Once personal data enters an AI, you rarely get it back.
It takes seconds, with no hacker and no alarm. A customer's personal data goes to an AI tool and an answer comes back. The data may stay behind, and months later no one can say where. For a bank, NBFC or fintech, that breaks customer trust first and becomes a regulatory problem second.
Three reasons this cannot wait
Gone is gone
Text sent to an AI tool or vendor may be stored, logged or reused. Deletion is hard to confirm, and you cannot call back what has already left.
Penalties in crores
India's DPDP Act allows penalties that can run into hundreds of crores of rupees, for failing to take reasonable security safeguards or to notify a breach.
72 hours
The DPDP Rules expect detailed breach information to reach the Data Protection Board within 72 hours of becoming aware. You cannot report what you cannot trace.
For a regulated institution, the first test is often a question, not a penalty: show us where the data went.
Penalties are statutory maximums, decided case by case after an inquiry. Not legal advice. Sources
Digital Personal Data Protection Act, 2023 (No. 22 of 2023), sections 8(5) and 8(6) and the Schedule, for the penalty ceilings. Digital Personal Data Protection Rules, 2025 (G.S.R. 846(E)), rule 7(2), for the 72-hour breach report.
That is why we are building VeriFlow.
VeriFlow is being built to keep a sealed record of the journeys customer data takes through AI tools and vendors.
Know which vendors and AI tools received customer data.
A sealed record an auditor can check on their own.
Evidence of the controls that acted, not only a written policy.
And it tells you plainly what it could not see.
An evidence layer, not another gateway
You already run tools that act on customer data. VeriFlow is designed to sit beside them: they do the work, and it keeps a record of what they did that an auditor can check.
They detect, redact, block or process.
Sent by API. No raw customer data needed.
Which flows exist, what happened, what is unseen.
A sealed export, read against a rulebook.
VeriFlow is
- A register of AI and vendor data flows
- A sealed record of what your controls did
- A coverage report that lists its own blind spots
- Evidence mapped to the clauses of a rulebook
VeriFlow is not
- A gateway or proxy for your traffic
- A detection or DLP engine
- A replacement for your GRC tool
- A compliance certificate or legal advice
The email nobody wants on a Friday
Which vendors received customer call transcripts last quarter, and can you show us?
It sounds like one question. The answer lives in contact-centre tools, a document reader, three AI assistants and a vendor list that was last updated in spring.
Two ways it can go
Chasing
- Emails to six teams and four vendors
- A spreadsheet that is out of date by Monday
- "We do not train on your data" quoted as if it were proof
- A list with a footnote saying "we think"
Opening the record
- Filter journeys by vendor, purpose and quarter
- See what was redacted, blocked or sent as is
- Hand over a sealed export the auditor can check alone
- A plain list of what the record could not see
Try to rewrite history
Each entry is sealed together with the seal before it. Change one entry after the fact and every seal after it stops matching. Try it below. These are sample entries, and the seals are computed live in your browser.
The auditor does not need our platform to run this check. The checking tool is being built to be separate and open.
We tell you what we cannot see
A record that only shows good news is not worth much. The report is built to list blind spots next to the journeys it covers, so a gap lowers the coverage figure instead of hiding in it.
A silent source
A tool that stops reporting is flagged, not assumed fine.
An unlisted AI tool
Traffic to an AI service that is not on the register shows up as unknown.
Promises, graded
One scale everywhere, including the sample report: unsupported, then self-declared, then document-backed, then independently verified. Only the last needs an outside check. These are sample promises.
Seen, Stopped, Vouched
Promises are graded above. These three words grade the journey itself: how much we know about what happened to the data.
We observed the journey and recorded it. We did not control it.
A control acted on the journey, such as redacting or blocking, and left a record of the decision.
An independent party checked the evidence or the control. It is the same idea as an independently verified promise.
Reads the local rulebook
The record keeps facts. Regulatory profiles read those facts against specific clauses, so the same record can serve different rulebooks and sectors. India comes first.
DPDP Act and Rules
Evidence about security safeguards, breach assessment, processor accountability and who received a person's data.
RBI data governance
Evidence about third-party sharing, lineage, retention and audit. The RBI text is a draft, so we treat it as one.
Other regimes later
New profiles are data files, not rewrites. We plan to test the design against a second jurisdiction before it is fixed.
RBI material is a consultation draft, so paragraph numbers may change. Sources
Digital Personal Data Protection Act, 2023 and Rules, 2025. Reserve Bank of India, Guidance on Regulatory Expectations for Data Governance, a draft circulated for consultation (July 2026), paragraphs 60 to 63 on third-party data sharing. Paragraph numbers may change in the final text.
We describe a report as evidence mapped to a clause. Whether an institution complies is a legal conclusion for its counsel and regulator. This is not legal advice.
Everyone gets asked a different question
One record, four readers. These are typical questions, paraphrased, not quotes from customers. The answers describe what we are building towards.
Which AI tools can reach customer data, and are they all on our list?
Coverage view with unregistered AI endpoints and silent sources flagged.
Can we show which vendors received personal data, and why?
Journeys by vendor, purpose and period, read against the rulebook.
Does the vendor really do what the contract says?
Each promise graded from unsupported to independently verified.
Can we check this without relying on the team that built it?
A sealed export and a checking tool that does not depend on VeriFlow.
What your risk team will ask
Straight answers, each with its status. Designed means decided in our design, not yet built. Planned means intended, not final. Open means we want your input.
What does VeriFlow itself store?
References and hashes, not names, account numbers or transcripts. A person appears as a keyed reference, so the record and the report hold no raw customer data.
Where does it run?
Planned: India-region hosting, separate keys per customer, and a copy of the seals held by you. Whether it runs as a service or inside your environment is something we want to settle with design partners.
Can we review it as a vendor?
A pilot pack for your vendor-risk team: a deployment guide, a description of how data is processed, and a list of known limits, before any pilot starts.
Which AI does it cover?
AI services and vendors that your gateways, proxies, DLP tools or applications report: model APIs, assistants, document readers, vendor tools with AI features. A flow nobody reports shows as a blind spot. Staff use of public chatbots is covered only if one of your tools reports it.
How does an auditor check it?
Three steps with a separate, open tool: recompute the seals, check the dated anchors (an independent timestamp plus a copy you hold), and look for gaps in the sequence.
What if the rules change?
Rulebooks are versioned data files. The RBI text is a draft, and we re-check the wording before every pilot.
We are young, so we show our working
VeriFlow is early. Instead of asking for faith, we commit to how we build it.
You can check our work
The checking tool is being built to be separate and open, so an auditor does not rely on our word.
Plain wording
Reports say "evidence mapped to a clause". They never say "compliant" or "zero exposure".
Your data stays out
The record is designed to hold references and hashes, not raw customer data. Our research asks for no customer data and nothing under NDA.
Build it with us
VeriFlow is being built with design partners. We are talking to people at banks, NBFCs and fintechs about the last time someone asked them to show where customer data went.
Start with a research interview
About 10 minutes, mostly tap-to-choose. Your answers shape the first use case and the report format. Leave out customer data and anything under NDA.
Take the interviewDesign partners get a say in the first use case and a first look at the report format. We will not describe you as a customer without your written agreement.
Questions, or want to talk first?
Write to us about the research pilot and what you would like to try. Please do not send customer data or documents under NDA by email.
veritastech@zohomail.in