The delays, extra costs and exceptions inside my decisions

Our 2025 recovery exercise used 12,000 test instructions in an isolated environment. No customer money moved. The application restarted after forty-two minutes, which met the engineering team's opening expectation. At fifty-seven minutes, however, operations could reconcile only 11,040 instructions. The remaining 960 had uncertain processing status. Calling the service recovered at the first timestamp would have concealed the work that still prevented a controlled return to normal processing.

The infrastructure lead wanted recognition for the successful restart. The service owner would not authorise further processing until the uncertain population was understood. I supported that refusal and asked the team to preserve the records instead of clearing the queue to make the dashboard look normal. We traced the interrupted transitions, compared acknowledgements and reconciled the test ledger. Every instruction was accounted for after two hours and forty-six minutes. The longer duration became the result presented to management, alongside the earlier technical milestones.

I approved CNY6.4 million for better status evidence, reconciliation tooling and operating exercises. Four months of planned feature work moved back because the same specialists were needed. Commercial sponsors challenged that delay; they were entitled to see its consequences. I explained the unresolved service condition rather than arguing that an infrastructure project should automatically outrank their requests. The operations team also changed the authority required before releasing held instructions, with risk colleagues reviewing the proposed controls.

A later exercise using an equivalent bounded scenario accounted for all instructions in seventy-eight minutes. That was a useful improvement over two hours and forty-six minutes, but it did not establish that every outage could be resolved within the same period. We retained the scenario, data population and remaining assumptions in the report. I would involve the service owner earlier when setting an exercise's completion criteria. The most consequential finding was not a slow component; it was our initial willingness to use a technical timestamp for a banking outcome it did not yet describe.

A rehearsal for replacing our term-deposit platform covered 250,000 records with CNY7.8 billion of principal. The aggregate principal reconciled. A business review then found that 6,200 records would receive the wrong renewal treatment under the proposed mapping. The arithmetic success had obscured a difference in what the customer had instructed the bank to do at maturity. That distinction made the migration unacceptable in its current form.

The programme sponsor was concerned about extending parallel operation and losing the agreed launch window. Branch operations wanted evidence that each renewal category retained its intended meaning. I recommended a six-week postponement, costing CNY1.9 million in additional parallel operation. The CEO accepted the recommendation after finance and the service owner examined the revised plan. I did not describe the delay as free insurance: it consumed money and specialist attention that would otherwise have supported other work.

We paired conversion engineers with deposit specialists to examine the affected mapping rules. The 6,200 rehearsal records were corrected and retested before production approval. Reviewers also added instruction-level sampling to the balance reconciliation, with a named business owner for each exception category. Technology supplied the converted records and evidence; it did not decide that a customer's instruction could be reinterpreted to fit the new system. The release paper retained both the corrected findings and the possibility of distinct live exceptions.

During the actual transition, 249,983 records were accepted through the main process. Seventeen separate exceptions were held and resolved over the following two days. Those were production exceptions, not a remainder of the 6,200 rehearsal mapping errors. Keeping the populations separate helped management understand what the additional preparation had achieved and what work still occurred at cutover. I would now make semantic checks part of the first rehearsal specification rather than wait for a business reviewer to expose their absence. A matching total is essential evidence, but it cannot tell us whether the bank will honour the same instruction tomorrow.

Our loan-workflow application depended on a specialist supplier. Contract discussions emphasised possession of source files, which initially appeared to provide a credible alternative if the relationship ended. An internal reproduction exercise revealed nine processing rules that our team could not explain from the delivered material. We had files without sufficient knowledge to operate the same workflow independently. That finding changed the commercial discussion.

Procurement wanted to use the renewal to reduce the annual support charge. The lending operations owner was more concerned that an abrupt supplier change would disrupt work that remained poorly documented. I proposed a renewal with explicit capability-transfer sessions and practical demonstrations. Annual support rose from CNY9.6 million to CNY11.2 million. Management accepted the higher charge because a cheaper agreement would have left the most consequential uncertainty untouched. The additional expenditure was a deliberate purchase of knowledge and access, not an immediate cost reduction.

