Chief Technology Officer / India
Three occasions when the output needed a boundary
- Engineers
- 85
- Four engineering heads
- Corporate clients
- 58
- Software users; no client assets held
- Engineering budget
- ₹18 crore
- Approved annual operating remit
Decisions & consequences
01A forecast total that should have remained incomplete
During a review of an import change, our team examined 146 incoming source sets from seven pilot clients. Fourteen lacked a current opening balance, but a legacy default converted the missing value to zero. The resulting forecast could still display a complete-looking total. A zero might be a valid balance in another case; here it concealed the absence of evidence.
I asked the product director to pause the planned release and agree how the affected state should be represented. The commercial sponsor wanted to retain the existing behaviour with an additional note, arguing that some users preferred a continuous total. I recommended that the product mark the forecast as incomplete unless an authorised user supplied a clearly identified assumption. We also required the qualification to appear in exports, because clients often discussed the figures outside the application.
The approved work cost ₹28 lakh and took six weeks, including migration support for existing configurations. We tested the fourteen affected source patterns and additional examples in which a genuine zero was supplied. The revised behaviour kept those conditions distinct and recorded the owner of any manual assumption. A planned charting enhancement moved into the following release so the same engineering team could complete the work. We did not claim to know what decisions clients would otherwise have made from the earlier display.
The correction revealed that our original design review had focused on whether imported values could be calculated, rather than what an absent value meant. I now ask for that meaning before agreeing a default. The result was not a prohibition on estimates: clients could still construct scenarios. It was a clearer separation between supplied evidence, an owned assumption and a gap that the software should not quietly fill on their behalf.
A consolidated view without inherited access by accident
A pilot client wanted a group-level cash-planning view that combined several subsidiaries. The proposed design used the existing corporate hierarchy to determine which detail a user could open. During review, we found that the hierarchy recorded ownership relationships, while the client's permission decisions depended on role and purpose. Reusing one structure for both would have made the feature easier to implement but harder to justify.
I recommended a separate permission review before accepting the consolidated view. The sponsor was concerned that this would delay a board-reporting demonstration already promised by the commercial team. We agreed to demonstrate a restricted aggregate view while withholding entity-level detail until the authorised client owners had approved the access model. The review covered forty-two pilot user permissions; twelve lacked a documented justification and were removed from the proposed allocation.
Engineering and migration work cost ₹17 lakh and extended the detailed-view release by four weeks. We tested role changes, a user moving between entities and a subsidiary leaving the reporting group with synthetic records. The client then reviewed the intended permissions using its own governance process. Our work did not determine the legal entitlement of every user or establish that no other access weakness existed. It implemented and demonstrated the approved boundaries for this feature.
The aggregate demonstration met only part of the original request, and the commercial team had to explain that narrower scope. I should have raised the distinction between hierarchy and permission during discovery rather than relying on an implementation review to expose it. The case changed our design checklist: organisational relationships may help navigate a product, but authority to inspect underlying detail needs its own evidence. The inconvenience of an extra review was real, and it belonged in the delivery plan rather than being treated as a surprise caused by cautious engineers.
Allowing the projection to decline a number
We piloted an automated cash projection with nine corporate clients. The feature used historical receipt and payment patterns to prepare a planning view for client review. Three pilot businesses had recently changed their operating pattern through a new product line, a business-unit transfer or revised customer terms. The available history could not adequately describe those changes, yet the prototype still produced a precise-looking figure.
The product team wanted to publish the number with a caution note and collect feedback. I argued that the proposed behaviour would make the model's confidence appear greater than the evidence. We agreed to hold automatic output for the affected cases and provide a scenario route in which the client's authorised user supplied assumptions with a reason and review date. The remaining six pilots continued under the agreed evaluation conditions; their use did not establish that the method would work for every business.
We allocated ₹21 lakh to the qualification workflow, scenario history and additional evaluation. The decision delayed the broad release by five weeks. Two of the three affected clients used the manual scenario route during the pilot, while one retained its existing planning process. Those outcomes were included in the commercial review rather than counted as three successful automated deployments. We also separated the projection from confirmed cash balances in the product's explanation and exports.
My learning was that uncertainty needs a supported operating path, not just an apologetic message. A client should understand why a figure is withheld and what evidence or judgement is needed next. The revised feature still depends on users identifying relevant business changes; it cannot recognise every event that makes history less useful. We recorded that limitation and continued evaluation. The engineering decision was to make an unsupported answer visible as unsupported, while preserving a practical way for the client to construct and own an alternative scenario.