Sending us evidence privately

What happens to evidence a company sends Roiva privately: it never counts toward the AI in Practice Index, and is used only for the separate Verified AI Rating.

The AI in Practice Index is built only from public evidence. Evidence a company sends us privately never counts toward the index, in any edition: not toward a score, a leader decision or a profile. If you've found a public source we missed, you don't need these guidelines: write to data@roiva.ai with a link to it, and see the method for how corrections work.

What is private evidence for?

Private evidence can be used for one thing only: the Roiva Verified AI Rating, a separate, opt-in rating. It assesses evidence a company holds that isn't public, such as records of an AI use running or of a Responsible AI practice operating. It is never part of the index, and an index result never depends on it. The rating's fee is $2,500, $7,500 or $15,000, set only by the company's annual revenue and the same for every result. It is paid when the company signs the rating agreement, before the review starts. The rating page sets out its terms. The index itself stays free: any company in an edition's study can ask for its own score and the sources behind it, and claim, correct or remove its profile, at no charge. To apply for the rating, write to data@roiva.ai; we'll send the agreement, the questionnaire and a secure upload link. Please don't send documents by email.

What can a company send?

A company names an accountable executive, who confirms the evidence is accurate, current and the company's to share, and the company signs an agreement with Roiva covering confidentiality, use and retention first. It then answers the rating's questionnaire: for each criterion it claims, the information below, the AI use it belongs to, and the documents that show it.

Any document from the company's own records that shows the information counts, if it is dated and current on the rating date. The documents named below are only examples. A reviewer reads each document whole and checks that it shows the information required; we don't take screen-shares or live demonstrations. A practice also needs records of it operating, not only a policy describing it.

For Level 4, those records go back further, and the AI inventory is checked against the books:

  • Six months of operation. Each practice Level 4 requires (AI inventory, risk and impact assessment before deployment, monitoring in production, incident handling, oversight of third-party AI, and AI literacy and training) needs records of it operating across the six months before the rating date: at least one dated six or more months before it, and at least one dated within the three months before it. Each risk and impact assessment is dated before its system went live, and at least one is dated six or more months before the rating date. These older records count even though they aren't current.
  • The inventory against the books. The company sends a list, from its accounts payable or general ledger, of the AI products and services it paid for in the 12 months before the rating date, including AI features bought from a general vendor where the invoice or contract names them. Its finance officer signs the list off. A reviewer checks that every item appears in the AI inventory and among the vendors in the company's oversight of third-party AI. If one is missing, neither practice meets Level 4 until the company updates its records.
  • A practice without these still counts. It counts toward the rating score, and only its Level 4 requirement waits.
  • Practices shown in public. A practice that public sources show in place can meet its Level 4 requirement without the six months of records. The check against the books still applies to the AI inventory and the oversight of third-party AI.
Criterion Information required Documents that often show it
A use of AI exists its name; the process and business unit it serves; what the AI does; a named owner (a role) a project charter, a system description, a vendor contract with pricing removed
A use is live the go-live date; a usage measure (cases, transactions or users) covering at least 30 days within the 6 months before the rating date, with its period a go-live record, a usage report or usage-log extract, dated screenshots
A use is at scale the process's eligible volume in the business unit and the share the use handles (at least half), for the same period; or each business unit it runs in, with its go-live date a volume report with the process's total eligible volume; deployment records for each unit
AI policy its title and owner; the approving body and date; in force on the rating date; it names AI the policy
Responsible AI policy as for an AI policy, plus the commitments it makes (fairness, transparency, human oversight or consumer protection) Responsible AI principles, an NAIC Model Bulletin AIS program document
Board oversight the board or committee; AI named in its remit, or on an agenda with the meeting date; in place on the rating date a committee charter, a minutes extract, a board-pack agenda
AI leader the person's name, title and reporting line; their accountability for AI; in the role on the rating date an organization chart, a role description
AI inventory the number of AI systems listed; what is recorded for each (owner, purpose, risk level, built in house or bought); the date last updated, and a version or change log from 6 or more months before the rating date; for Level 4, a list from accounts payable or the general ledger of the AI products and services paid for in the 12 months before the rating date, signed off by a finance officer, each of which appears in the inventory an AI inventory or register extract, current and from 6 or more months earlier; an accounts-payable or ledger extract
Risk and impact assessment before deployment which deployed systems were assessed, and when relative to go-live; the risk level found; who approved risk or impact assessments for deployed systems
Testing and validation, including bias testing the systems tested; what was tested (accuracy, robustness, bias); the dates; the results recorded test and validation reports, including bias tests
Monitoring in production what is monitored (performance, drift); how often; the owner; the latest report date monitoring or drift reports
Incident handling the escalation route and its owner; that incidents are logged; the latest entry or review date an incident procedure and log extract
Oversight of third-party AI the AI vendors in scope, including every vendor on the accounts-payable list for the AI inventory; the due-diligence steps; the contract terms (audit rights, data use); the dates. Or that all AI is built in house, as the inventory and the accounts-payable list show AI vendor due-diligence files, contract terms
Human oversight of consequential decisions which decisions; who reviews them, at which step; the override or appeal route; how many were reviewed in the period a review procedure, human-review or appeal logs
Data governance the data sources of the AI systems; the quality and lineage checks; the owner; the date data quality and lineage records
AI literacy and training the program; who must take it; completions (a number or share of staff); the period training records and completion rates
Transparency to the people AI decisions affect which decisions people are told about; where and how they are told; what an adverse decision explains consumer notices and adverse-decision letters that name AI
Workforce transition and staff voice the roles AI changed and what happened to them (redeployed, reskilled, absorbed growth, attrition, eliminated), with numbers; how staff were informed or consulted before rollout, and when redeployment or reskilling records; notices, surveys or consultation records from before rollout
A measured result the figure and its unit; the period it covers; the method and baseline; the use a KPI report
Level 5 figures: a result in cash or hours, net of its costs a source-system extract the reviewer recomputes to within 10%; the fully loaded costs for the same period; the revenue of the business units the uses serve; a finance officer's sign-off; for a figure in hours, where the released time or roles went a method note, the baseline, a data extract, a cost schedule, segment or management accounts, headcount or HR records, the sign-off