We assigned bank engineers to reproduce each rule in a controlled setting, with lending specialists checking the business interpretation. Supplier staff explained dependencies and reviewed the results. Commercial and legal colleagues handled the agreement; my responsibility was the technical evidence behind its delivery. We distinguished attendance at a workshop from a successful reproduction, and recorded the versions and supporting components used. This prevented a completed training schedule from becoming the sole measure of transfer.

After eight months, our team had reproduced seven of the nine rules. Two still required supplier participation. I reported that incomplete position to the technology committee rather than announce that the bank had achieved independence. The work improved our ability to question changes and plan alternatives, but it did not create a ready replacement for the whole service. I would start the reproduction exercise before the renewal negotiation next time, when both parties have more room to price the remaining work. A contract can provide rights; the bank still has to establish what its people can actually do with them.

An analysis of customer records produced 39,000 candidate pairs for possible matching. The initial proposal treated that population as a substantial opportunity to simplify records quickly. Branch colleagues cautioned that similar names and contact details did not establish that two records belonged to the same person. I supported a fixed review sample of 4,000 pairs before any broader programme was approved. We assigned twelve reviewers over six months at a cost of CNY2.8 million.

The reviewed sample contained 2,760 pairs assessed as the same customer and 1,240 assessed as distinct customers. Even the first group did not automatically authorise a merger. Account authorities and supporting records still had to satisfy the bank's approved process. Operations approved and merged 2,100 pairs. Another 660 preliminary matches remained pending those checks. The 1,240 distinct pairs stayed separate. These categories reconciled to the reviewed population and remained visible in the management paper.

The data team wanted a larger automated run once the matching logic had been refined. The service owner asked what would happen when a later exception contradicted an earlier match. I recommended that technology improve the review evidence and audit trail before increasing volume. That decision retained manual cost and slowed the proposed cleanup. It also preserved the distinction between identifying a likely relationship and authorising a change to banking records. Compliance and operations retained their own decisions; the CIO did not certify an identity or account authority through an algorithm.

We did not extrapolate the sample's proportions to the remaining 35,000 candidate pairs. Their composition could differ, and the first group had been chosen for a bounded review rather than as a statistical representation of the whole queue. The practical outcome was a better controlled decision process and 2,100 approved mergers, with unresolved work still identified. I would describe that evidence before discussing the size of the original opportunity. A large candidate count can make a programme appear valuable while saying very little about how many changes the bank should actually make.

For 2026, business sponsors brought twenty-seven technology requests with a combined estimated cost of CNY112 million. Our authorised change envelope was CNY92 million, within a total technology authorisation of CNY190 million that also covered CNY98 million of operations. The gap was not only financial. Several proposals needed the same experienced reviewers and branch teams, so approving every request with a smaller allocation would have left an implausible delivery plan.

I asked sponsors to examine dependencies together and state what would be delayed if their initiative took precedence. Seven requests with estimated costs of CNY24 million were deferred. Twenty initiatives received CNY88 million, leaving CNY4 million of contingency inside the change envelope. Finance reviewed the estimates and affordability. Business owners retained responsibility for the benefits they expected. I was responsible for explaining the technical sequence and the capacity assumptions rather than claiming ownership of every commercial outcome.

Two sponsors challenged the deferrals because their individual cases remained attractive. I agreed that a sound proposal could still wait when the organisation lacked capacity to execute it alongside other commitments. We recorded the reasons and the conditions for reconsideration, including whether an existing dependency had cleared. The CNY24 million was deferred requested expenditure, not cash already spent and subsequently recovered. Describing it as a saving would have misrepresented both the decision and the future demand.

Of fourteen milestones due in the first half, eleven were completed and three remained late. I kept that result in the portfolio review instead of using the smaller approved list as proof that prioritisation had solved delivery. Sponsors could see where their own readiness, a technical dependency or a revised estimate required action. I would expose the shared specialist constraint earlier in the next planning cycle, before every request acquired a public delivery date. The decision made the plan more honest and affordable; it did not remove the need to manage unfinished work throughout the year.