November 19, 2026

Agent Washing in Financial Services: How to Substantiate AI Agent Claims Before Examiners and Investors Test Them

Key Takeaways

  • Agent washing is marketing an assistant, a workflow automation or a chatbot as an autonomous AI agent.
  • Agentic claims are easier to test than earlier AI claims: either the system acted in a system of record under its own plan, or a person did, and the logs show which.
  • The SEC and DOJ cases against Presto Automation and Nate’s founder already treat false autonomy claims as securities law violations, even without the “agent washing” label.
  • In financial services, agentic claims get tested in investor disclosures, regulatory exams, vendor due diligence and customer marketing, often all at once.
  • Substantiate each claim across six dimensions (autonomy, tool and system access, decision scope, human checkpoints, evidenced outcomes and failure handling) with operational evidence, not a demo.
  • Start keeping execution traces, evaluation results, guardrail logs and override rates the day an agent goes live.

Calling a product an AI agent is a claim someone can check. A system either plans and takes actions in a system of record on its own, or it drafts something and waits for a person to act. In financial services, that check now comes from four directions: SEC examiners reviewing registrants’ AI representations, investors reading earnings call transcripts, procurement teams running vendor due diligence, and plaintiffs’ lawyers reading everything the first three produced.

Our earlier analysis, AI in Finance: The Rise and Risks of AI Washing, covered firms that overstated their use of AI. This article covers the next version of the same problem. It defines agent washing, shows how the existing enforcement record already applies to it, and gives investor relations, disclosure counsel, vendor risk teams and AI program owners in banking, insurance and asset management a test they can run on any agentic claim, their vendors’ or their own, before an examiner, auditor or plaintiff runs it for them.

What Is Agent Washing?

Agent washing is marketing an assistant, a workflow automation or a chatbot as an autonomous AI agent. Gartner used the term in in June 2025, describing it as the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities. In the same forecast, Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI offer the real thing.

The products being relabeled are often useful, but the problem is the claim. FINRA’s 2026 Annual Regulatory Oversight Report defines AI agents as “systems or programs that are capable of autonomously performing and completing tasks on behalf of a user.” A product that cannot complete a task without a person carrying out the final step does not meet that definition, whatever the pitch deck says.

What the product is What it does Who takes the action
Chatbot or assistant Answers questions, retrieves documents, drafts text A person, every time
Copilot Suggests the next step inside an application A person approves each step
RPA or workflow automation Runs a fixed, pre-scripted sequence The script, with no planning; it stops when inputs change
AI agent Takes a goal, plans the steps, chooses tools, writes to systems of record within delegated authority, and handles its own errors The agent, inside limits a person set

Agent washing happens when anything in the first three rows is sold, disclosed or reported as the fourth.

How Is Agent Washing Different From AI Washing?

Agentic claims are easier to test than AI claims, so the consequences arrive faster. “We use AI in our investment process” describes an input, and “AI” is loosely defined enough that a firm can argue about what counts. “Our agent resolves claims end to end” describes the behavior. Either the agent wrote the claim decision to the system of record under its own plan, or a person did. The logs show which.

Three features make agentic claims sharper than earlier AI claims:

  • They come with numbers. Agentic marketing and investor messaging lean on automation rates, straight-through processing rates, cases closed without human touch and headcount avoided. Each number can be recomputed from execution records, and examiners and plaintiffs can ask for those records.
  • They describe the operating model. A claim that agents now handle onboarding, reconciliations or first-notice-of-loss speaks to the cost base and staffing, which is what makes it material to investors.
  • They create supervisory obligations whether true or false. If the agent is real, the firm has to supervise it. Snell & Wilmer reads FINRA’s 2026 report as pulling AI agents into Rule 3110 (Supervision), Rule 3120 (Supervisory Control Systems) and books-and-records preservation. If the agent is not real, the claim is misleading. Either way, the firm needs operational evidence of what the system does.

AI washing mostly exposed what a firm said about its technology. Agent washing also exposes what the firm can prove about its controls.

The AI-Washing Enforcement Record Already Applies to Agent Claims

No regulator has yet brought a case labeled “agent washing’’. The two AI-washing cases that best fit financial services’ agent claims, Presto Automation and Nate, were both about claimed autonomy: tasks supposedly completed without human intervention, backed by rates that, according to the SEC, misstated how much of the work humans did.

