Chief Data Officer / India
Keep the inconvenient population inside the account
- Four heads within the insurer
- 43 data colleagues
- ₹8.6 crore FY2026–27 expense; ₹2.4 crore separate improvement programme
- Field statistics to enterprise data leadership
- 29 years
- Five contiguous appointments from July 1997
- Company context
- 19 branches
- 920 employees and ₹1,860 crore FY26 gross written premium; not personal authority
Decisions & consequences
01Withdrawing the deterioration claim I attached to a correct total
In 2022 I approved a callback paper that described a falling combined rate as deterioration in the service. In the first period, 800 of 1,000 digital requests and 180 of 300 assisted requests were completed within their promised slot. The combined rate was 980 of 1,300, approximately 75.4%. In the second period the counts were 450 of 500 digital and 650 of 1,000 assisted requests, or 1,100 of 1,500 overall, approximately 73.3%. The aggregate had fallen. That alone did not explain the performance of either channel.
Pradeep Qelmavi asked for the conditional rates. Digital had moved from 80% to 90%, and assisted from 60% to 65%. The second period contained a much larger assisted share. The service sponsor preferred the single company headline because it corresponded to the actual combined experience and was easy to compare. That was a reasonable reason to retain the aggregate. It was not a reason to remove the changing composition or imply that both operations had worsened. I had approved a stronger interpretation than the report supported.
I authorised ₹52 lakh of reconstruction and reporting work. We retained the overall rate, both sets of counts and the channel rates, with the definition of the promised slot and the limits of the comparison. Operations retained responsibility for the callbacks and for examining staffing, demand and other explanations. The data account did not establish which intervention caused either rate to change, and it did not convert the missed promises into acceptable service merely because the conditional rates improved.
We gave up the simplicity of one directional judgement and withdrew my original heading. The delivered result was a clearer account of the two observed populations, not a claim that service had improved in every relevant respect. The first period contained 320 missed promises and the second 400; those absolute counts remained visible too. Next time I would require the composition review before approving the management interpretation, rather than accept reconciliation as sufficient evidence that the story attached to the total was also sound.
Refusing to fill missing turnover with an invented business declaration
A 2023 stewardship review examined 2,400 small-business files used in a defined underwriting-support workflow. Of those, 420 had no recorded annual-turnover declaration, ninety explicitly declared zero and 1,890 contained a positive amount. An analytical preparation rule proposed replacing the missing entries with a typical amount derived from other files. That would have made the field complete. It would also have created a value that the business had not supplied and made it easy for a later user to mistake an analytical assumption for a customer declaration.
The delivery lead wanted the filled dataset because the downstream comparison expected a numeric field. Underwriting colleagues asked instead whether the amount was supplied, confirmed or inferred. I supported preserving those distinctions. A declared zero was not the same state as absence, and neither state established that the declaration had been independently verified. The purpose of the field and the authority to act on it needed to remain with the responsible professional owners, not with a convenient preparation rule.
We spent ₹48 lakh on source review and a confirmation route. Of the 420 missing entries, 312 were subsequently confirmed through that route and 108 remained unresolved. The later values retained their confirmation status; the original absence was not rewritten as though the value had always been known. Underwriting owners kept an assisted route for the unresolved population and determined what further material was needed. My team did not issue an automatic decline, alter a premium or claim that a zero declaration was fraudulent.
The outcome was a less mechanically complete but more intelligible dataset. We accepted continuing work for the 108 unresolved cases and the delay involved in obtaining confirmation. I did not describe 312 confirmations as proof of universal data accuracy, nor claim that removing the proposed imputation improved loss experience. The cost bought a clearer relationship between a recorded statement and its source. I would specify that relationship at intake in a future workflow, before an empty numeric field creates pressure to solve a business uncertainty with a default value.
Withholding a table whose hidden value remained calculable
In 2024 a proposed partner benchmarking release grouped a defined set of insured business records into aggregate cells. The initial design hid small cells but retained exact row totals. In one synthetic review example, a row total of twenty-four sat beside visible counts of eleven and ten and one suppressed cell. A recipient could subtract the visible values and recover the hidden count of three. The symbol had concealed a display value without concealing the inference the rest of the table allowed.
The partnership sponsor wanted to issue the table on the agreed date because no names or individual rows appeared in it. I understood why the aggregate format seemed sufficiently distant from the source records. Pradeep's review showed why that impression was inadequate. We needed to examine the table together with its margins and related planned views, not approve each cell in isolation. Legal and privacy specialists retained the assessment of the actual release and its authorised use. The arithmetic example was not a claim that every small count identifies an individual.
I recommended withholding the proposed output and funded ₹34 lakh of redesign and review. The revised authorised release used coarser groupings and reduced breakdowns, with related outputs examined together. We did not keep an exact margin merely to preserve the original appearance when it undermined the chosen treatment. The partnership team accepted less analytical detail and a six-week delay. The underlying source remained subject to its own access and handling arrangements; a table redesign did not create permission for a different use.
The delivered result was a narrower reviewed release, not a universal assertion of anonymity or an assurance that all future combinations would be safe. We retained the output specification and the need to examine later additions. What I would change is the point of review. Disclosure questions should shape the proposed product before a commercial date is promised. The case reminded me that an apparently cautious presentation can still reveal its hidden value through an entirely ordinary calculation, and that data leadership includes examining those relationships before the format itself becomes the evidence for trust.
Declining a larger dataset after testing its usable contribution
In 2025 a supplier proposed an external business dataset at ₹1.6 crore a year. The commercial presentation emphasised record coverage and the number of populated fields. Our intended use was narrower: a defined stewardship workflow needed additional business attributes with suitable provenance and freshness. I agreed to a ₹27 lakh evaluation before recommending a recurring contract. The question was not whether the supplier possessed many records, but what new material could responsibly enter this particular workflow.
We reviewed 480 records against the agreed criteria. Seventy-two offered a newly supplied field beyond what our current account held. Of those, thirty-six met the current-use criteria for relevance, provenance and freshness, and thirty-six did not. The other 408 records added no new field for the defined comparison. These were findings for the reviewed set, not a representative estimate of value across the supplier's entire database. A new field also did not by itself establish a better underwriting or service decision.
The business sponsor preferred purchasing now and discovering additional uses later. The supplier's broader coverage might indeed have supported another purpose. I recommended declining this contract because the evidence assembled for the approved purpose did not justify the proposed recurring commitment. CEO Saira Dazquri accepted that recommendation while leaving future proposals open to their own purpose and review. Underwriting and legal owners retained decisions about appropriate use; my team did not turn a favourable field count into permission or professional authority.
The ₹27 lakh evaluation remained an incurred cost. The ₹1.6 crore annual figure was a declined proposal, not cash already spent and recovered or a quantified saving against an approved equivalent service. We also retained the unmet stewardship need rather than describe the decision as completion of the wider programme. I would agree an evaluation's decision criteria even earlier next time, before a supplier demonstration makes abundance feel like a business case. A data leader should be able to explain why a useful-looking purchase was not useful enough for the actual question, without declaring the material worthless for everyone else.