Decentralised News · Agentic Finance · Research cutoff 19 November 2026
By Heath Muchena
Direct answer: Before buying, funding or recommending an AI-agent startup, ask for evidence of an operating product, paid demand, financial reality, controllable permissions, security, data obligations and a viable exit. A pitch deck can explain ambition; it cannot substitute for transaction records, customer agreements, incident reports or reproducible demonstrations.
What Matters
- “The agent did something” is weaker evidence than a trace linking an authorised task to an accepted result.
- Revenue, transaction volume and token emissions are different measures and should be recorded separately.
- A security audit has a scope and date; a logo alone says little about a deployed application.
- Buyers and journalists need to distinguish evidence supplied, evidence independently tested and unanswered requests.
Scope of This Investigation
DN derives a twelve-record request from public NIST, OWASP, FTC and UK ICO material and from the economic distinctions explored in our own agent research. It is a proposed investigative protocol, not a claim that 12 named startups were audited or that the standards bodies endorse this exact list. We performed documentary research, not customer interviews, red-team tests or financial-statement audits.
The Signal: product claims outpace proof of economic delivery
Agent products increasingly combine model inference, third-party tools, customer data, payments and delegated authority. The most important commercial question is whether the company can demonstrate its full operating cycle: instruction, permission, attempt, outcome, correction and cost. Each stage can appear successful while another fails.
US FTC actions against allegedly misleading AI business opportunities underline why income and professional-performance claims need substantiation. Those proceedings are not a dataset measuring the failure rate of agents. The NIST and OWASP frameworks supply a vocabulary for testing different risk categories, but neither can certify an individual startup solely because its name appears in marketing. [1] [2] [3] [4]
DN Alpha Thesis
The scarce resource in agentic finance is not another system that can call a model. It is a repeatable evidentiary chain from delegated authority to an accepted economic outcome. That chain determines whether a product is useful, whether its claims are measurable and whether a buyer can recover when something goes wrong.
The strongest startups will not necessarily have the largest token communities or the flashiest demonstrations. They will make relevant records available, explain what those records cannot prove and let independent reviewers reproduce the basic results without trusting a founder’s narration.
The Twelve Evidence Records
| # | Evidence request | What it should contain | Weak substitute |
|---|---|---|---|
| 1 | Product reality | An accessible, versioned demonstration of one supported workflow, its input requirements, exceptions and output receipts | A polished demo video without matching logs |
| 2 | Customer demand | Paid-customer counts by period, renewal cohort and the definition of an active paying account | Registered wallets, trial signups or bot invocations presented as revenue |
| 3 | Revenue accounting | Recognition and cash collection definitions, net of refunds and pass-through expenses | GMV or token volume relabelled company earnings |
| 4 | Economic unit | Cost per accepted task including retries, model/API usage, integration and human rescue | Cost per successful model response only |
| 5 | Operational reliability | Incident history, downtime denominator, failed-job rate and recovery process | A 99.9% figure with no observation window |
| 6 | Permission design | Actual permission boundaries, signing authority, allowlists, budgets and independent stop control | A UI toggle presented as guaranteed revocation |
| 7 | Supply chain | Model, tool, connector and hosting dependencies with version/change notices | A diagram without ownership or update records |
| 8 | Security evidence | Threat model, testing coverage, incident response owner and reproducible test results | An audit logo without report scope |
| 9 | Data rights | Controller/processor roles, training/retention terms, sub-processors and deletion or export obligations | A broad assurance that no data is ever stored |
| 10 | Financial authority | Who can approve or block trades, transfers and external messages; how the action is recorded | A disclaimer offered in place of enforceable controls |
| 11 | Claims and marketing | The evidence behind success, savings, return or safety claims, including negative cases | Guaranteed passive income or cherry-picked wins |
| 12 | Exit and portability | Export format, contract termination, credential revocation, service migration and custody of customer records | An account-close button with no verified data handoff |
1. Product and acceptance are not synonyms
For a specific workflow, preserve the task description, model/tool version, allowed sources and expected output. A run may produce a response but fail the user’s intended outcome. If a research agent discovers ten documents, for instance, the customer may still reject the answer because it cites the wrong reporting period. Acceptance criteria must be described before the comparison, not selected after results are known.
Ask the startup to demonstrate an ordinary success and a controlled failure. The latter should show what the system recorded, whether it attempted an unauthorised action, how it communicated uncertainty and how control was restored. The demonstration should not require real customer secrets or financial transfers. A sandbox record can answer an engineering question, while leaving live reliability unmeasured.
2. Separate paying customers from promotional demand
A popular API, a social account with many followers and a large number of newly created wallets are not necessarily paying customers. A startup should define whether it counts accounts, humans, organisations, subscriptions, transactions or agent identities. It should distinguish free from paid and show the reporting window. Retention cannot be inferred from a lifetime signup total.
For revenue, request enough explanation to distinguish cash receipts, recognised revenue, gross transaction value and protocol fee flow. If a crypto protocol pays users tokens to create activity, those emissions are not the same as a third party paying for a useful service. A buyer does not need the founder’s personal account information to ask for the definitions and evidence supporting a material public claim.
3. Costs must include unsuccessful work
The apparent cost per prompt can look excellent while the actual workflow costs more. Repeated tool calls, rejected responses, data reconciliation, API failures, cloud resources, incident investigation and a human completing the unfinished task all consume resources. Define the unit as an accepted outcome, then account for every attempt attributable to the same requested job.
