The Philippine reference is specific, not a blanket approval
On 21 May 2024, the NPC issued Advisory Opinion No. 2024-005 on a proposed tool for analyzing calls and emails, scoring employees and identifying coaching opportunities at a Philippine BPO subsidiary. It treated service improvement as a possible legitimate interest while emphasizing necessity, proportionality and employee information. The opinion was limited to the facts submitted and did not adjudicate the parties’ rights.[1]
This is an operating note based on existing guidance, rechecked on 15 October 2026; it does not report a new rule. The workflow below is Kooya’s recommended design, not an NPC-approved implementation or a finding that any particular deployment is lawful. A BPO should assess its actual client arrangements, data and intended employment uses with its responsible advisers.
Define what the score is allowed to mean
Start with three separate outputs. An event detector might flag that a required explanation was not found. A quality reviewer might judge whether the explanation was understandable in context. A manager might decide what coaching follows. Those are different tasks with different evidence requirements.
For example, a hypothetical booking-support team wants to check whether an agent explained a cancellation condition. The model should identify the relevant segment, show the applicable rubric and mark missing or unclear evidence. It should not turn an uncertain transcript into a confident statement about the employee’s competence.
Kooya recommends starting with coaching support and excluding automatic pay, promotion or disciplinary decisions from the first deployment. That is a scope choice for a controlled trial, not a claim that human involvement alone resolves privacy or employment obligations.
- Name the exact behavior: for example, confirming the customer’s requested change.
- Define when the behavior is applicable and when it is not.
- Include an insufficient-evidence outcome instead of forcing every call into pass or fail.
- State which downstream decisions the score must not trigger.
Assess employee and customer data as separate concerns
The Data Privacy Act distinguishes lawful criteria for personal information in Section 12 from the rules for sensitive personal information in Section 13. A recording can contain several kinds of information, including health details or government identifiers. Do not assume a general legitimate-interest argument settles the basis for every item in a call.[2]
Map the route from recording to transcription, scoring, supervisor dashboard and any external model provider. Identify whose data each stage processes and which entity determines the purpose. Removing names from a transcript does not by itself establish that the recording or remaining context is anonymous.
As an implementation step, give privacy and security owners a concrete sample data-flow diagram and the proposed access and retention settings. Review customer information as well as employee information. A coaching dashboard should not expose entire customer records simply because the underlying support platform stores them.
- Restrict the reviewer to the evidence needed for the assigned coaching task.
- Document retention for recordings, transcripts, score explanations and dispute records separately.
- Identify any provider use beyond delivering the contracted analysis.
- Check that deletion and access changes reach derived records, not only the original audio.
Explain the workflow before introducing the dashboard
The NPC opinion emphasizes informing employees of the relevant policy before using the tool. It also discusses objections to processing based on consent or legitimate interest: continued processing requires the controller to establish and communicate an appropriate lawful basis or compelling reason. It does not give employers unrestricted discretion to dismiss objections.[1]
For the operating workflow, explain what the system evaluates, who can see results, how supervisors use them and where employees can raise concerns. Show an example of a corrected score and an example the system cannot assess. This makes the review process more concrete than an assurance that a person remains involved.
Keep the operational score-dispute route connected to, but distinct from, the privacy-rights process. A supervisor can correct a missed phrase without deciding a legal objection. The responsible privacy team should handle rights requests under the applicable requirements.
Test the transcript before testing the judgment
NIST’s published AI RMF 1.0 calls for measuring performance under conditions similar to deployment and evaluating fairness and bias. It is voluntary guidance. NIST currently notes that a revision is in progress; this note uses the published 1.0 framework as an evaluation reference.[3]
Build a reviewed sample from the languages, accents, channels and call types actually in scope. For a Philippine operation, that may include English, Filipino, mixed-language exchanges or regional-language material—but test the real mix rather than assuming every queue is multilingual. Include interruptions, low-volume speech and poor recordings when those conditions occur.
Separate two checks. First, did the transcript preserve the relevant words and identify the speaker correctly? Second, did the scoring logic apply the rubric correctly to that evidence? Record how often an audio or transcription problem becomes a false coaching flag. Aggregate agreement can conceal the calls that most need a human review.
Use enough examples to understand each condition you intend to support, and show the sample counts alongside results. A handful of difficult calls can reveal a failure mode; it cannot establish dependable performance across an entire workforce.
Make a disputed score easy to reconstruct
For each proposed flag, show the call segment, transcript excerpt, rubric version, model or configuration version and reviewer decision. If the model cannot point to supporting evidence, mark the flag unresolved. Avoid displaying an unexplained confidence number as a substitute for evidence.
In the booking example, imagine the customer interrupts just as the agent explains the condition. A transcript may omit the phrase or assign it to the wrong speaker. The supervisor should be able to listen, correct the evidence, change the score and record why. That correction should reach any coaching summary derived from the original result.
Give employees a defined route and response expectation for operational disputes. Limit access to the people who need it, and record unresolved cases separately from confirmed findings. Do not rank employees on unresolved flags while presenting those ranks as established performance.
Calibrate the rubric with people who do the work
Ask two qualified reviewers to score the same sample independently before treating their labels as a reference. Discuss disagreements: they may reveal an ambiguous rule rather than a model failure. Keep the resolved examples with the rubric so a later shift can understand the decision.
Avoid using inferred emotion as the sole evidence for courtesy or empathy. Prefer observable behavior tied to the service objective, such as acknowledging a stated concern or accurately explaining an available option. This is Kooya’s rubric recommendation; it makes the feedback easier to challenge and act on.
A useful coaching result names the behavior, supplies the evidence and proposes a practical improvement. A vague personality label gives the employee little to correct and the reviewer little to verify.
Run a shadow trial before scores affect performance records
Begin with a shadow workflow: authorized reviewers compare proposed flags with their ordinary review process, while AI results remain outside consequential performance decisions. Define the trial boundary, owner and stop conditions before it begins. The purpose is to learn whether the extra analysis improves coaching without creating an unmanageable correction queue.
Track false flags, missed applicable events, unresolved evidence, reviewer disagreements, corrections and review effort. Report results by the conditions you tested. Include time spent resolving errors when assessing any efficiency benefit; this article makes no claim about measured savings.
Pause or narrow the use if the tool repeatedly penalizes a language condition, loses supporting evidence or cannot propagate corrections. Recheck after changes to the model, transcription service or rubric. Broader use should follow a new decision about evidence and consequences, rather than happen because a dashboard makes ranking easy.
Use the AI pilot evaluation checklist to record the evidence and stop conditions for this bounded trial.