The citation looked convincing before we examined its status

## Ask what the source allows the answer to claim

The demonstration was persuasive because the answers did not appear to arrive from nowhere. They contained source links, and those links gave the response a visible basis. I initially treated that feature as a strong reason for confidence in our research-assistant pilot. The subsequent review showed why the presence of a citation is only the beginning of the examination. A source can exist while being superseded, inappropriate to the question or insufficient to support the answer attached to it. The confidence of the prose does not settle any of those matters.

At Vornqemi, qualified reviewers examined ninety-six questions in the ₹54 lakh pilot. Fifty-eight answers had current supporting sources. Twenty-two relied on superseded material, and sixteen lacked adequate support for the answer given. The categories described the reviewed set, not a universal performance rate for every possible question. They were sufficient to challenge the wider proposition we were considering. A service intended to help colleagues understand reference material needed a clearer account of which material should inform the answer and what should happen when the basis remained uncertain.

The distinction between a real source and a suitable source is particularly important in a continuing industrial business. Documents can remain valuable as historical records after they cease to support a current instruction or description. Their preservation does not mean that every later use is appropriate. A qualified person may need the earlier account to understand a decision, while another question requires the current approved position. I would not ask a language model or a search ranking to settle that professional distinction by implication. The source collection and the intended task need accountable owners.

The programme sponsor preferred to add a clearer disclaimer and proceed with the wider launch. He argued that employees already knew to exercise judgement and that the tool could still save time. That was a reasonable benefit to investigate. It did not remove the question of what the interface encouraged users to rely on. Yusuf Dorzemi objected to any suggestion that an assistant retrieving engineering-related material could therefore issue operating instructions. I supported narrowing the use rather than treating a cautionary sentence as a complete answer to the findings.

## Make the narrower use a real operating choice

The next stage was limited to research assistance with human review and an explicit account of source status. It was not authorised to control plant, prescribe technical work or replace the responsible specialist's decision. Those limits had to shape the actual service and how it was introduced, rather than appear only in an approval paper. A narrower description would have little value if demonstrations, prompts or surrounding communications continued to imply that the system could provide a settled operating answer. The user should be able to understand what kind of help is being offered.

We retained the full ₹54 lakh pilot cost. The decision to continue a restricted use did not recover that money or establish a return on it. Nor did the fifty-eight supported answers prove that the remaining work would be straightforward. Unsupported answers and superseded sources remained findings in the evaluation. They were not omitted to improve the result presented to executives. I wanted the team to be able to explain why a narrower use was still worth exploring without pretending that the original proposition had already been demonstrated successfully.

Human review also needs a practical meaning. It is not enough to add a person's name to the process if that person lacks the competence, time or access required to examine the answer. The appropriate reviewer depends on the material and decision involved. My responsibility as CIO was to make our information service support that review and preserve its boundaries. I did not become the technical authority for thermal operations by sponsoring the software. Qualified engineering, operating and other owners retained responsibility for the conclusions within their respective fields.

The restricted stage required more effort from users than the first demonstration suggested. They had to inspect the basis of an answer rather than treat a fluent paragraph as a completed result. That could reduce the apparent time benefit for some tasks. I considered it a finding to understand, not an inconvenience to hide. If a service saves drafting time but creates substantial verification work, the company needs that fuller account before deciding where it is useful. A benefit should be examined in relation to the actual task, including the work left with people after the system responds.

## Keep responsibility with the people interpreting the material

By the October review, we had not established a quantified productivity return or a later defect rate. The limited evaluation supported a change in permitted use and further work, not a broad claim that the research problem had been solved. Management should be comfortable reporting that kind of outcome when it is what the evidence supports. The company had learned something consequential about the proposition it was considering. That learning would be weakened if the report converted the decision to narrow the service into proof that the narrower service was already effective in every intended setting.

The episode made me reconsider what I ask to see before an executive demonstration. A smooth ordinary question is useful, but it can leave the important boundaries unexplored. I would want qualified users to examine questions involving superseded material, insufficient evidence and situations outside the proposed scope. The purpose is not to create an adversarial performance in which the technology must appear flawless. It is to understand where the service helps, where it needs assistance and where it should not be used. Those distinctions should inform the proposition before the demonstration establishes an expectation of authority.

The underlying reference collection deserves attention as well. A tool can expose weaknesses that were already present in how documents were maintained or understood. That does not make every error the source owner's fault, nor does it mean improving the collection will eliminate every misleading answer. It means the programme should avoid describing the issue as if one component alone could settle it. Source material, retrieval, presentation, review and intended decision interact. Management needs owners for the relevant work and an account of the remaining limitations rather than an assurance attached to one attractive feature.

For a board, I would ask how management has connected the proposed use to the evidence from evaluation. What did the reviewers examine? What counted as support? How were unsuitable sources and unanswered questions retained in the result? Which responsibilities remain with qualified people, and does the actual service preserve those responsibilities? These are questions about the enterprise's reliance on information. They do not require directors to choose a model or approve individual prompts, and they should not invite the board to take over the professional decisions the tool is meant to assist.

I also want leaders to own the expectations they create. I had treated visible citations as a stronger basis for confidence than the review justified. Correcting that view was part of the work, not a matter to leave to a disclaimer written by somebody else. A research aid may remain useful without being an authority. In this case, the responsible next step was to narrow the service, retain the costs and unresolved findings, and make the source question explicit. The link beneath an answer should invite examination of its basis, not become a visual substitute for that examination.