Imagine a common scenario in financial services. A team deploys AI to review contracts: hundreds of pages, repetitive clauses, and routine work that normally takes employees days. The model identifies the risks and references to correct paragraphs. Yet the output doesn’t add up.
It turns out the problem is in the contracts folder, where files sit side by side named contract_v1.docx, contract_v2.docx, contract_final.docx, contract_final2.docx, and contract_finalfinal_THIS_ONE.docx.
The AI selected the most recently saved file. But that file was an older version that someone had accidentally opened and auto saved last week.
The model did exactly what it was designed to do. The process, however, let it down.
When the Model Isn’t the Problem
In my previous article, I argued that AI failures often begin outside the model – in the governance, ownership and decision-making structures surrounding it. But governance is only part of the problem. Even a well-governed AI system will struggle if the process beneath it depends on outdated documents, informal exceptions or approval paths designed for a much slower pace of work.
When an AI system produces a poor result, the model is usually the first suspect. Perhaps it hallucinated, misunderstood the task, or received a weak prompt. But sometimes the model did exactly what it was supposed to do. The failure occurred in the process that supplied its information or in the organization that could not act on its output.
Deploying AI reveals process problems that always existed but were never visible – like a diagnostic tool. It exposes weaknesses that people previously compensated for without thinking: missing context, unclear ownership and informal workarounds that were never documented.
Speed Without Capacity
AI does not only accelerate output. It also accelerates escalations, requests for decisions and the discovery of problems. A model flags an anomaly in real time, not on Friday afternoon when someone finally finds the time to go through the logs.
Organizations are not set up for this pace. Approval processes, escalation paths, decision-making authority – all of it was calibrated to human speed. When AI produces outputs faster than the organization can absorb them, three things can happen:
- AI waits for a decision, erasing the efficiency gain.
- Employees ignore its output and return to what they believe works – the old process.
- Employees push decisions forward without properly checking them because they feel pressured to maintain speed.
Some employees return to manual checks because those feel safer. Others approve results they have had no time to verify because they feel pressured to maintain the speed promised by the technology. In both cases, the organization pays for faster output without becoming faster at making responsible decisions.
Ownership can also become blurred. Who is responsible for reviewing a flagged anomaly, deciding whether the model is wrong and escalating the issue when several teams are involved? Without a clear answer, the AI continues running in the background, producing outputs that nobody truly owns.
This reflects a broader organizational problem. As a Forbes analysis from late 2025 observed, when technology changes workflows faster than an organization can absorb, the result is not efficiency; it is overwork. AI may accelerate one stage of a process, but its value still depends on whether the surrounding organization can absorb and act on what it produces.
The Silent Knowledge Nobody Wrote Down
Imagine that Martin knows the company’s contract with a supplier includes an exception for payments below €2,000. It was agreed verbally three years ago but never documented. When the company deploys AI to automate payment approvals, the system blocks the payment – correctly, according to the written contract.
This is tacit knowledge: organizational knowledge that lives in people’s heads rather than in systems. Every company has it. And most companies have no idea how much of it they have.
AI can only use knowledge that has been made accessible to it. Martin’s exception does not exist from the system’s perspective. Without that context, its decision may be technically correct but operationally wrong.
McKinsey has identified the same challenge in agentic AI deployments. Building effective AI agents requires companies to codify expert practices that may exist in standard procedures – or only as tacit knowledge in employees’ heads. In other words: AI deployment may therefore be the first time an organization sees just how much the process depends on what Martin knows.
The False Source of Truth
The most dangerous AI output is not always an obviously incorrect one. It may be a credible, professionally presented answer produced from outdated information.
For years, many departments have operated with multiple versions of the same document because Jane from legal always knows which one is authoritative. AI, unfortunately, does not.
Unless it has been given rules for distinguishing between versions, the system has no reason to question the document it receives. And so the AI review runs correctly – on the wrong document. That may be worse than no review at all because the output looks credible.
An experienced employee may know where the truth is and can navigate through chaos. A system, however, needs a reliable way to identify which information is current, approved and relevant.
Tacit knowledge, unreliable documents, and approval processes built for human-speed output appear to be separate problems. But they share the same root cause. None of them were created by AI, but AI makes them harder to ignore.
Questions to Answer Before You Build an Agent
Before deploying an AI agent, an organization should be able to answer three questions.
Which decisions depend on knowledge that has never been documented?
This means identifying the exceptions, shortcuts and judgement calls that experienced employees apply without consciously describing them. Interviews alone may not reveal all of this knowledge. Organizations may need to observe how work is actually performed and compare it with the official process.
How will the system identify information that is current, approved and relevant?
Giving an agent access to more documents does not solve the problem if it cannot distinguish an approved contract from a draft or an active policy from an outdated one. Versioning, ownership, approval status and retention rules must be clear enough for the system to identify which source should govern its decision.
Can existing approval and escalation processes absorb the volume and speed of its output?
Teams should estimate not only how much work the agent can complete, but also how many reviews, exceptions and escalations that work may create. The NIST AI Risk Management Framework recommends clearly defining roles and responsibilities for human oversight of AI systems. These need to be designed for the expected volume before the system is deployed – not after employees become overwhelmed by it.
These are not questions to answer once and forget. Processes change, documents age and undocumented exceptions accumulate again. Readiness therefore has to be maintained, not merely established before launch.
AI Readiness Starts with the Process
AI readiness is not only a test of the technology. It is a test of whether the organization has made its own processes explicit enough for the technology to operate within them.
Preparing an organization for AI therefore involves more than selecting a model or building an agent. It means clarifying ownership, identifying authoritative information, documenting exceptions and redesigning how outputs are reviewed, escalated and acted upon.
This is difficult work, but a failed deployment will eventually force the organization to do it anyway. The only variable is timing.
