AGENTPR™ RESPONSIBLE AI
Trustworthy media intelligence isn't a checkbox. It's an engineering discipline.
This page documents how we build, review, and disclose every AI-generated brief that leaves the AGENTPR™ platform.
The seven commitments
A skimmable summary of how AGENTPR™ builds, reviews, and discloses every AI-generated brief.
Engineering of Trust™
Methodology, transparency, and accountability — repeated the same way every brief.
Read section
Analytical assurance
Three claim-confidence levels are supported by five evidence-provenance labels, task-specific model-confidence scores where applicable, and a separate human-review status.
Read section
Human oversight
Briefs auto-flag for analyst review on low confidence, sarcasm, or crisis risk.
Read section
Data handling
What we collect, how long we keep it, and who can access it.
Read section
Third-party AI
Powered by Claude (Anthropic). Customer data is not used to train models.
Read section
Feedback loop
Report a wrong, biased, or unsafe output — every report is reviewed by a human.
Read section
Glassbox AI Policy
Our commitment to inspectable AI — full policy now published.
Read section
1.
Our framework — Engineering of Trust™
Engineering of Trust™ is the discipline behind every AGENTPR™ brief. It sits on three pillars: methodology (the GASD™ pipeline — Gather, Analyse, Synthesise, Deliver, run by four specialised agents), transparency (every source we read is shown; every source we filter out is explained), and accountability (distinct provenance, claim-confidence, model-confidence and human-review signals, with a clear paper trail from raw signal to final recommendation).
Trust isn't a marketing claim. It's a series of engineering decisions repeated the same way for every brand, every brief, every week.
Methodology
The GASD™ pipeline. Four specialised agents. Eight structured sweeps. The same sequence every time.
Transparency
Every source read is shown. Every source filtered is explained. If you can't trace it, we don't ship it.
Accountability
Distinct assurance layers, human-review workflows, and a clear paper trail from raw signal to final recommendation.
2.
Explainability — how analytical assurance works
Every material analytical claim in an AGENTPR™ brief carries a High, Medium or Low confidence level so analysts and decision-makers can weigh it appropriately:
High confidence
Corroborated across multiple credible sources with consistent sentiment and clear context.
Medium confidence
Supported by fewer sources, mixed signals, or partially ambiguous context. Worth acting on with care.
Low confidence
Thin sourcing, conflicting signals, or sarcasm / cultural-nuance risk. Surfaced for human review before any external action.
Each brief also shows source attribution per claim, the full list of sources read, and the list of items filtered out (and why). If you can't trace it, we don't ship it.
How AGENTPR™ communicates analytical certainty
| Assurance layer | Question answered | AGENTPR™ expression |
|---|---|---|
| Evidence provenance | Where did the evidence come from? | Monitored, Web-Verified, Print-Verified, AI-Sourced, AI-Inferred |
| Claim confidence | How strongly is the claim supported? | High, Medium, Low |
| Model confidence | How certain was the model about a defined machine task? | Percentage or task score, where applicable |
| Human review | Must professional judgement intervene? | Standard Review, Human Review Recommended, Human Review Required |
These layers operate together but are not interchangeable. A high model-confidence score does not convert an AI-sourced item into verified evidence and does not override a human-review requirement.
Model confidence is a machine-generated task score. It is not evidence verification, claim confidence or permission to publish. A high model-confidence score must never bypass required human review.
3.
Human oversight — when and why we flag for review
AGENTPR™ is an analyst-augmentation system, not an autonomous decision-maker. The platform assigns a human-review workflow status when any of the following are detected:
- Low overall confidence on a key claim
- Sarcasm or irony risk in sentiment classification
- Cultural-nuance ambiguity (idiom, code-switching, region-specific framing)
- Crisis or reputation-risk indicators above threshold
- Conflicting signals between sentiment, emotion, and intent classifiers
- Entity-identity uncertainty
- Insufficient or contradictory evidence
Standard Review applies to routine outputs. Human Review Recommended calls for focused professional assessment before consequential use. Human Review Required blocks publication or external action until an authorised reviewer approves or rejects the output.
Responsible AI in practice
GASKI election intelligence
In the GASKI election-intelligence implementation, AI-assisted polling-unit sheet transcription is separated from publication authority. Machine-read records carry task-level confidence scores and remain pending until an authorised human reviewer verifies or rejects them against the original INEC sheet. Discrepancy flags indicate records requiring review; they are not presented as proof of wrongdoing.
GASKI reports figures attributed to INEC and does not declare election results.
4.
Data — what we collect, how long we keep it, who has access
Collect
Account details, the brand briefs you run, the public sources we read on your behalf, and basic usage telemetry needed to run the platform.
Retain
Brief outputs are retained for the life of your subscription so you can revisit history. You can request deletion at any time.
Access
Only you and authorised members of your workspace. Internal access is role-gated, logged, and limited to support and abuse investigations.
Full detail — including legal basis, regional storage, and your rights — lives in our Privacy Policy.
5.
Third-party AI — Claude (Anthropic) powers our analysis layer
AGENTPR™'s analysis and synthesis agents (ZALI and NZE) are powered by Claude, built by Anthropic. We selected Claude for its safety posture, long-context reasoning, and suitability for high-stakes reputation work.
Customer data sent to Claude is processed under Anthropic's commercial terms and is not used to train their models.
6.
Feedback — how to report concerns or incorrect outputs
If you spot a brief that's wrong, biased, or unsafe, tell us. Every report is reviewed by a human and used to improve the methodology and guardrails.
Email support@useagentpr.com with the subject "Responsible AI feedback"
Please include the brief ID (if available), the specific claim, and what you believe is incorrect.
AI Output Contestation
You believe a specific score or finding about your organisation is incorrect.
Email: support@useagentpr.com — subject "AI Output Contestation"
Response: Human review within 14 business days
Data Protection
Exercise a data right or report a privacy concern.
Email: support@useagentpr.com — subject "Privacy / Data Protection"
Response: 30 days. Breaches reported to NDPC within 72 hours.
7.
Our Glassbox AI Policy
Glassbox AI is the principle that any AI system shaping public narrative must be inspectable — sources, methodology, and decisions visible end to end.
Glassbox AI Policy — Published
The full policy documents seven principles: the Glassbox principle, data handling, how our AI makes decisions, analytical assurance, prohibited uses, your rights, and how to raise a concern.
Accountability
- Accountable Executive
- Dr. Celestine N. Achi — Founder & CEO, Cihan Media Communications
- Technical Co-owner
- Orimolade Oluwamuyemi, FIIM — Strategic Technology Adviser
- Relevant Nigerian Data Protection Regulator
- Nigeria Data Protection Commission (NDPC)
- Assessment Baseline
- v1.0 — June 2026
This identifies the relevant national regulatory authority. It does not imply that the NDPC has approved, certified or endorsed AGENTPR™.
Meaning Intelligence™ research status
The initial research preprint, The Register Gap: A Meaning Intelligence Framework for Nigerian Public Discourse, was published in June 2026. Expanded corpus development, multi-model validation, analyst calibration and continuous benchmark evaluation remain ongoing.
Read the preprintHow does your AI deployment score?
Run Free RAI Assessment