Date Matter Agency What was claimed What the agency alleged or found Outcome
Sep 29, 2026 Cryptoaiml and TSAI Pro SEC AI trading signals with 98% accuracy; rentable AI trading bots with guaranteed profits No trading platform, no trades, no bots, per the SEC Pending, S.D.N.Y.
Aug 26, 2026 GenesisAI Corp. SEC Revenue, valuation and customer demand for an AI model marketplace Marketplace never commercially viable; claimed waitlist did not exist, per the SEC Settled; injunction; former CEO paid $50,000 disgorgement plus interest and a $50,000 penalty
Apr 9, 2025 Albert Saniger (Nate, Inc.) SEC and DOJ (S.D.N.Y.) App used AI to complete online purchases without human involvement Purchases completed manually by contract workers, largely in the Philippines; over $42 million raised Parallel civil complaint and criminal indictment for securities and wire fraud
Jan 14, 2025 Presto Automation SEC Presto Voice eliminated human order-taking at drive-thrus; high rates of orders completed without human intervention Most orders needed offshore human agents; the reported rate counted orders completed without restaurant staff Cease-and-desist order; no penalty, citing the company’s finances
Jun 11, 2024 Ilit Raz (Joonko) SEC and DOJ AI-powered recruitment platform Platform did not work as described, per the SEC complaint Parallel securities fraud and wire fraud charges
Mar 18, 2024 Delphia and Global Predictions SEC Delphia: client data fed its AI investment process. Global Predictions: “first regulated AI financial advisor” Neither had the capability claimed, per the SEC orders Settled; $225,000 and $175,000 civil penalties under the Advisers Act and Marketing Rule

Four lessons carry straight into agentic claims:

  • The hidden human is discoverable. Presto’s and Nate’s claims failed because the people doing the work showed up in staffing, contracts and internal messages.
  • The metric is evidence against you if it measures the wrong thing. According to the Presto order as summarized by Cooley, the company’s “non-intervention” rate counted intervention by restaurant staff and left out its own offshore agents.
  • Disclosure controls are part of the charge. The SEC also found Presto had no established process for drafting, reviewing or approving its periodic reports (D&O Diary).
  • Private plaintiffs don’t wait for regulators. DLA Piper notes the SEC brought no AI-focused actions in the first half of 2026 while private AI-related securities filings continued. Cornerstone Research counted 15 AI-related securities class actions in 2024 and 16 in 2025 (Bloomberg Law).

Where Agentic Claims Surface and Get Scrutinized in Financial Services

The same agentic claim usually appears in four places, and each has its own reviewer. A sentence that starts in a vendor’s RFP response can end up in a bank’s earnings call script, then in an examiner’s document request.

Venue Where the claim appears Who tests it What they ask for
Investor disclosures and earnings calls 10-K and 10-Q risk factors and MD&A, earnings call remarks on efficiency, investor day decks, press releases SEC Enforcement, including the Cyber and Emerging Technologies Unit created in February 2025 with AI-related fraud among its priorities; shareholder plaintiffs after a price drop The basis for each statement, the disclosure controls that reviewed it, and whether stated automation rates match the records
Regulatory examinations Responses to exam requests, written supervisory procedures, AI inventories SEC Division of Examinations; FINRA Under its fiscal 2026 priorities, the SEC reviews the accuracy of registrants’ representations about AI capabilities and whether firms have policies to monitor or supervise their AI use. FINRA’s 2026 report flags agents acting without human validation or beyond their intended authority (Global Relay)
Procurement and vendor due diligence RFP responses, security and AI questionnaires, statements of work, product documentation Procurement, third-party risk, model risk, information security Proof the product does what the questionnaire says, in an environment like yours, plus the monitoring the buyer will need to supervise it
Customer-facing marketing Websites, client communications, social posts, product pages SEC Marketing Rule exams for advisers; FINRA communications review for broker-dealers; the FTC for consumer claims Substantiation on file before the claim is published. Delphia and Global Predictions were charged under the Marketing Rule for claims in advertising

Banks have one more gap to plan for. SR 26-2, the April 2026 model risk guidance from the Federal Reserve, OCC and FDIC, places generative and agentic AI outside its scope, so the evidence trail for agents has to come from the bank’s own AI governance program. Our SR 26-2 control framework for generative and agentic AI maps what that program needs to cover.

