The loop that catches broken experiences and repairs them is heading toward automation. How far it ultimately goes comes down to one thing: the ground truth data informing it.
My team does something specific for Fullstory's largest customers. When an experience breaks, a checkout that fails, a form silently rejects input, rage clicks spike on a page that looked fine yesterday, we find the cause and help fix it. We investigate incidents, build the analysis, and sometimes write remediative code ourselves. Find and fix is the loop we live in, and Fullstory has been good at it since the early days of session replay.
The cycle I see is similar across the customer base: a routine conversion analysis flags an error firing during online checkout far more often than it should. Session intelligence says it has to do with a third-party backend issue and surfaces what the numbers had hidden: your customers are often so loyal that they invent their own workarounds just to place an order. But workarounds are not solutions, so a fix gets shipped, funnel health is confirmed in Fullstory, and teams make this type of hygiene a quarterly habit. In the process, millions in annualized revenue can often be recovered.
That’s the story for the boardroom. What I want to talk about is the machinery behind it, because that's where AI is changing things fastest.
The loop has always looked the same
Every find-and-fix win fundamentally moves through the same steps:
Detect: An anomaly trips a threshold: a conversion drop in Adobe, a spike in rage clicks, or an error in Fullstory.
Diagnose: Someone moves from knowing a problem exists to knowing its cause, using triage dashboards, session replay, and funnel analysis.
Prioritize: Product and engineering weigh business impact against effort and write it into the backlog.
Fix: Engineers write and test the change, QA verifies it, and the patch ships.
Refine: The team retunes the alert so the next occurrence is caught earlier, with fewer false positives.
Only the first step runs on its own. The rest still require human work. That’s the part AI is starting to change, and it is worth being precise about where, because the honest answer is it’s not moving at the same pace.
The finding half is already faster and cheaper
Diagnosis is where AI is earning its place today. The old version of this had one of my analysts jumping across disconnected platforms, building custom charts, filtering error logs, cross-referencing funnels, and watching replay after replay to spot a single pattern. Hours, sometimes days.
What used to take that long, an AI agent now does in seconds by querying Fullstory directly through MCP. It runs analyses across the full dataset in parallel, evaluates funnel permutations simultaneously, and catches subtle behavioral patterns the human eye routinely misses. This is running against real behavioral data today, not sitting on a roadmap.
Ground truth data is why it works
Here's what decides whether any of this holds up: an AI investigator is only as good as the data beneath it. Feed it incomplete data, and it does not hand you incomplete answers. It hands you confident, wrong ones.
That failure takes two shapes.
Tagged systems only capture events that someone thought to define in advance, so when a bug hits an untagged element or an invisible overlay blocking a button, the AI sees a user stop and blames intent rather than the defect it failed to see.
Sampled systems keep a fraction of traffic, so a low-frequency, high-impact bug on a specific device and browser either disappears or gets inflated into a false enterprise-wide trend. And when an agent hits a gap in either, it fills the blank with probability, blaming page-load speed for a drop-off that was really an un-instrumented API error.
Fullcapture is the system that removes the gap. Every scroll, hover, rage click, dead click, and field change is captured at the DOM level, without manual tagging, giving the AI the full context leading up to the anomaly rather than just a sketch. Most AI guesses at why a user failed. Fullstory gives it the evidence.
Ground truth is not a data-quality nicety. It is the difference between an agent that diagnoses the real problem and one that hallucinates a plausible one. That is the ballgame.
That same fidelity is why the economics work. Better data is not only more accurate, but it’s also cheaper to reason over. When working from sparse or unstructured data, an agent burns reasoning tokens chasing missing context down dead ends. When fed exact, deterministic event streams straight from the source, the agent reaches the root cause in far fewer loops. Resolution shifts from a cost measured in analyst hours to one measured in compute.
Where this is headed
So where does that leave the fix itself? My honest answer: writing and shipping production code today is still a human-led function, and at the enterprise scale I work in, it will be for the foreseeable future.
AI is genuinely strong at pinpointing a root cause and proposing a fix. Handing it the keys to deploy that fix unsupervised is a bar that our customers' security, compliance, and QA teams are not going to lower casually.
But the direction is not in question. One step at a time, every part of this loop that used to require a person is being handed to systems that can do it faster. The finding half is already there. The fixing half follows as trust is earned, and trust here is earned on data. A team hands more of the loop to AI only once it is confident the diagnosis is right every time, the guardrails catch a bad fix before it ships, and there is a track record to point to. Precision rests on complete, unsampled behavioral data.
From hunting for the problem to being handed it. From reports on what happened to systems that act on what is happening. Find and fix has been our craft for a decade, and AI is making it better, faster, and cheaper, starting now and widening from here. The ceiling on how far it goes is set by one thing: the quality of the behavioral data underneath it.
