AI Usage & Disclosure Matrix

Aperim Pty Ltd (ABN 46150699737; ACN 150699737) is incorporated in New South Wales, Australia and operates the aicial brand. In this matrix, “aicial” refers to that brand; “we”, “us” and “our” refer to Aperim Pty Ltd; and “you” refers to the person or organisation reading this matrix, including a customer, prospective customer or other individual affected by our use of artificial intelligence.

This matrix is the technical companion to our Responsible AI Usage Policy. It names each AI-enabled feature aicial operates or plans to operate, or that a customer may choose to enable, and states what each one does, the general nature of the data it uses, its risk classification, its human-oversight level and how we disclose it to a user. That policy states the principles, governance and disclosure standard that apply to every AI system we operate; this matrix applies them feature by feature. Where this matrix and that policy appear to differ, that policy governs, as its section on this matrix states.

This matrix applies to our website, waitlist and services, on the same basis as that policy, including our planned self-serve software service once it is made available. Every feature listed below is one of three things: a feature we operate today, as part of our current productised services; a feature of our planned self-serve software service, which is not yet available; or, once that service is available, a feature a customer may choose to enable. Each entry below states which of these applies to it, and where a feature depends on a customer actively turning it on, this matrix says so.

For each AI-enabled feature, this matrix states what the feature is, its purpose, the general nature of the data it takes as input, the general nature of the data it produces as output, its risk classification, the level of human oversight that applies to it, and how we disclose it to a user. These are general descriptions of categories of data, not an exhaustive technical specification. Our Privacy Policy is the authoritative statement of how we collect, use, disclose, retain and protect personal information within any of these systems.

Risk classification uses the same four-tier risk framework that policy uses to describe the European Union’s AI Act (Regulation (EU) 2024/1689) — unacceptable, high, limited and minimal risk — for the reasons its section on risk classification explains: that framework is external, well understood and independently legible to a customer, regulator or investor, not because every aicial feature is subject to EU jurisdiction. We do not operate any feature that would fall within the unacceptable-risk tier.

Human oversight is described below using four terms of our own, chosen to reflect the spirit of the human-oversight commitments set out in that policy’s section on human oversight and contestability, which states the underlying principle rather than this specific four-level terminology:

• In-the-loop. A person must take an affirmative action, such as an approval, before the system’s output takes effect. The system cannot act by itself.
• On-the-loop. The system, or a person acting on its output, can proceed without a case-by-case approval, but a person monitors its operation and can intervene, adjust or stop it.
• Out-of-the-loop. The system performs a specific function without case-by-case human involvement in an individual instance, subject to the human-set rules, system-level oversight and correction described in this section.
• In-command. A person retains the authority to override, disable or reconfigure the system at any time, whatever its day-to-day oversight level. This applies to every feature in this matrix: Aperim Pty Ltd’s leadership remains in command of each one, consistent with the accountability described in that policy’s governance section.

Each feature entry below states the oversight level that applies to its ordinary operation, in addition to the constant in-command authority described above.

Disclosure states how we actually tell a user they are dealing with an AI system or AI-generated content for that feature, consistent with the standard set out in that policy’s disclosure commitments section.

What it is. The system behind aicial’s cross-platform performance reports. It generates benchmark comparisons, diagnoses reach and engagement changes, and produces the diagnostic narrative a customer receives in a Social Performance Audit, Outcome-Proof Pack or Benchmark & Strategy Engagement — our current productised services.

Purpose. Diagnostic and informational: explaining what is happening in a customer’s own performance and how it compares to a stated benchmark. It does not take action on a customer’s accounts.

Data inputs. The customer’s own connected-account performance data, within the scopes the customer authorises, together with consented, licensed and lawfully public aggregate data used to build the benchmark comparison.

Data outputs. Benchmark reports and diagnostic narrative delivered to the customer, describing the customer’s own performance and how it compares to the benchmark.

Risk classification. Limited risk. The output is informational and reviewed by the customer; it is not a high-stakes automated decision about a person.

Human oversight. Out-of-the-loop. The engine produces each report from human-set methodology and rules, without case-by-case approval of an individual report before it is generated. The customer then reviews the report and decides what to do with it; the engine does not act on a customer’s accounts by itself.

Disclosure. Labelled as AI-assisted analysis within the delivered report. Every benchmark states its sample size and coverage limits, and any modelled or estimated figure in the report — including modelled attribution — is labelled as modelled or estimated, consistent with the honest-numbers principle set out in that policy.

What it is. The system that proposes content variants, informed by the customer’s own past content performance, and tests them through the customer’s own experiments on their own connected accounts, suggesting when to publish them. This feature forms part of our planned self-serve software service, which is not yet available.

Purpose. Assists a customer in creating and timing their own content. It does not publish anything by itself.

Data inputs. The customer’s own past content performance data and the customer’s own brand and content material.

