CUSTOMER SERVICE: The JANUS Framework

GUARDIAN OF DIGITAL GATEWAYS

Introducing:
The JANUS Framework™

The same system scores both sides of the gate.

This sentry is named for Janus, the Roman mythological god of doorways, transitions, and duality, who is famously depicted with two faces looking in opposite directions at once. Within the portfolio, he is unique: he is the only sentry named after a Roman mythological god, and the only one whose harm flows in two directions simultaneously.

Janus was the Roman god of beginnings and endings, of doorways, gates, and passages. He was invoked at the start of every new undertaking and at the close of every completed one. He is depicted with two faces, one looking forward and one looking back, or one looking inward and one looking outward, depending on the tradition.

He was one of the most important gods in the Roman pantheon. His temple in the Forum had doors that stood open in times of war and closed in times of peace. He was the guardian of the threshold, the one who watched both sides of the gate at once.

He is not a historical person. He is a principle given a name. And the principle is exactly what this sentry governs: any system that mediates an interaction between two parties must be accountable to both of them.

His legacy demonstrates that a gate is not a wall. It is a threshold. And a threshold has two sides.

THE ALGORITHMIC THREAT

A customer service algorithm does not arrive in a call center as a disciplinary tool. It arrives as a quality score. The customer on one end is routed, deflected, and surveyed. The agent on the other end is scored, ranked, and coached. The same operational framework evaluates both. The same framework fails both. No one can see it.

The same reality governs sentiment analysis, escalation logic, chatbot deflection, call routing, and post-call quality assurance. When the logic remains buried within a black-box model, the harm runs in two directions. The customer cannot explain why they were sent in circles. The agent cannot explain why their performance baseline dropped. The regulator has no visible violation to cite. The company has no explicit liability on the books.

Yet the liability is real. The downstream systemic harm is now documented.

Sentiment analysis models calibrated on standard speech profiles systematically misread accents, regional dialects, and cadence patterns that fall outside the initial training distribution. The customer who speaks differently is flagged as hostile. The agent who handled the call correctly is penalized for an escalation or a frustration they did not cause.

Quality scoring tools trained on historical performance data replicate the evaluation anomalies that have historically advantaged certain teams and marginalized others. Agents with accents receive lower quality scores for identical work. Agents who code-switch are rated as less professional. Agents whose demographic profile matches the customer's are consistently rated as more personable.

Algorithmic routing structures systematically send specific callers to longer queues, fewer escalations, and less experienced agents. The pattern correlates directly with zip code, language preference, and account history. The enterprise cannot defend the routing logic because no one documented the underlying data provenance or model parameters.

In extreme environments, these tools go beyond scoring. Emotion recognition systems deployed in call centers attempt to infer employee internal emotional states directly from biometric vocal frequencies. The EU AI Act Article 5(1)(f) bans this practice outright. It prohibits the placement on the market or use of AI systems to infer emotions within the workplace and educational institutions. This statutory boundary is a strict prohibition rather than an abstract compliance recommendation.

The prohibition is absolute for employees. It does not extend to customers. The same voice-analytics engine that is prohibited when pointed at an agent is permitted, with strict safeguards, when pointed at a caller. The law draws the line by population, not by technology.

That distinction is what makes this sector uniquely exposed. The same platform can be compliant in one direction and illegal in the other. The company that does not map both directions is running a prohibited system and a high-risk system on the same infrastructure.

The systemic failure is structural. Customer service models that share training pipelines, weights, or third-party dependencies with widely deployed foundation models propagate baked-in bias across entire industries. A single vendor's unverified quality scoring engine silently becomes an entire industry's evaluation standard.

The customer has no visibility into the score. The agent has no mechanism to audit the weights. The enterprise has no forensically defensible paper trail to prove the model was governed.

Until now.

THE JANUS FRAMEWORK SERVICES

WHAT WENT IN

1. Data Provenance Audit

We document the provenance of the training data behind every sentiment analysis, quality scoring, routing, and emotion recognition model. We verify every data point at the source, trace what the model encoded at the weight level, and produce a baseline record that satisfies the evidentiary standards of FTC Act Section 5 claims, state consumer protection actions, Title VII claims, and EU AI Act compliance.

2. Synaptic Silencing Audit

Before you can govern a customer service model, you must identify what needs to be silenced. We document what the model carries at the weight level before it reaches a caller or an agent. We cross-reference the developer's AB 2013 disclosure, record the fine-tuning history, and produce the pre-deployment baseline record.

WHAT IT DOES

3. Sentiment Analysis Bias Audit

We audit the sentiment and tone detection models used to score customer interactions. We verify that the model does not systematically misread accents, regional dialects, and cadence patterns that fall outside the initial training distribution. We document the pattern when the model penalizes the agent for a sentiment it misclassified.