Plans, targets and forecasts don't count. Neither do general statements about AI.

Level 5 figures take a stricter check than documents. A figure counts only once a reviewer with accounting competence has recomputed it from the company's source-system extract, to within 10%, checked its costs against invoices or the general ledger, and received a finance officer's signed confirmation. It is counted net of its costs. The rating page sets out the full check.

What shouldn't a company send?

  • Personal data about customers, policyholders, patients or employees. Remove it first.
  • Protected health information.
  • Trade secrets the evidence doesn't need, such as model weights, prompts or pricing.
  • Material non-public information. A listed company should consider whether a figure is material before sending it.

How is private evidence handled?

  • Used only to rate the company that sent it. Never in the index, its research or its corrections; never in benchmarks, averages or peer comparisons without separate written consent; never in sales, marketing or Roiva's product; and never to train AI models.
  • Kept confidential. Submissions are held apart from the index's data, encrypted, and readable only by the named reviewers and the AI agents working for them. Submissions are never used in sales or marketing, and we never tell anyone, inside or outside Roiva, which companies applied, except as needed to review them.
  • Reviewed by people. AI agents (Claude, made by Anthropic) may do a first pass, such as matching each item to what it shows and checking dates. They never decide whether an item counts: a Roiva reviewer with no commercial interest in the company reviews every item and records whether it is verified. Level 5 figures are decided by a reviewer with accounting competence.
  • What "verified" means. Every scorecard, the one the company receives and the one it publishes if it chooses to, carries both meanings. For a figure in cash or hours: "Verified means Roiva's reviewer recomputed the figure, to within 10%, from source-system data the company provided; checked its costs against the company's invoices or ledger; and received a signed confirmation from the company's finance officer. It is the company's measurement. Roiva didn't audit or examine the company's financial statements or controls, didn't test its source data, and doesn't establish that AI caused the result." For every other item: "For every other item, 'verified' means a Roiva reviewer read the company's documents and found that they show the information the method requires. Roiva didn't test the company's controls or systems, or confirm the documents beyond reading them."
  • AI processing. The agents run under Anthropic's commercial terms. Anthropic, PBC processes submissions for us as a sub-processor. It doesn't use them to train its models, and may keep them for a limited time for safety monitoring, as its terms set out. A company can ask for a people-only review, at any time before its review starts, at no extra cost. Its documents then never go to an AI agent.
  • No trading. Roiva, our staff, our reviewers and their households don't trade in a company's securities from receiving its submission until 30 days after the rating decision, and after that for as long as we know we hold material non-public information about it.
  • Personal or health data sent by mistake. If we find personal data or protected health information in a submission, we delete that part, tell the company, and don't use it.
  • Deleted on a schedule. Documents are deleted 90 days after a rating expires or is withdrawn, or 30 days after a company withdraws its submission. A submission with no decision after six months lapses, after a month's notice, and is treated as withdrawn.
  • A short record is kept. For 3 years after the documents are deleted, we keep a verification record: for each item, the criterion, who reviewed it, the date, the outcome and a fingerprint of the file. It holds no figures unless the company published them.
  • Deleting everything. A company can ask us to delete everything at any time. That withdraws its rating, and its documents are deleted at once. The verification record is deleted too, unless the rating was published. Then we keep it for its 3 years, so we can show how a public rating was decided.
  • Never published without consent. We never say that a company applied, was rated or withdrew, or publish anything it sent, unless the company agrees in writing. The one exception: once a rating a company published or announced is no longer valid, because it expired, was withdrawn or its scorecard was unpublished, the page its badge links to shows only that, with the date. If the company keeps showing the badge or announcing the rating, we may say the same to anyone who asks, or publicly. We never publish our reasons for withdrawing a rating.
  • If we get it wrong. If we find we haven't handled a submission as this page says, we tell the company that sent it, say what happened and what we've done about it, and correct any rating it affected.