By Brandon Aday
Founder, Aday Interactive, Inc. · 20+ years in enterprise systems and AI · Published November 16, 2026 · 9 min read
The short answer
An AI incident is not a breach. It is a wrong output that reached someone who acted on it. Most firm incident plans do not cover it. The plan you need answers four things: how you find out, how you stop it spreading, how you determine what else was affected, and who tells the client.
Your firm probably has a plan for a data breach and a plan for the office being unreachable. Now consider a different Tuesday. An AI-assisted summary went to a client last week with a materially wrong figure, the client made a decision on it, and you are finding out now because they mentioned it in passing. Nothing was breached. Nothing was down. And almost no professional firm has a written plan for what happens in the next hour.
A breach plan assumes unauthorized access, which frames everything after it: contain the access, assess exposure, notify per the applicable rule. A continuity plan assumes something stopped working, so the response is restoration.
An AI incident matches neither. Access was authorized. Nothing stopped working. The system performed normally and produced something wrong, and the error traveled through ordinary channels with ordinary approval. None of your existing triggers fire, which is precisely why these incidents surface late.
In most firms the answer is that a client tells you, which is the worst available option. Better detection is not exotic. Sampling AI-assisted outputs for review on a schedule catches patterns before they compound. Logging what the system produced means a question has an answer. A simple internal route for staff who think an output looked wrong matters more than any tooling, and it only works if raising a concern is treated as useful rather than obstructive.
The goal is modest and worth stating plainly: find out before the client does. That single sentence is the design requirement for everything in this section.
The instinct is to investigate. The correct first move is to stop production of anything similar. If an AI-assisted process produced one bad output, the same process is very likely still running and producing more.
So the first action is a pause on that workflow, not an inquiry into root cause. This is the step firms skip, because pausing feels disproportionate before you understand the problem. It is not disproportionate. It is the only action that stops the count of affected clients from rising while you work.
One wrong output is rarely alone. If the cause was a flawed source document, everything drawing on that document is suspect. If it was a prompt or configuration change, everything since that change is in scope. If it was a model update, the window is everything since the update.
Answering this requires records. A firm that logged what its AI produced can define the affected set in an afternoon. A firm that did not is reduced to reviewing everything or guessing, and guessing about client-facing errors is not a position any firm wants to be in when explaining itself later.
This decision must exist in writing in advance, because in the moment it is uncomfortable and delay is the default. Who makes the disclosure call, what threshold requires it, who speaks to the client, and what gets documented.
For regulated firms there may be obligations attached to the underlying error regardless of how it was produced. The AI involvement does not create a new duty, and it does not relieve an existing one. That question belongs to your professional responsibility counsel, and the time to ask is before an incident, not during.
The other half of readiness is finding weaknesses yourself. Give the system inputs designed to break it: questions whose answer does not exist, requests that push against policy, edge cases your process rarely sees, deliberately misleading phrasing.
You are testing two things. Does it refuse or flag when it should, and does it fail loudly rather than quietly? A system that produces a confident plausible answer to an unanswerable question has told you something important, and far better on a Tuesday afternoon than in a client matter.
Our AI Governance Audit includes the incident path: how a wrong output would be detected, who has authority to pause a workflow, what records exist to scope the blast radius, and who owns client communication.
Most firms discover the same three gaps. Nobody can pause the workflow without a technical person who may be unavailable. The logs needed to scope the problem were never turned on. And the disclosure decision has no named owner. All three are cheap to fix beforehand and expensive to resolve mid-incident.
Any AI-assisted output that was wrong in a way that mattered and reached someone who acted on it. It is distinct from a breach, because access was authorized and nothing failed visibly. That is why existing incident plans usually do not trigger.
Pause the workflow that produced it, before investigating. If the process generated one bad output it is probably still generating more. Stopping production first keeps the affected set from growing while you diagnose.
Logs. If you recorded what the AI produced and when, you can scope by cause: everything drawing on a flawed source, everything since a configuration change, everything since a model update. Without logs you are choosing between reviewing everything and guessing.
That depends on the underlying error and your professional obligations, not on the fact that AI was involved. AI does not create a new duty and does not relieve an existing one. Decide the threshold and the owner in advance with your professional responsibility counsel.
Deliberately trying to make it fail: unanswerable questions, policy-violating requests, misleading phrasing, rare edge cases. You are checking whether it refuses when it should and whether it fails loudly rather than producing a confident wrong answer. Doing this on purpose beats discovering it in a client matter.
Informational and educational purposes only
This article reflects Aday Interactive, Inc.'s views on marketing and technology architecture for professional-services firms as of the publication date. Aday Interactive, Inc. is not a law firm, and Brandon Aday is not a lawyer or legal counsel. It is not a substitute for advice from a licensed professional in your jurisdiction and does not create any professional relationship between you and Aday Interactive, Inc. Rules, statutes, checklists, and AI-engine behavior referenced here can change; verify the current versions and consult qualified counsel before acting. Where the article discusses regulated professional practice, those references are for informational and educational purposes only and do not constitute legal, medical, tax, financial, or investment advice. Consult a licensed professional in your jurisdiction before acting on anything you read here.