4. Quality Scoring Audit

We audit the models that generate agent quality scores. We verify that the criteria are job-related, that the model does not default to highest-weight historical patterns, and that the training data does not contain proxies for race, national origin, language, or gender. When a scoring system is found to be discriminatory, we document the finding in a litigation-ready format.

5. Routing and Escalation Audit

We audit the models that determine which callers reach which agents, how long they wait, and whether their issue is escalated. We verify that routing is based on actual service needs, not on account history, zip code, or language preference proxies. We document the pattern when the model systematically deflects certain callers.

6. Emotion Recognition Prohibition Audit

We audit every system that identifies or infers emotion from biometric data. We distinguish between three categories: employee-facing emotion recognition, which is prohibited under EU AI Act Article 5(1)(f); customer-facing emotion recognition, which is permitted; and text-based sentiment analysis, which falls outside the prohibition entirely. We verify that no employee-facing system is inferring emotional state from voice, facial expression, or physiological signal. Where the line is blurred, we map the system architecture to the legal standard.

7. Voice Biometric Consent Audit

Where emotion recognition is permitted, we audit the consent architecture behind every voiceprint captured. We verify that the collection complies with Illinois BIPA, the Texas Capture or Use of Biometric Identifier Act, and the Washington My Health My Data Act. We document whether consent was obtained, whether the data is retained and deleted according to the law, and whether the customer was informed.

8. Sentiment Analysis Distinction Audit

We audit the vendor's classification of its own product. A tool branded as "sentiment analysis" may in fact be emotion recognition if it uses vocal features to infer anger, stress, or vulnerability. We verify that the system's actual capabilities match its marketing claims. When a tool crosses the line from text sentiment into biometric inference, we document the crossover.

9. EU AI Act Article 5(1)(f) Compliance Audit

For companies with European operations or European customers, we audit the entire customer service stack against Article 5(1)(f). We verify that no employee-facing system is using biometric data to infer emotion. We verify that the workplace definition is applied correctly, including to contractors, trainees, volunteers, and remote workers. We verify that the recruitment and selection process is compliant.

We also audit the customer-facing systems against the high-risk classification under Annex III. We document the transparency disclosure architecture, the registration obligations, and the conformity assessment timeline.

The employee-facing prohibition applies now. The high-risk obligations apply from December 2027. The two timelines are mapped separately.

10. Inference Safeguard Validation

We validate that human review requirements for final termination, discipline, and quality-based pay decisions are actually operational. The system should hold irreversible decisions for human review while allowing routine scoring to proceed. We document whether the verification layer exists, whether it functions, and whether it was bypassed.

HOW FAR IT REACHES

11. Systemic Correlation Assessment

We measure whether the customer service model shares training data, architecture, or third-party dependencies with widely deployed foundation models. We track behavioral drift across customer cohorts and agent cohorts and generate the substantial-modification evaluation.

12. Vendor Disclosure and Accountability

We map the algorithmic supply chain behind the company's customer service stack. We demand vendor disclosure of model logic, training data provenance, and fairness audits. Where vendors refuse disclosure, we document the refusal as evidence of concealment.

WHAT WE DO ABOUT IT

13. Proactive Compliance Architecture

For companies that want to govern the system before the FTC or the state Attorney General does.

We establish a Customer Service Algorithmic Integrity Protocol that governs the design, deployment, and monitoring of sentiment, quality, routing, and emotion recognition algorithms. This includes pre-deployment bias testing against both customer and agent populations, ongoing model drift monitoring, customer and agent grievance pathways that bypass the algorithm, mandatory human-in-the-loop requirements for all high-stakes decisions, and FTC, EEOC, BIPA, and EU AI Act readiness.

14. The Janus Redress Initiative

For customers, agents, and advocacy organizations.

We provide a Probabilistic Harm Audit that uses econometric counterfactual baselines to isolate the Algorithmic Increment of Harm in a specific routing, scoring, or termination decision. This quantifies the harm for consumer protection claims, employment discrimination claims, and impact litigation.

THE FRUIT OF THE POISONOUS MERKLE TREE™

The remedy doctrine. What happens when the same system fails both sides.

A customer service algorithm does not produce a single bad interaction. It produces a tree of downstream consequences on both sides of the gate. The misread sentiment leads to the low score. The low score leads to the coaching. The coaching leads to the discipline. The discipline leads to the termination. The termination leads to the lost income. The lost income leads to the lost home. On the customer's side, the deflection leads to the unresolved issue. The unresolved issue leads to the escalation. The escalation leads to the complaint. The complaint leads to the regulator. The regulator leads to the fine.

