Chief Data Officer / India
The clearer map needed more distinctions, not fewer
- Data employees
- 47
- Four heads; nine contractors separately counted
- Annual data-function expense
- ₹11.8 crore
- Separate from insurer premium and professional approval authority
- Survey and insurance-data experience
- 32 years
- Source evidence, geographic versions and location uncertainty
Decisions & consequences
012018: the same marker concealed three different location standards
I approved a property-location view that represented 180 records with an identical precise point symbol. The interface was intended to support building-level screening. A source review led by stewardship colleagues distinguished 120 survey-verified positions, forty-two postal-area centres and eighteen positions whose provenance had not been established. All records had something that could be plotted, but the evidence did not support treating them as equally accurate descriptions of individual buildings. My approval had allowed the presentation to conceal that difference.
The delivery team proposed retaining the map with a general caution below it. I withdrew the view from the intended screening workflow and allocated ₹44 lakh over nine weeks to source review, representation and use controls. Surveying and business professionals helped define what evidence the task required. We separated verified positions from area-level references and withheld the eighteen unresolved positions from a definitive building-location account. The forty-two postal-area centres remained useful only for purposes compatible with their coarser meaning; they were not relabelled as surveyed buildings.
The revised view showed 120 positions with their survey provenance, forty-two area references with explicit limitations and eighteen unresolved records outside detailed screening. Those counts describe the reviewed set, not the entire insurer portfolio. No claim is made that all records were subsequently surveyed or that the review changed a customer’s cover. Underwriting and other professional owners retained responsibility for their decisions. The programme expense was not presented as an avoided claims loss derived from a more attractive map.
The difficult correction concerned my own standard of approval. I had focused on whether the display was coherent, without asking whether a user would infer a common precision that the sources did not possess. A caution could not repair every downstream assumption once the uniform symbol had become the working truth. I now ask how differences in provenance remain visible at the point of use, and whether an unresolved location should be withheld from a task rather than included merely to make the view complete.
2023: the latest boundary map could not silently rewrite an earlier report
A reconstruction exercise examined 96 property records for a defined earlier reporting period. The current geographic classification used a newer boundary version than the one appropriate to that period. Fifty-eight records retained the same classification across the versions, twenty-six had a different historical classification and twelve lacked sufficient version evidence to resolve the question. The properties had not necessarily moved. The relation between each position and the reporting boundary had changed, or could not yet be established.
The initial proposal was to apply the latest map to all 96 records because it was available in the central service and would produce a complete table. I asked the team to preserve the historical question instead. We allocated ₹61 lakh over twelve weeks to reconstruct the relevant source versions and document the reporting relationship. The responsible reporting owner reviewed the intended period and treatment, while data colleagues established the provenance of the maps and joins. We did not decide a tax, legal or insurance classification simply by selecting a technical version.
The reviewed output supported eighty-four historical classifications: the fifty-eight unchanged cases and the twenty-six whose earlier classification differed. Twelve remained unresolved and were withheld from a definitive historical account pending further evidence and the reporting owner’s decision. The report distinguished an actual change in the underlying property from a change in the geographic reference. It did not conceal an error in an earlier report or claim that every historical question had now been answered. No financial restatement amount is invented from the classification work.
The decision preserved a distinction that an always-current data service can make easy to forget. Updated information is valuable, but the newest version is not automatically the answer to every question about the past. The CDO account needed to explain why a less complete historical table was more faithful to the evidence than a complete table answering a different question. The remaining twelve records were part of that explanation, not a gap to fill with today’s convenient relationship.
2025: a sharp flood boundary still left uncertain classifications
A defined property-screening exercise initially placed seventy of 320 sites inside a modelled flood boundary and 250 outside it. The proposed dashboard would have displayed those categories as definitive property-level answers. Hydrological specialists and location stewards asked how uncertainty in the source, position and boundary affected sites near the line. Under the documented sensitivity exercise, sixty-three remained inside and 227 remained outside across the examined variations. Thirty were unstable: seven from the original inside group and twenty-three from the original outside group.
The business sponsor wanted to retain a simple two-colour map for ease of use. I recommended an explicit unresolved category and a review of the intended decision before release. We spent ₹56 lakh over ten weeks on the sensitivity work and revised account. The specialists defined the relevant modelling assumptions; my data function preserved the inputs, classifications and limitations. We did not select a universal safety distance or claim that the sensitivity range represented every possible flood event.
The revised screening account retained sixty-three stable-inside,227 stable-outside and thirty uncertain classifications for that exercise. Stable-outside did not mean safe from flooding, and an inside result did not establish a claim, insurance cover or expected financial loss. Professional owners assessed how this information could support their work and which sites required further investigation. The thirty uncertain cases were not redistributed merely to restore a binary chart. No reduction in insured losses or premium improvement is attributed to the changed display.
The useful outcome was a more honest decision aid, with the uncertainty placed where the user encountered the classification. A sharp graphic line had implied a degree of certainty that the underlying evidence did not provide. I wanted the leadership discussion to confront that limitation rather than hide it in a methodology appendix. The resulting map was less simple, but the organisation could see which conclusion remained stable under the examined assumptions and which needed professional attention before it was used.