A startup selling a $200 monthly service to an organisation with twenty users needs a different cost model from one offering $200 per completed financial reconciliation. Before comparing prices, state what the customer is entitled to request and how exceptions are billed. If the product charges a flat amount but its underlying model expenses are usage-dependent, the high-volume customer segment may alter the economics sharply.
4. The permission diagram is an economic document
An agent acting on a customer’s behalf needs a boundary that the model cannot casually rewrite. Ask which entity owns the account and signs the transaction, which tokens or instruments can be moved, what daily and per-action limits exist, and where an independent stop or revocation mechanism operates. Test whether the control applies at signing or only in the interface.
The OWASP Agentic Applications framework explicitly addresses agent goal hijack, tool misuse, identity and privilege abuse and supply-chain issues. The existence of those categories is not evidence of a compromise at a particular startup. It does justify asking how the company tests the trust boundary between an untrusted instruction and a consequential tool call. [2]
5. Security reports have dates and deployments
An audit may cover one smart contract or SDK revision and leave later application changes outside scope. Record the audit’s publisher, assessed commit or deployed address, test environment, exclusions, findings, remediation and any independent retest. A marketing badge without those fields should be treated as a statement awaiting support.
Ask for a security incident-response owner and for a simple description of how compromised credentials are revoked. A product may have excellent model performance and still offer unacceptable operational exposure if it holds more authority than the task needs. Data retention and deletion should be reviewed separately from source-code vulnerability testing.
6. Contracts determine who can change the result
Agent startups often rely on cloud hosts, model providers, wallet services, data vendors and orchestration layers. A diagram should name which provider can change pricing, rate limits, data access, supported functionality or terms. It should also distinguish the startup’s own service promise from commitments made by upstream suppliers.
For relevant UK data processing, ICO guidance describes controller and processor responsibilities, sub-processors and end-of-contract provisions; it is under review following legislative changes. A company serving buyers in multiple countries must not confuse that specific guidance with a universal law. The investigative request is for the actual terms and a dated statement of the roles that apply. [5]
7. Make adverse evidence part of the review
Request cancellation and refund counts where relevant, incident severity and customer complaint categories. Smaller companies may not have enough observations to estimate a stable failure rate. In that case the correct finding is insufficient history, not perfect reliability. An honest startup should be able to describe at least what it has instrumented and how it will know when a change worsens outcomes.
A transparent founder may rightly withhold private customer contracts or commercially sensitive details. That does not invalidate the company. The analyst can accept appropriately redacted evidence, independent attestations or demonstrations performed under mutually agreed controls. The evidence standard should be proportional to the harm a mistake could cause.
DN Startup Evidence Coverage Register
Use these twelve checks to inventory the records obtained. A checked box means you have a dated and reviewable record, not that its content has been independently verified or that the startup is safe.
This instrument is local and does not transmit or store data. It is a completeness inventory, not a security grade or investment score. A full 12/12 can still contain misleading, outdated or false records.
Three outcomes a responsible investigation can reach
Evidence substantially supports the claim
The product is available, the performance measure is defined and supported by recent observations, failure modes are disclosed and contractual controls match the intended deployment. State the actual supported scope, not a claim of universal safety.
Useful product, unresolved exposure
The system may work on a limited workflow, but the requested financial or permission evidence is missing. A buyer can still trial the product with constrained read-only permissions and no sensitive customer data, provided the applicable terms permit it. The conclusion is suitability for a bounded experiment, not full deployment.
Claim contradicted by evidence
A promised return, cost, permission boundary or cancellation right conflicts with an operative contract or a repeatable test. Preserve the conflict and request a specific correction. Never turn a founder’s assertion into a scored benchmark simply because it was presented in a polished slide deck.
Methodology and Limits
This is an original DN research protocol based on the primary and standard-setting documents listed below, plus transparent distinctions between transactions, revenue, user counts, permissions and outcomes. It is intended for buyers, editors, procurement teams and startup researchers. It is not an exhaustive security audit, a financial audit, a legal opinion or a rating of a particular company.
No startup interviews, proprietary records, wallet connections or payments were performed. The twelve records are DN’s proposed evidence requests, not an official NIST, FTC, OWASP or ICO checklist. We do not claim that counting supplied documents measures truthfulness or safety.
Primary Documents
- NIST AI RMF Generative AI Profile. Voluntary US risk framework, published July 2024; useful for governance and measurement, not a product certification.
- OWASP Top 10 for Agentic Applications 2026. Community security risk framework published December 2025; identifies common classes of misuse and privilege issues.
- FTC Operation AI Comply. US enforcement action illustrates importance of substantiated claims; allegations are not a statistical measure of startup failure.
- FTC Air AI enforcement. US agency allegations about AI resale earnings and refund promises, not a finding about all AI companies.
- ICO controllers and processors. UK-specific data protection contract guidance under review; commercial contracts require fact-specific legal assessment.
Frequently Asked Questions
How do I check whether an AI agent startup has real revenue?
Request a precise definition of paying customers, recognised revenue, customer receipts and fees, separated from transaction volume and emissions.
Does a smart contract audit prove an agent is safe?
No. Its scope, version, date and findings matter, and the broader application may expose other risks.
Should an AI agent provider disclose private customer data?
No. Evidence should be appropriately redacted or independently attested, avoiding unnecessary personal information and secrets.
Is 12 out of 12 a safety score?
No. It means the required categories of evidence were received according to the reviewer’s judgement. Each document must still be checked for quality and applicability.
Continue the Research
See DN’s Know Your Agent research and agent repeatability study for complementary frameworks. Research is educational and no investment or security guarantee is made.