Every branch of that tree is tainted by the original algorithmic error. And every branch is still bearing fruit on both sides.

The Fruit of the Poisonous Merkle Tree™ treats algorithmic taint as a traceable structural condition. The doctrine provides a five-part remedy that goes beyond monetary damages.

Algorithmic Disgorgement. The company must surrender the outputs it derived from the tainted model. Not just the algorithm. Every score, every routing decision, every quality determination, every termination that flowed from it. The taint travels. The disgorgement follows the taint.

Forced Structural Retraining. The company must rebuild the system that produced the harm. Not patch it. Not audit it. Rebuild it. The training corpus, the scoring thresholds, the routing logic, the human review pathway. The correction is not a memo. It is a new architecture.

Downstream Notification. Every customer who was misrouted and every agent who was mis-scored must be notified. The agent who was terminated knows. The agent who was quietly denied a raise does not. The customer whose complaint was deflected into a loop does not. The doctrine requires the company to find them and tell them.

Supply-Chain Traceability. Every downstream system that ingested the tainted output must be identified. The workforce management platform. The payroll system. The performance review tool. The customer relationship management system. The taint travels. The remedy maps where it went.

Harm Apportionment. The harm is not the company's alone. The vendor built the model. The integrator deployed it. The company operated it. The customer and the agent both bore the cost. Every party in the chain is assigned a share of the remedy proportionate to the role it played. This is where the Right to Be Probable™ does its work. The Algorithmic Increment of Harm isolates what each party contributed.

The doctrine is not punitive. It is structural. The taint travels. The remedy follows. And under JANUS, it follows in both directions.

THE JANUS LEGAL FRAMEWORK

FTC Act Section 5. The federal statute that prohibits unfair or deceptive acts or practices in commerce. The anchor for claims that algorithmic routing or deflection constitutes a deceptive practice.

Title VII of the Civil Rights Act. The federal statute that prohibits employment discrimination. The anchor for agent-side claims that quality scoring, discipline, and termination were discriminatory.

EU AI Act Article 5(1)(f). The prohibition on AI systems that infer emotions from biometric data in the workplace and in education. In effect since February 2, 2025. Applies to employees, contractors, trainees, volunteers, and applicants. No consent exception. Narrow exceptions for medical and safety uses only.

EU AI Act Annex III, Point 1(c). The classification of emotion recognition systems as high-risk when used on customers. Transparency obligations apply from August 2, 2026. Full high-risk compliance obligations apply from December 2, 2027.

EU AI Act Article 50(3). The transparency requirement for deployers of emotion recognition systems. The deployer must inform the natural persons exposed to the system. The disclosure must be clear and accessible.

The Illinois Biometric Information Privacy Act. The state statute that requires informed consent for the collection, storage, and use of biometric identifiers, including voiceprints and facial scans. The template for similar statutes in Texas and Washington.

State Two-Party Consent Laws. The state statutes that require all parties to consent to the recording of a telephone call. California, Illinois, Pennsylvania, and others.

State Consumer Protection Statutes. The state analogues to FTC Act Section 5, which allow state attorneys general and private plaintiffs to bring claims for unfair or deceptive practices.

The Fruit of the Poisonous Merkle Tree™. The remedy doctrine. Five-part structural remedy for algorithmic taint in customer service.

Regulatory Judo™. The disclosure engine. The right to know whether AI was involved in the decision.

Eight instruments. One record.

WHO THIS SERVES

The JANUS Framework serves the entire customer service ecosystem. Companies that want to build defensible systems. Customers and agents that want to hold those systems accountable. And the attorneys, regulators, and advocates who bring the cases to a legal forum.

For Companies and Contact Centers

Chief Customer Officers. Contact Center Directors. Quality Assurance Leaders. Workforce Management Teams. General Counsel. Compliance Officers. BPO Leaders.

If you want to audit your sentiment, quality, routing, and emotion recognition systems before the FTC or the state Attorney General finds them, we apply the DigitalRAS forensic engine, the Sentiment Analysis Bias Audit, and the Quality Scoring Audit to identify and remediate exposure before it becomes a liability.

For CX Tech and AI Developers

Conversation Analytics Vendors. Sentiment Analysis Developers. Quality Management Platform Builders. Emotion Recognition Developers. Product Counsel. Regulatory Affairs Teams.

If you are building the models that contact centers deploy, you carry obligations under FTC Act Section 5, Title VII, and EU AI Act Article 5 before your model ever reaches a call. We built the weight-level governance architecture that makes your model defensible at the point of sale and at the point of subpoena.

