AI will expose whether organizations actually learn from mistakes or simply get better at hiding them.
Failure is inevitable. The more important question is whether an organization can see what went wrong clearly enough to change what happens next.
I came to this view before I started building AI systems. In enterprise operations, I watched people quietly hold weak processes together by remembering missing steps, moving information between systems, catching bad data, and knowing which exceptions mattered. The work got done, but that success could hide how fragile the process actually was.
When workarounds hide the problem
Workarounds are often useful because the work still has to get done. The problem is what the organization learns from them.
Bradley Morrison described workarounds as both a solution to an immediate problem and a mask of the system weakness underneath it. When the workaround succeeds, the original weakness can remain in place because the visible outcome still looks successful. [1]
That is easy to recognize in operations. A handoff can be broken while an experienced employee keeps repairing it, or two systems can fail to share information while people move it manually. Over time, the company can start mistaking successful recovery for a healthy process.
Blame can make failure harder to see
Failure can also disappear when people do not recognize the value of sharing it. Lauren Eskreis-Winkler and Ayelet Fishbach found across five studies that people were reluctant to share failures. In controlled tasks, participants undershared failure even when it contained objectively more information than comparison experiences; highlighting that information increased willingness to share it. [2]
Google's Site Reliability Engineering practice deals with the same problem through blameless postmortems. Google also notes that public shaming can motivate people to cover up facts critical to understanding and preventing recurrence. [3] The point is not to remove accountability. It is to understand what happened, identify what in the system allowed it, and make sure the review produces follow-up action instead of ending with blame or documentation alone.
That is the distinction I think many organizations miss. Making it safer to surface a mistake does not mean lowering standards; it means holding the organization accountable for learning from the mistake. If the same failure keeps happening and the only thing that changes is who gets blamed, the organization has not fixed much.
AI will expose what was already there
This matters more as AI moves from helping with isolated tasks to participating in real workflows. IBM's 2026 survey of 1,000 senior executives found that only 37% of AI initiatives had delivered the business outcomes senior leaders expected by the end of 2025. Executives ranked failure to redesign business processes as the biggest obstacle to getting AI to work at scale, and more than three in four said using AI in isolated parts of a workflow creates limited value unless the broader process is redesigned. [4]
A separate IBM workforce study found that 36% of executives said unclear accountability complicates AI deployment. It argues that ownership, validation, exceptions, escalation, and override need to be clear inside the workflow itself. [5]
Those findings point to a simple problem: AI does not remove the need for a well-run organization. If ownership is unclear, a model does not decide who should own the work. If a handoff loses context, a more capable model does not automatically repair the handoff. If people are afraid to report problems, better monitoring does not guarantee that anyone will act on what it finds.
AI can make weak processes easier to see, but it can also help those same processes run faster. Adding capability without fixing the process underneath it can scale the problem along with the output.
Failure has to change something
NIST's AI Risk Management Framework makes this an organizational issue as much as a technical one. Its guidance calls for clear responsibility, a culture where important decisions can be questioned, and processes for reporting incidents and near misses so past failures can inform future work. [6]
A useful failure should therefore leave something behind. Maybe the workflow changes, ownership becomes clearer, a permission gets narrower, a test gets added, or a manual step is removed. The answer will differ by system, but the next attempt should not begin from exactly the same state as the last one.
In September, I wrote that failure should change the system that produced it rather than disappear into a log. The same idea applies to organizations: failure → evidence → correction → changed process → check whether it worked.
The goal is not a company where nothing ever goes wrong. The goal is a company where problems stay visible long enough to be understood and fixed, and where the same mistake becomes harder to repeat.
That matters for trustworthy AI because a control only matters if people are willing to act on what it finds. An organization can have policies, audits, dashboards, and approval gates and still fail to learn if bad news is buried, explained away, or pushed back onto the person who surfaced it. The strongest organizations will not be the ones that never fail. They will be the ones that learn fast enough to make the same failure harder to repeat.
Sources
- [1] Bradley Morrison, "The problem with workarounds is that they work: The persistence of resource shortages," Journal of Operations Management, 2015.
https://www.sciencedirect.com/science/article/pii/S0272696315000650 - [2] Lauren Eskreis-Winkler and Ayelet Fishbach, "Hidden failures," Organizational Behavior and Human Decision Processes, 2020.
https://www.sciencedirect.com/science/article/pii/S0749597818302747 - [3] Google Site Reliability Engineering, "Postmortem Culture: Learning from Failure."
https://sre.google/sre-book/postmortem-culture/
https://sre.google/workbook/postmortem-culture/ - [4] IBM Institute for Business Value, "Redesign for enterprise AI," July 2026.
https://www.ibm.com/downloads/documents/us-en/16ddab5aa1547458 - [5] IBM Institute for Business Value, "2026 CHRO Study: Designing the Thinking Organization," 2026.
https://www.ibm.com/thought-leadership/institute-business-value/c-suite-study/chro - [6] National Institute of Standards and Technology, AI Risk Management Framework, GOVERN function and Playbook.
https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
https://airc.nist.gov/airmf-resources/playbook/govern/