Data outputs. Content variants and scheduling suggestions presented to the customer for review, and the results of any experiment the customer approves and runs on their own connected account.

Risk classification. Limited risk.

Human oversight. In-the-loop. Nothing publishes to a connected account, including a variant used in an experiment, until the customer reviews and approves it.

Disclosure. Each suggestion is labelled within the product interface as AI-generated, and every content variant it generates also carries the machine-readable content-provenance marking described in that policy’s disclosure commitments section. Consistent with the no-manipulation principle set out in that policy, every suggestion stays confined to the customer’s own honest content, timing and format choices.

What it is. The system that looks for accounts that may be copying a customer’s name or likeness, and for content that may present a deepfake risk to a customer’s identity. This feature forms part of our planned self-serve software service, which is not yet available.

Purpose. Defensive. It alerts a customer to a potential impersonation or deepfake risk so the customer can assess and act on it — we monitor, and the customer decides what happens next.

Data inputs. Public account and profile data relevant to detecting a potential match, together with the customer’s own reference material submitted under the express, separate biometric consent our Privacy Policy describes. We compare that reference material against handles, metadata and perceptual-hash fingerprints of public content to identify likely impersonation or deepfake risk; it does not involve biometric search over third-party or scraped media.

Data outputs. An alert to the customer describing a potential impersonation or deepfake match, for the customer to assess.

Risk classification. Limited risk.

Human oversight. On-the-loop for monitoring and alerting: we monitor on an ongoing basis and alert the customer, who decides what happens next, without a case-by-case approval before each alert. In-the-loop before we take an external step on the customer’s behalf, such as a takedown request: a person reviews every adverse finding and approves that specific step before we take it, consistent with the human-oversight commitment set out in that policy.

Disclosure. Every alert is clearly labelled as an automated detection, not a confirmed finding.

What it is. The system that aggregates niche and competitor performance signals from consented, licensed and lawfully public data to build the benchmark statistics used across aicial’s reports, including in our current Social Performance Audit, Outcome-Proof Pack and Benchmark & Strategy Engagement.

Purpose. Honest coverage and benchmark reporting: placing a customer’s own performance within a wider aggregate picture.

Data inputs. Consented, licensed and lawfully public data, aggregated across many accounts, consistent with the sources our Privacy Policy describes, together with deal, rate and payment-experience data a customer chooses to contribute in exchange for access to the resulting aggregate benchmarks.

Data outputs. Aggregate benchmark statistics with a stated sample size. Except for a customer’s own connected accounts, or public figures where we have obtained legal advice supporting that analysis, this layer does not produce a profile-level report about an identifiable individual.

Risk classification. Minimal risk for the aggregation itself, because it produces aggregate statistical output rather than an assessment of an identifiable individual. Limited risk for the narrow public-figure case described below, because that case can produce a profile-level analysis of an identified individual, gated by a legal-advice approval before use. This is a different question from whether an activity involves profiling for privacy-law purposes: our Privacy Policy’s section on AI and personal information explains separately when benchmarking-related processing amounts to profiling, and that section’s protections apply regardless of this classification.

Human oversight. Out-of-the-loop for the aggregation itself: a fully automated statistical calculation that does not need case-by-case human review to produce an aggregate figure. In-the-loop for the narrow public-figure case described above: a person must obtain legal advice supporting that specific analysis and approve it before we act on it, so this output does not proceed by itself.

Disclosure. Every benchmark we publish states its sample size and coverage limits, consistent with the honest-numbers principle set out in that policy.

What it is. An optional automation that drafts and sends replies to messages on a customer’s own connected account, available where a customer chooses to enable it. This feature forms part of our planned self-serve software service, which is not yet available.

Purpose. Care and support automation on a customer’s own connected accounts.

Data inputs. The customer’s own account and message content.

Data outputs. Automated reply content sent on the customer’s behalf.

Risk classification. Limited risk.

Human oversight. On-the-loop. A customer must expressly opt in before this feature operates on their account; once enabled, it drafts and sends replies without a case-by-case approval of each one, and the customer monitors its behaviour and can review, adjust or turn it off at any time.

Disclosure. Every automated reply discloses that it is automated. This is not optional and is not something a customer can turn off while the feature remains enabled, consistent with the disclosure commitment set out in that policy.

We review this matrix on the same cadence as the Responsible AI Usage Policy: at least once a year, and sooner if we launch a materially new AI-enabled feature or a law that governs us changes, consistent with that policy’s review and updates section. The version, effective date and last-reviewed date shown at the top of this page identify the current matrix.

If a feature’s classification or oversight level changes as the feature, its use or the applicable law changes, we update this matrix and, where the change is material, that policy as well.

For questions about this matrix, to raise a concern about a specific AI-enabled feature, or to ask a person to review an AI-influenced output, contact:

Aperim Pty Ltd, which operates the aicial brand
ABN 46150699737
ACN 150699737
New South Wales, Australia
Email: ethics@aicial.com