Two consequences follow for financial institutions. First, a buyer that repeats a vendor’s agentic claim to clients or investors owns that claim; the vendor’s deck is not a defense. Second, IR, legal, procurement and the AI program office need to answer from one evidence base. When the earnings call says “agents” and the exam response says “human-reviewed tool,” the gap between them is the finding.

The Substantiation Test: Six Dimensions of an Agentic Claim

An agentic claim holds up when each of its six dimensions is backed by an operational artifact: a record the system produced while working, not a document written about it. Issuers can run the test on every agentic sentence before it goes into a filing, script or client communication. Buyers can send it to vendors and ask for the artifact instead of a yes or no.

Dimension The claim sounds like What would have to be true Artifact that proves it Red flag
Autonomy “The agent handles reconciliations end to end.” The agent receives a goal and decides the steps. A measured share of tasks finishes with no human action at any step. Execution traces showing each planned step and action with timestamps; an automation rate computed from those traces, with “human intervention” defined to include every person, in-house or outsourced The rate excludes some humans, as Presto’s did; the only proof is a demo
Tool and system access “The agent updates the policy administration system.” The agent holds scoped credentials to write to named systems of record, and those writes carry the agent’s identity. A permission inventory per agent (tools, read or write, environments); system-of-record audit logs that show the agent’s service identity as the author “Integrates with” means read-only; agent output lands in a queue a person releases
Decision scope “The agent approves claims under $5,000.” A written delegation of authority sets what the agent may decide, and runtime controls enforce it. The delegation document mapped to policy rules; guardrail configuration; logs of actions the guardrails blocked Scope lives only in the prompt; no blocked action has ever been recorded
Human checkpoints “A human reviews every high-risk decision.” Named checkpoints fire every time the condition is met, and a reviewer sees enough to disagree. Approval logs with reviewer, timestamp and decision; the human override rate, meaning the share of agent proposals a person changed or rejected, trended over time An override rate near 0% suggests rubber-stamping; a very high rate means people are doing the work, as at Nate
Evidenced outcomes “Agents cut onboarding time by 60%.” A baseline, a measurement period, a defined population and the same metric definition before and after. Pre-deployment evaluation results on a representative test set; production metrics computed with the same definitions; both dated A vendor case study with no baseline; pilot numbers quoted as production results
Failure handling “The agent escalates when it is unsure.” Failure modes are defined, with a fallback, an escalation path and a way to stop or roll back the agent. Intervention and incident logs; escalation records; error rates by type; records of kill-switch and rollback tests No failure has ever been logged

The human override rate deserves particular attention. It is the one number that tests autonomy and oversight at once. A rate near zero tells an examiner the checkpoint may not be real. A rate near 100% tells an investor the agent may not be real. Track it from the first day in production, and be ready to explain whichever way it moves.

None of these artifacts appear on their own. The agent has to be built to produce them, starting with traces: our guide on what to instrument in an AI agent explains which spans and decision data to capture. Human checkpoints deserve the same scrutiny, because one that works on paper can still fail in practice, as we explain in Agentic AI Risk Management and the Limits of Human-in-the-Loop Oversight.

The figures in the “claim sounds like” column are illustrative examples, not data.

Why a Demo Is Not Evidence

A demo shows one path, on data the presenter chose, on a day the presenter chose. It cannot show a rate, a failure or who acted in production last quarter. The SEC’s complaint against Saniger alleges he staged product demonstrations that made Nate’s AI appear to work autonomously (Harris Beach Murtha).

Examiners and plaintiffs look at behavior over a period and buyers should ask for the same thing:

  • Production execution traces for a recent period, sampled by the buyer rather than curated by the vendor
  • Evaluation results on the buyer’s own data or a test set built to match it
  • The human override rate and intervention log for the same period
  • The definitions behind every automation or accuracy figure in the proposal

A vendor that can only offer a demo, a case study and a reference call has not yet produced evidence. That may be acceptable for a pilot. It is not enough to support a public statement that the firm runs on agents.

For more questions to put to monitoring vendors, see From Dashboards to Decisions: A Buyer’s Guide to Agentic AI Monitoring.

The Evidence to Keep From the Day an Agent Goes Live

Evidence collected after a claim is public looks like a defense. Evidence collected from go-live is a record. The artifacts below cost little to capture at launch and a great deal to reconstruct a year later, when the person who configured the agent may have left.

