, 2025. www.trychroma.com
  5. Anthropic Help Center, “Retrieval augmented generation (RAG) for Projects,” 2026. support.claude.com
  6. Anthropic Help Center, “Uploading files to Claude,” 2026. support.claude.com
  7. Anthropic, “Pricing,” Claude API documentation, list prices for Claude Opus 5.5, read October 23, 2026. platform.claude.com
  8. Anthropic, “Using the Messages API,” Claude API documentation: the API is stateless and each request sends the full conversation history. platform.claude.com
  9. NIST SP 800-53 Revision 5 and its control catalog spreadsheet. We counted 492 pages and 1,189 controls and enhancements, withdrawn items included. csrc.nist.gov
  10. Confidential supervisory information rules: 12 CFR 261.20 (Federal Reserve), 12 CFR 4.37 (OCC), 12 CFR 309.6 (FDIC) and 12 CFR 792.30 (NCUA). www.law.cornell.edu
  11. Dahl, Magesh, Suzgun and Ho, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” Journal of Legal Analysis, 2024. The study tested 2023 models. arxiv.org
  12. Magesh et al., “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” Stanford RegLab and HAI, 2024. arxiv.org
  13. Board of Governors of the Federal Reserve System, SR 26-2, “Revised Guidance on Model Risk Management,” April 17, 2026. www.federalreserve.gov
  14. Office of the Comptroller of the Currency, Bulletin 2026-13, April 17, 2026. www.occ.gov
  15. Islam et al., “FinanceBench: A New Benchmark for Financial Question Answering,” Patronus AI, 2023. arxiv.org
  16. Xie et al., “FinBen: A Holistic Financial Benchmark for Large Language Models,” 2024. arxiv.org
  17. Vals AI, Finance Agent benchmark, page updated June 4, 2026. www.vals.ai
  18. Gokhan et al., “RIRAG: Regulatory Information Retrieval and Answer Generation,” 2024. arxiv.org
  19. Hudson, “Meta-Benchmarks for Financial-Services LLM Evaluation,” 2026. arxiv.org

Bring AI into the record. Keep people in charge.

Bring your controls, a policy, and one regulator paragraph that gives you trouble. We target two-week onboarding.

Join the waitlist →See the platform