For Customers, Agents, and Advocates

Customers Misrouted or Deflected. Agents Mis-Scored or Terminated. Call Center Workers. Customer Service Unions. Consumer Protection Organizations. Worker Advocacy Organizations. State and Federal Regulators.

If your routing, your score, or your termination was determined by a system you were never allowed to see, you may have a claim. The record already exists. Let us read it.

For Consumer Protection and Employment Attorneys

We serve as co-counsel and forensic support on FTC Act, state consumer protection, Title VII, and BIPA cases that turn on what the algorithm did and how the record proves it.

THE JANUS ECOSYSTEM

The JANUS Framework is the two-way customer service configuration of DigitalRAS. It is one sentry in a coordinated system of algorithmic accountability. Two related engines extend its reach across the customer and worker lifecycle.

THE HATTIE TAKE™

The Creative & Entertainment Sentry

Where JANUS governs the platform that delivers the service, HATTIE governs the platform that replicates the performer. Both operate at the interface between the institution and the individual. Both are audited by the same engine.

→ Explore The HATTIE Take™

THE MARTHA INQUIRY™

The Media & Journalism Sentry

Where JANUS audits the algorithm that scores the interaction, MARTHA audits the algorithm that determines what the public sees. Both govern the systems that mediate between institutions and the people they serve.

→ Explore The MARTHA Inquiry™

THE WALLACE INITIATIVE™

The Employment Sentry

Where JANUS audits the quality score that determines discipline, WALLACE audits the hiring model that determines who gets the job. The agent's performance score and the applicant's screening score are produced by the same family of systems.

→ Explore The WALLACE Initiative™

REGULATORY JUDO™

The Disclosure Request Engine

LAUNCHING SOON

Both the customer and the agent have the right to know whether AI was involved in the routing, the scoring, or the decision. Regulatory Judo generates the legally compliant disclosure request and builds the record that feeds the pattern that feeds the case.

→ Join the Regulatory Judo Waitlist

THE CULTURE & INFORMATION CLUSTER

HATTIE. MARTHA. JANUS.

Three sentries. One cluster. One record.

WHY THIS IS 1 OF 1

Most governance programs treat AI risk as a policy problem. They produce frameworks, checklists, and statements of principle. The policies exist. The training exists. The safeguards exist. None of it is enough.

The Architecture treats AI risk as an architecture problem. It governs the model, the process, and the record simultaneously. It accounts for both the technical failure and the human oversight that permits it.

System failures require a multi-layered explanation. The technical breakdown pairs with automation bias. It compounds with siloed review. It feeds on a lack of imagination regarding system vulnerabilities. It thrives under an incentive structure that prioritizes speed over verification.

The model fails because it executes inside an operational framework that was never built to catch it.

The Architecture is built at the weight level. It produces a record that is forensically defensible. It is designed to be tested.

No other legal framework currently combines sentiment analysis forensic auditing with FTC Act enforcement, quality scoring review with Title VII protection, routing and escalation auditing with consumer protection, and biometric emotion recognition auditing with EU AI Act Article 5(1)(f) compliance.

No other legal framework audits both sides of the interaction. JANUS audits the agent side against the Article 5(1)(f) prohibition, audits the customer side against the permitted use standard, and audits the vendor's own classification to verify which side the system is actually on.

No other legal framework audits both sides of the call on separate compliance timelines. JANUS audits the employee-facing prohibition, which is already in force, and the customer-facing high-risk obligations, which are on a December 2027 runway. The immediate risk and the future risk are mapped separately.

No other legal framework combines weight-level customer service forensics with a dedicated two-sided practice. JANUS produces the record. Runtime Governance enforces it.

The JANUS Framework is not a generic AI ethics framework. It is a weaponized legal instrument built specifically for the customer service sector, where algorithmic scoring is already widespread, largely unregulated, and generating consumer protection and employment exposure at an accelerating rate.

It is the only sentry in the portfolio named for a god. It is the only one that governs two parties at once. It is the only one where the legal standard differs depending on which side of the call the system is analyzing.

The gate turns in two directions. So does the record.

Thirteen sentries. One engine. One record.

The regulators are already scaling. The JANUS Framework ensures you are not left behind.

The difference does not show up in a slide deck. It shows up in discovery.

THE RECORD SPEAKS FOR ITSELF.
WE MAKE SURE IT IS HEARD.

Every call. Every score. Every routing decision. Every termination. The algorithm generated a record of everything it did. Most of that record has never been examined. Most of it has never been challenged.

We examine it. We challenge it. We litigate it.

Button: Request a Confidential Consultation

SCHEDULE A CONSULTATION

TIANGAY KEMOKAI LAW, P.C.