← All news and guides
Data protection

DPDP Case Study: An AI Hiring Platform, From Audit to Production in Three Weeks

How we audited an AI hiring platform against the DPDP Act, found 25 gaps, and shipped consent, notice, erasure and staff-access fixes to production in three weeks.

Mono Team ·

A dark typographic card reading: DPDP case study. Audit to production in three weeks. An AI hiring platform.

Short answer: we audited an Indian AI hiring platform against the DPDP Act, found 25 gaps, and shipped the fixes to production in three weeks. Candidates now give call consent themselves and the platform can prove it. The privacy notice is written from what the code actually does. One routine finds, exports and erases a person everywhere their data lives, and staff have two-factor sign-in with every recording view logged.

This is the story of that work: what we found, the order we fixed it in, and what we'd tell another team starting the same job. The platform is not named. If you want the method behind it, see our DPDP gap assessment and DPDP implementation services.

What does the platform do, and why does DPDP matter so much for it?

The platform helps companies hire engineers in India. Candidates sign up, upload a CV, apply to roles, take proctored assessments on camera, and get AI screening calls. Recruiters also find candidates themselves and import them in bulk. Employers send in roles through an intake form.

That means a lot of personal data, much of it recorded, and most of it processed by AI providers outside India:

  • Candidate profiles, CVs and job applications
  • Camera, microphone and screen recordings from assessments, with AI scores
  • Phone call recordings, transcripts and AI summaries from screening calls
  • Profiles that recruiters found or imported, of people who never contacted the platform
  • Employer enquiries, newsletter sign-ups and support requests

Hiring data is high stakes for the people in it. A candidate who is recorded, scored by an AI and called without agreeing to it has every reason to complain, and under DPDP the burden of proving consent sits with the company, not the candidate.

What did the DPDP audit find?

The audit was a read-only review of the code, database migrations and configuration. We didn't touch production data. We traced each kind of personal data from where it is collected to every place it is stored or sent, then checked each step against the Act and the DPDP Rules 2025.

It found 25 gaps: six we rated P0 (likely non-compliant today, or a major exposure), nine P1 (must close before the core duties start on 13 May 2027) and ten P2 (hygiene). Some protections already existed: tables for call consent and do-not-call requests, a check before each call with quiet hours, an AI and recording disclosure at the start of calls, self-service account export and deletion, and row-level security on every table.

The gaps fell into five themes:

  • Nothing was ever deleted. Video, call audio, transcripts and AI verdicts were kept forever, and deleting a person left large pieces behind.
  • Notices were missing or wrong. The privacy policy was three sentences. The assessment room said recordings went only to the hiring team, while AI providers also processed them.
  • Consent was recorded by staff, not given by candidates. For screening calls, a recruiter picked the consent source from a dropdown.
  • AI providers were unconstrained. CVs and transcripts could be routed to any model provider, including ones that keep or train on what they receive.
  • Staff access was all or nothing. One staff password opened every candidate's data, and nothing recorded who watched which recording.

What did we fix first?

We started with anything exposing data right now, because those fixes were small and the risk was live. In the first week:

  • Analytics consent. Session replay ran on every public page, including the application forms, with no consent prompt. It now loads only after the visitor clicks Accept, never runs on forms, assessments or staff pages, and can be withdrawn from the footer.
  • AI data routing. Every AI request now goes only to zero-retention endpoints that do not train on inputs, with a test that checks the setting on every model the platform uses.
  • Logs. The voice service logged full phone numbers and the first words of each call. Phone numbers are now masked, and call events are cut down to what operations need. Assessment access links are scrubbed from error reports.
  • The matching opt-out. Onboarding showed a pre-ticked box to appear in job matching. Unticking it saved a setting that no code ever read, so candidates who opted out were still matched. It is now an unticked opt-out that works.

The last one is a pattern worth looking for in any product: a control that looks like it respects a choice but does nothing. It is worse than having no control, because the notice promises something the system doesn't do.

This was the biggest P0. Screening calls are AI-led and recorded, and telephony providers in India treat calls without consent as cold calls. Before the fix, a recruiter recorded consent on the candidate's behalf, and a bulk import stamped the import time as the moment of consent.

The fix had four parts:

  • Consent at application. The application form has an unticked, optional box. Ticking it records the time, the application as evidence, and the notice version shown. It covers calls about that one job only, and the system enforces that scope before every call.
  • Opt-in links for everyone else. Imported and recruiter-found candidates get a single-use link tied to one person, phone number and job. It expires in 30 days and is stored only as a hash. Saying "no thanks" adds the number to the do-not-call list.
  • Staff entries need proof. Recruiters can no longer pick the self-given sources. A manual or imported consent needs evidence and a real consent date, never the import time, and is flagged as staff-attested.
  • Easy to stop. Every call tells people they can say "remove me" at any time, and the opt-out detection understands Hindi and Hinglish phrasing as well as English.

