Most working methods stay dark to the organization. The task is finding the few worth lifting out and testing. Illustration generated with OpenAI.
The AI-Native Company · Part 3 of 5
Employees are developing AI working habits beyond the tools their employers provide. In October 2025, TechRadar covered a LayerX report that attributed 82% of observed pastes into generative AI tools to unmanaged personal accounts. That is a finding about pasting activity in the vendor’s observed environment, not the share of all enterprise AI work. It nevertheless highlights a practical blind spot: some methods that employees use to complete work may never appear in the company’s approved-tool dashboard.
That blind spot includes both risk and useful knowledge. An account manager may have found an effective way to assemble a renewal brief. A recruiter may have developed a better method for comparing evidence against a role’s requirements. Another employee may simply be producing more convincing errors... Management cannot tell which is which from usage counts, and inspecting personal chat histories is not a sound discovery strategy. The problem is identifying transferable methods, together with the judgment needed to use them, through evidence people can appropriately share.
Microsoft and LinkedIn’s 2024 survey found that 78% of AI users brought their own tools to work. This measured people using tools their organization did not provide; it did not measure their share of AI interactions. For an operator, the implication is that sanctioned adoption data may be incomplete. A company can fund a new reporting application while employees already possess a useful method, or promote a popular technique whose review costs exceed its savings. Both mistakes arise from confusing visibility with value.
Three discovery methods work better together than any one works alone: voluntary demonstrations of working practice, analysis of recurring operational demand, and controlled pilots of promising candidates. Demonstrations reveal how people work. Operational data shows whether the outcome matters often enough to justify investment. Pilots establish whether the method transfers beyond its original creator. None is sufficient alone. A compelling demonstration can be a one-off success, while a high-volume task may still be too variable or poorly understood to productize.
- Start by asking employees to demonstrate a recent task using material they are authorized to share. Request the input, the accepted result, the corrections they made and an example that failed. Include the steps outside the assistant: finding data, checking calculations, consulting a colleague and moving the output into a business system. This reveals the actual method rather than just its prompt. Participation will be more useful when the purpose is clear and the company provides time and appropriate tools for it. A voluntary sample will still overrepresent enthusiasts, so it should not be treated as a complete inventory.
- Next, connect those demonstrations to operational demand. CRM records, service tickets, reporting calendars and project systems can show how often an outcome is needed, who needs it and where work waits or comes back for correction. Interview teams that rarely use AI as well as the enthusiasts. The commercially attractive candidate is a recurring outcome with a transferable method and a meaningful improvement available. Frequent prompting is a weak proxy: a difficult, low-value task can generate more messages than a highly valuable process that runs quietly once a month.
- Then test the candidate with someone other than its creator, using cases that were not used to develop the method (brace for impact during this phase). Compare accepted output quality, elapsed time, review effort and rework with the current approach. Keep the required quality threshold consistent. This is where tacit knowledge becomes visible: the expert may recognize an unreliable source or an unusual customer condition without mentioning it in the prompt. Capture those decisions before packaging the process. A method that transfers only after substantial expert intervention may still be useful, but its business case must include that intervention.
The counterargument is that employees already know whether a tool saves them time. Sometimes they do; perception alone is still a poor investment gate. In a 2025 randomized study of 16 experienced open-source developers completing 246 tasks, METR researchers found that AI access increased completion time by 19%, although participants later estimated a 20% reduction. As the paper puts it, “AI tooling slowed developers down.” This concerned particular developers, repositories and early-2025 tools. It is evidence for measuring outcomes, not a claim about every current AI application.
For a COO, RevOps leader or practice head, the useful discipline is two assessments kept separate. Productization potential covers demand, transferability and economic value. Readiness covers data access, evaluation, permissions, exception handling and ownership. A valuable process can be unready because its source data is unreliable. An easy process can be too small to matter. Give each promising candidate a process owner, a baseline and a specific uncertainty the pilot must resolve. Avoid an aggregate score that allows strong demand to conceal an unacceptable data-access problem.
What this produces is a better investment queue. Consider an illustrative process with 1,000 monthly cases, of which 60% qualify for automation. At 70% adoption, that is 420 executions. If each produces $8 of validated net operating benefit before platform and maintenance costs, the monthly benefit is $3,360; $600 in recurring costs reduces it to $2,760 before implementation costs (those are assumptions, not a benchmark). Where the benefit represents staff time, management must establish how that capacity will be used before calling it cash savings.
The next issue starts where the pilot ends. A method can show economic value and still be unsuitable for unattended execution. We will look at progressive autonomy: what evidence justifies moving from recommendations to approved actions, then to execution within defined limits, and what should cause that authority to shrink again.
Florian Boymond is the founder and CEO of Stackmint, focused on turning enterprise expertise into dependable AI operations.
The AI-Native Company series
- Part 1: Top-down vs. bottom-up AI: what are we actually changing?
- Part 2: The assistant as the front door to company workflows
Originally published on LinkedIn on 2026-10-08.