Chief Information Officer / India
Keep the unresolved cases inside the result
- Commercial-insurance employer
- 680 employees
- Twenty-two branches; ₹1,540 crore FY26 gross written premium
- Technology organisation
- 72 employees
- Five heads and eighteen contractors counted separately
- First external appointment under consideration
- 0 statutory seats
- July 2026 fictional IICA registration; assessment preparation continuing
Decisions & consequences
01The same amount was not enough to identify a retry
In 2022, operations reviewed 238 queried or repeated receipt messages within a population of 12,600 messages. Some callers had retried because they did not receive a response. Others had sent separate valid instructions with similar values. The proposed shortcut was to suppress messages sharing an amount and customer contact within a short period. That would have made the technical queue smaller, but it could also have removed a legitimate second instruction. A similar payload did not establish the sender's intent.
The reviewed group contained 142 genuine retries, 84 distinct instructions and twelve cases requiring reconciliation. Dinesh Dovqari wanted a service that colleagues could use without repeatedly asking technology what had happened. Finance wanted a dependable receipt record, while authorised insurance specialists retained any decision about the contract. We defined a request reference scoped to the intended action, with the recorded result linked consistently to that action. A reused reference with different details required review; an uncertain result could not be relabelled complete merely to clear the queue.
The project cost ₹78 lakh and added ₹12 lakh annual support expense. It included handling lost responses and delayed messages, with operations and finance checking the resulting record. The twelve ambiguous cases were reconciled through their appropriate process before being marked resolved. We did not describe them as twelve duplicate payments or imply that receipt processing itself established cover. The reviewed counts described the messages examined, not a universal failure rate across every transaction the insurer handled.
The main tradeoff was a more explicit service contract and continuing reconciliation work in place of a convenient similarity rule. Some integration partners had to retain a stable reference when retrying instead of generating a new one each time. That required explanation and testing. The result was a clearer connection between an instruction and its recorded outcome, without pretending that every uncertain situation could be settled automatically. I would now ask for the retry story at the beginning of an integration proposal, including what the caller is told when the response is missing, rather than treat it as an error-handling detail to be added after the ordinary journey works.
A shared office did not make two customers one entity
A 2023 customer-record project proposed 740 merges using similarities in names, addresses, telephone numbers and contact people. The programme expected a tidier customer view and less repeated administration. Commercial-insurance colleagues warned that several organisations could use the same office or intermediary contact while remaining distinct legal customers. A merge could therefore make the database look cleaner while confusing which entity a record described. I supported reviewing the proposed relationships before releasing the automatic merge rule.
We allocated ₹62 lakh to business review, evidence handling and the revised workflow. Authorised colleagues determined the relevant entity relationships; technology displayed candidate links and retained the basis for accepted or rejected matches. A shared contact could be recorded as a relationship without combining the customer identities. We also kept unresolved cases visible, because leaving two records separate temporarily was preferable to making an unsupported assertion merely to complete the migration count.
The review classified 196 proposed merges as the same entity, 487 as distinct entities and 57 as unresolved. Those categories totalled the original 740 proposals. The 196 accepted combinations followed the approved process, the 487 remained separate and the 57 were held for further evidence. We did not call the unresolved cases fraudulent, infer credit quality or change policy and claims decisions from the matching result. The project addressed the identity represented by the record, with other professional responsibilities remaining where they belonged.
The commercial sponsor reasonably objected that the additional review delayed a simpler customer view and increased expense. I agreed that duplicate records created real work, but argued that an incorrect combination could create a different and more consequential problem. The revised design allowed colleagues to see relevant relationships without requiring a single identity where the evidence did not support one. The outcome was not a claim that the entire customer population had been verified. It was a controlled decision on the 740 proposed merges, with 57 still unresolved. The case taught me to ask what business assertion a data-cleaning action makes before celebrating the reduction in record count.
The newest entry was not always the right view of the date
In 2024, a review of 180 approved insurance changes found 37 displayed in the wrong business-time view. The system favoured the most recently entered record. That worked when events were entered in the same order in which they applied, but it could misrepresent a late entry, a correction or a future-effective change. I had accepted the simplified design because it made the history easier to implement. The implementation preserved a timestamp while losing a distinction colleagues needed to interpret the record.
We funded ₹92 lakh to represent the approved effective date and the recorded date separately, with clear views for the relevant tasks. Underwriting and operations owners defined the source instruction and authorisation required. The system change did not permit a user to create retrospective cover or alter a contractual consequence merely by choosing another date. It preserved the business decision and when the record became available, so an authorised colleague could reconstruct the appropriate account rather than rely on whichever entry happened to be newest.
The test set contained 102 scenarios involving late entries, corrections, future-effective changes and ordinary sequences. Eighty-eight were supported by the revised path. Fourteen legacy scenarios still required an assisted process and remained outside the claim of completed automated support. We reported those fourteen alongside the release result. The earlier 37 findings and the later 102 scenarios were different kinds of evidence; they were not presented as a comparable before-and-after defect rate or proof that every historical record was now correct.
The difficult choice was to retain two understandable dates and some assisted work rather than offer one reassuring timeline. Colleagues needed guidance on which view answered their question, and support had to preserve the authorised source when resolving exceptions. The board saw the expense, the supported scenarios and the remaining scope together. My correction was to recognise that simplicity in a technical representation can transfer complexity to the person interpreting it. I would now ask what must be reconstructed after a late correction before approving a history model, rather than assume that sorting by entry time supplies the business meaning of the sequence.
Four unfinished keyboard tasks stayed in the release account
A 2025 review examined sixty representative tasks in an internal insurance service application using keyboard operation. Twenty-one were blocked at some point in the working sequence. The visual demonstration had looked complete, but users could not reliably reach or operate certain controls, follow the focus or finish an error-recovery path without a mouse. The problem was not established by the appearance of the page. It became visible when people attempted the actual task through a different interaction method.
We approved a ₹1.2 crore change programme with appropriate users and accessibility specialists involved in review. The work covered complete task sequences, including opening records, editing, confirming and recovering from errors. We did not classify a task as complete merely because its first control could be reached. Operations also needed a practical route for work that remained unsupported, with a responsible owner and enough capacity to avoid making the affected colleague repeatedly improvise assistance.
On retesting the same sixty tasks, 56 passed and four still required an assisted route. We retained ₹24 lakh annual support capacity for that route while further work continued. The four were shown explicitly in the release account; they were not removed from the denominator to produce a perfect completion claim. The result concerned those tasks under the stated test conditions. It was not a certification of the whole application against every accessibility requirement or evidence about a particular employee's health or ability.
The sponsor wanted to close the programme after the main changes, arguing that the remaining tasks were infrequent. I accepted that frequency could affect sequencing, but it did not make a blocked task disappear for the person who needed it. The funded assistance and unfinished scope therefore remained part of the operating plan. The case changed how I ask for a demonstration: who can complete the work, through which interaction path, and what happens where the answer is still incomplete? Technology success should include the retained effort needed to make the service usable, rather than leave that effort outside the project account because it happens after release.