Most companies do not have an AI tool problem. They have a decision-quality problem caused by scattered, stale and inconsistently governed information. The commercial choice is therefore not which assistant to licence first, but whether the organisation can make its own data reliable enough for AI to support real work.
The model is not the bottleneck
The pattern we keep seeing is that impressive demonstrations create the wrong sense of progress. An assistant can summarize a meeting in seconds, draft a response or answer a question across documents. That proves the interface works. It does not prove the organisation has a dependable operating capability.
The difference is the data beneath the answer. If policies conflict, ownership is unclear and access rules have accumulated without review, AI will make those weaknesses easier to discover and harder to ignore. It may still produce fluent work. Fluency is not the same as a sound decision.
This is especially clear in Microsoft 365 environments. Copilot follows a user’s existing permissions rather than granting new ones. That is a sensible security design, but it shifts the practical risk to permissions that may no longer reflect the business. Files shared broadly for an old project can remain technically available long after the project has ended. Before AI, finding them required effort and prior knowledge. With AI, a natural-language request can expose the same underlying access more quickly. Independent guidance describes this as years of quiet oversharing becoming instantly discoverable.
Consider a finance director preparing for a board meeting. She asks an internal assistant to explain the company’s latest workforce costs. The assistant retrieves a current budget, an outdated planning file and a spreadsheet containing restricted compensation details because all three are accessible within her account. The response may be coherent, yet the decision-maker cannot easily tell which source is authoritative or why sensitive information appeared.
The failure is not that the assistant misunderstood the question. The failure is that the organisation never made source ownership, document status and access intent explicit. Improving the model would not solve that. A governed data foundation could: authoritative sources would be identifiable, stale material would have a lifecycle, and sensitive content would have enforceable boundaries.
That work is less visible than launching a chatbot, and often more valuable. It reduces the time people spend reconciling versions, searching for approvals and asking colleagues where the real answer lives. It also changes who owns information quality. Data governance stops being a periodic technology exercise and becomes part of how finance, legal, human resources and operations run their work.
The strongest objection is cost. Cleaning permissions and content can delay deployment, while a pilot can show value immediately. That objection is valid when governance becomes an excuse for indefinite review. It is weak when governance is treated as product infrastructure. A narrow, monitored rollout can expose the highest-risk data while producing evidence about which controls matter. The aim is not perfect information before use. It is controlled usefulness with a clear path to improvement.
Our view is that the competitive advantage will not come from owning a slightly better assistant. It will come from having information that can be safely turned into decisions at lower coordination cost. The AI race is therefore an organisational race: the companies that prepare their data will not simply generate faster answers. They will spend less effort questioning whether those answers can be trusted.