When One Employee Succeeds with AI, Why Can’t the Organisation Simply Replicate It?

Geoffrey Chen

After AI Enters the Workflow · Season Two: “From Personal Tool to Organisational Capability” · Article 1

A senior analyst begins using AI to prepare a weekly report. She knows where the relevant figures are stored, which numbers are provisional, and which conclusions must not be disclosed. AI helps her arrange tables, compare changes and draft the narrative. Her experience tells her which plausible passages to delete. Work that once occupied a day can now be completed in half a day.

Management sees the result and gives the same tool to the whole department. Within weeks, outcomes diverge. Some employees rely on obsolete files. Some cannot recognise missing data. Others adopt causal explanations that the AI supplied without evidence. The number of reports rises, but so does the review burden. The original success has not been replicated; one person’s skilled practice has become many people’s unstable experiments.

The problem is not that the other employees lack a clever prompt. The initial success depended on conditions that were never written down: domain knowledge, familiarity with records, habits of judgement, sensitivity to errors and informal ways of obtaining corrections. The organisation copied the tool but not the structure of work around it.

Individual success often rests on tacit conditions

At least five kinds of hidden resource may support a strong individual result.

First is task knowledge. The analyst understands which decision the report serves, not merely its stated format.

Second is organisational knowledge. She knows which system is authoritative, which spreadsheet is often late and who can confirm an exception.

Third is quality judgement. She can detect a change of definition, date or causal meaning inside an otherwise ordinary sentence.

Fourth is a network of relationships. When information is incomplete, she knows whom to ask and possesses enough standing to request a correction.

Fifth is personal accountability. She checks carefully because her name and explanation will be attached to the final report.

None of these conditions resides automatically in the model. They may not appear in the prompt. They are distributed across experience, institutions, relationships and working memory. Sending the same prompt template to every employee does not transmit them.

A pilot proves possible value, not stable operation

Individual experiments are excellent for discovering opportunities. They can reveal that a particular step is faster, a set of documents can be organised or a recurring task can be reduced. They rarely answer all the questions needed for organisational adoption.

Will the process still work across different employees, data and peak periods? Can a new staff member detect errors? How is sensitive information controlled? Will an update to the tool change results? Who can continue the process when the original user leaves? If a complaint arises, can the organisation reconstruct the inputs and approvals?

These questions mark the move from demonstration to operation. NIST’s AI Risk Management Framework places governance, mapping, measurement and management across the AI lifecycle precisely because one successful output cannot establish a durable capability. NIST AI RMF Core

ISO/IEC 42001 likewise treats AI as a management-system problem. It specifies requirements for establishing, implementing, maintaining and continually improving organisational policies and processes around AI, rather than treating product selection as sufficient. ISO, “AI management systems”

Why individual gains do not automatically become organisational change

A randomised field experiment across 66 firms and 7,137 knowledge workers gave selected employees access to generative AI integrated into the office applications they already used. Among those who used it, participants later spent about two fewer hours a week on email and reduced work outside normal hours. Yet the study did not detect changes in the quantity or composition of tasks from individual-level provision alone. NBER, “Shifting Work Patterns with Generative AI”

The study does not settle every form of AI adoption, but it clarifies an important distinction. Time saved by an individual is not the same thing as organisational redesign. Two free hours may create a more sustainable working week, or they may be absorbed by additional meetings, email and output expectations. If upstream requests, downstream approvals and performance measures remain unchanged, local speed does not automatically transform the larger system.

OECD research on AI adoption in firms similarly treats skills, data, organisational capability and governance as conditions of adoption. Availability of a capable model is only one factor. OECD, “The Adoption of Artificial Intelligence in Firms”

Extract the work structure instead of copying the behaviour

The organisation should not begin with “How do we make everyone work like this analyst?” It should ask, “Which conditions made this success possible, and which of them can be made explicit?”

The individual practice can be converted into a use-case description: the task’s purpose; the sources of input; the authoritative records; the steps assigned to AI; the points at which the user exercises judgement; common failure modes; evidence of completion; required approvals; and the route for reversal.

Some conditions can be implemented technically. The system can retrieve the current version, restrict sensitive data, require sources for material claims and preserve an activity trace. Other conditions must remain attached to roles and training: recognising anomalies, handling exceptions, communicating with affected people and adopting the final decision.

The thing to be replicated is therefore not the employee’s wording. It is an explained chain of work.

Establish the smallest repeatable unit first

Organisational adoption does not require immediate deployment to everyone. A better unit is a bounded use case, not a general AI doorway.

The minimum unit should include one defined task, permitted sources, an output standard, a test set, a business owner, a technical support route and stopping conditions. Employees with different levels of experience should run it in a controlled but realistic setting. The organisation can then compare quality, time, review demand and unusual cases.

If only the original expert obtains good results, the lesson is not necessarily that every employee should become equally expert. It may show that the system still relies on professional knowledge that has not been represented. The organisation can narrow the task, structure the source material, improve the interface or retain expert review.

Repeatability does not mean that every user must produce identical prose. It means that facts, rules and consequential decisions remain stable across users.

Scaling also requires a decision about who receives the benefit

Individuals often adopt AI out of curiosity or a desire to reduce their own burden. Organisational deployment changes performance expectations, job boundaries and supervision. If employees believe that every hour saved will simply produce a higher quota, or that sharing effective practices will be used to weaken their roles, they may hide actual use or withhold the tacit knowledge needed for safe adoption.

An organisation should therefore state what the gain is for: less overtime, better quality, previously neglected work, higher capacity or some combination. Employees should participate in use-case design because they know the exceptions and workarounds absent from formal process maps.

Diffusing AI is not only a technical transfer. It is a renegotiation of work.

Conclusion: replicate an explainable capability, not individual speed

Individual success matters because it reveals a possibility. It is evidence for discovery, not yet for scaled operation.

A genuine organisational capability must continue when the original user is absent. It must achieve a defined standard across different users and records. When failure occurs, the organisation must be able to locate the evidence, owner and remedy.

The conclusion is:

An organisation cannot replicate individual AI success merely by distributing the same tool. It must convert the task knowledge, authoritative sources, judgement points, verification methods and responsibilities behind that success into a repeatable work structure.

Once that structure exists, AI can become an organisational capability. Until then, rapid rollout may reproduce not the success, but the contingencies that the original expert quietly absorbed.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.