A Strategic Guide to the Agentic AI Audit walks through what auditors ask for when they reconstruct an agent’s decision months later.

For every agent in production (issuers and buyers alike)

  • An inventory entry naming the owner, purpose, systems of record touched, permissions, delegation of authority, version and go-live date
  • A pre-deployment evaluation report, dated and tied to the exact version deployed
  • Execution traces kept under the firm’s records-retention schedule
  • Guardrail configuration and a log of every action it blocked
  • Approval and override logs, with the override rate reported monthly
  • Incident, escalation and rollback records
  • A change log of model, prompt, tool and permission changes, with who approved each

For issuers making public statements

  • A claims register that maps every public agentic statement to the artifact behind it
  • Review of agentic statements inside disclosure controls, repeated before each filing and earnings call
  • Metric definitions (automation rate, straight-through rate, hours saved) fixed before their first public use
  • A consistency check across investor materials, marketing and exam responses

For buyers in procurement and vendor risk

  • Contract rights to production logs, evaluation results and notice of material model or tool changes
  • Re-evaluation after each significant vendor update
  • A rule that no vendor’s agentic claim is repeated to clients or investors until the firm holds its own evidence

How Lumenova AI Produces the Evidence

Lumenova AI gives financial institutions the operational record that turns an agentic claim into a supported one. The platform produces evidence as agents run, so IR, legal, procurement and the AI program office answer from the same dated record.

  • Continuous observability shows what agents did. Execution traces, tool calls, guardrail interventions and human overrides are captured in production, so automation and override rates come from records rather than estimates. This covers the autonomy and tool access rows of the test.
  • Adversarial testing shows how agents fail. AI Security attacks the agent the way an adversary would, turns a finding into a draft policy a person activates, and replays the original attacks to prove the fix holds. That gives the failure-handling row its evidence: a failure found, fixed and retested, on record.
  • Pre-deployment evaluations quantify capability. Agents are tested against representative scenarios before launch and after each change, producing dated results tied to a version. That is the baseline an outcome claim needs.
  • Runtime policy enforcement checks agent traffic against approved policy. The Lumenova AI Platform holds the policy and the approved use cases. The AI Gateway checks the traffic routed through it against them, and each decision keeps the rule, the reason and the input it saw.
  • A centralized governance workflow keeps the record auditable. The AI Inventory & Registry ties each agent to its owner, permissions, approvals, evaluations and change history, so the answer to an examiner’s request already exists before the request arrives.

For the full approach, check out our article on AI Agent Observability: Executive Guide to Governance & Risk.

The result is a claims process that runs in the right order: evidence first, statement second.

Test the Claim Before Someone Else Does

Every agentic statement a financial institution makes will eventually be compared with what its agents did. Run the six-dimension test on your own disclosures and on every vendor proposal now, while the gaps are still internal findings.

To see where your agent oversight and evidence trail stand today, take the Agentic AI Risk and Governance Assessment, or book a demo to see how Lumenova AI records what your agents actually do.

Frequently Asked Questions

Agent washing is marketing an assistant, a workflow automation or a chatbot as an autonomous AI agent. Gartner described it in 2025 as rebranding AI assistants, RPA and chatbots without substantial agentic capabilities.

AI washing overstates whether or how a firm uses AI. Agent washing overstates what a system does on its own. Agentic claims describe actions in systems of record, often with an automation rate attached, so execution logs can confirm or refute them.

Ask who takes the final action. A real agent plans steps and writes to systems of record under its own identity, within limits a person set. Ask for production execution traces, a permission inventory, evaluation results on your data and the human override rate, not a demo.



Agentic statements in filings, earnings calls and marketing are representations of fact about operations. The SEC’s fiscal 2026 exam priorities include reviewing the accuracy of registrants’ AI representations, and private plaintiffs continue to file AI-related securities class actions. A claim without operational evidence behind it is exposed in both channels.



Not under that name. The SEC’s Presto Automation order and the SEC and DOJ actions against Nate’s founder both centered on claims that tasks were completed without human intervention, which is the core agentic claim.



Operational evidence produced while the agent works: execution traces, pre-deployment evaluation results, guardrail and intervention logs, approval and override records, and incident records, all dated and tied to the deployed version.




Related topics: AI EthicsAI Transparency

Make your AI ethical, transparent, and compliant - with Lumenova AI

Book your demo