We also hardened the form against abuse: consent only applies to the phone number stored on the cited application, suppressed numbers stay suppressed, and submissions are rate-limited.

How did we rewrite the privacy notice?

We wrote the new notice from the code, not from a template. For each flow we listed what is collected, why, the legal basis, which providers receive it and where they are, how long it is kept, and how to object. The result is a 13-section notice with a version number. A short notice line now appears at every point where data is collected, and the assessment rooms carry an accurate notice about camera and screen recording and AI processing.

The version matters as much as the text. Each consent and each assessment session stores the notice version the person saw, so if the notice changes, the platform can still show what someone agreed to.

How do you find, export and erase one person across every system?

Before the audit, a deletion request meant hand-written database queries across about 15 tables, and it still missed uploaded files and call recordings held by the telephony provider. Account deletion left job applications and their CVs behind, and some tables had no delete path at all.

We replaced all of that with one routine. Given an email, phone number or profile link, it finds every row, every stored file and every call recording linked to that person. It can export them as a single file, or erase them in a safe order: files first, then child records, then the account. It supports a dry run and reports any step that fails.

Staff run it from a console that shows what was found, lets them tick records that are possibly the same person, and needs the identifier typed again before anything is erased. Every search, export and erasure is logged with who ran it. A public form on the privacy page takes requests from anyone, with a 90-day due date set automatically.

Erasure keeps exactly three things, each for a documented reason: the do-not-call record, so the person is never called again; a minimised proof of the consent they gave; and the staff sign-in security log.

How did we tighten staff access?

Staff now have two-factor sign-in. Once a staff member enrols, it is checked on every admin page and every admin API route, and making it mandatory for every account is the final step. An append-only log records who played a recording, read a transcript, opened an assessment review or downloaded a CV, and admins can review it. Camera and screen recordings can only be read by staff, and a candidate's browser can only upload to its own open session.

Role-based limits inside the database, such as which staff roles may see recordings at all, are the next step and wait on a decision about who needs what.

How did we ship it without breaking a live product?

Each fix shipped as its own change, went through staging first, and was reviewed again before release. About half of all the commits were review fixes made after the first version, such as hardening the consent form against forged submissions and making sure self-service deletion could only ever touch the caller's own account.

Database changes were tested against a local copy of the database before they touched production, then checked again on production after release.

Summary: what we found, what we built, and the evidence

AreaWhat we foundWhat we builtEvidence
Call consentWhat we found: Recruiters recorded consent for candidatesWhat we built: Candidate-given consent scoped to one job, opt-in links, proof for staff entriesEvidence: A consent record per number with source, time, proof and notice version
Privacy noticeWhat we found: Three sentences; a false statement in assessmentsWhat we built: A versioned notice written from the code, shown at every collection pointEvidence: Notice version stored with each consent and assessment
AnalyticsWhat we found: Session replay on every page, no consentWhat we built: Loads only after Accept, never on forms, withdrawableEvidence: Tests that nothing is sent before consent
AI providersWhat we found: Data could go to providers that train on itWhat we built: Zero-retention, no-training routing on every requestEvidence: A test on every model tier
Access and erasureWhat we found: Manual queries that missed files and recordingsWhat we built: One routine to find, export and erase a personEvidence: A log of every request and who ran it
Staff accessWhat we found: One password opened everythingWhat we built: Two-factor sign-in and an access logEvidence: An access log admins can review
LogsWhat we found: Phone numbers and speech in logsWhat we built: Masking and scrubbingEvidence: Tests for each scrubber
Breach responseWhat we found: No planWhat we built: A runbook with notice templatesEvidence: The runbook

What would we tell another team starting DPDP work?

  • Map the data before you fix anything. Most of our findings came from following one kind of data end to end, not from a checklist.
  • Fix live exposure first. Analytics, AI routing and logs took days, and they were leaking data every hour they waited.
  • Make consent provable. If staff can record consent for someone, assume a regulator will ask how you know it is real.
  • Write the notice from the code. Every false sentence in a notice is a finding waiting to happen.
  • Build one erasure routine. Deletion scattered across features always misses something, usually files and vendor copies.

FAQ

How long does DPDP compliance work take for a product like this?

For this platform, three weeks from the start of the audit to the fixes running in production. The audit took the first days, the live exposures were fixed in the first week, and consent, notices, rights requests and staff security were built over the next two weeks.

When do the DPDP duties in this case study apply?

The core duties, including notice, consent, security safeguards, breach reporting and data principal rights, apply from 13 May 2027, eighteen months after the DPDP Rules were notified on 13 November 2025.

Does a hiring platform need consent to call candidates?

For AI-led, recorded screening calls, the safest basis is consent the candidate gives themselves, for a stated purpose, with proof you can produce. Under the DPDP Act the burden of proving consent is on the company, and telephony providers in India also treat calls without consent as cold calls.

Find out what DPDP changes for your platform.