The Mars Climate Orbiter entered the Martian atmosphere at the wrong altitude and burned up. The investigation found a single cause: Lockheed's ground software output thrust calculations in pound-force seconds, and NASA's navigation software expected newton-seconds. Both programs worked perfectly on their own terms. The spacecraft burned anyway because the data passing between them was in the wrong shape.
This is the situation most companies are walking into with AI in customer success.
The market has decided that the answer to scattered signal and agents working blind is to bolt an AI layer on top of what's already there: AI ticket triage and churn prediction, AI copilots that brief CSMs on the last six interactions and AI sentiment analysis on the ticket the agent is reading.
Each of these tools works; but they only work on the data they're given.
The model learns what the plumbing carries
A churn prediction model is a function that takes inputs and returns a probability. If the inputs are NPS responses and support ticket counts, the model predicts churn from those two signals and nothing else.
The model hasn't learned that the champion who signed the deal left three weeks ago, because nobody piped LinkedIn departure signals into the training data. It hasn't learned that the renewal contact stopped opening CSM emails two months ago, because the email engagement data lives in Outreach and the model reads from Snowflake. It hasn't learned that a billing dispute was resolved by finance without a ticket, because finance doesn't write to the customer record.
The model's confidence score looks the same, whether it's seeing the full picture or a sliver of it. That's the dangerous part. A human agent looking at a thin file at least knows the file is thin. An AI agent produces a 0.82 probability and moves on.
Wrong, faster
A junior CSM who doesn't know an account is in trouble takes a week to figure it out, and in that week something might happen that wakes them up: an email thread, a Slack ping from someone else on the team who picked up a weird tone on yesterday's call.
An AI copilot that doesn't know an account is in trouble confidently suggests the next-best-action within 400 milliseconds, surfaces the wrong talking points, and drafts a follow-up email that assumes everything is fine. The CSM hits send, the customer reads a note that ignores the actual state of the relationship, and the trajectory gets worse.
Speed and confidence on bad inputs is the same blindness, running at 10,000 requests per second.
Selection bias, now at scale
The previous piece covered how CS teams work with biased samples: NPS respondents who cluster at the extremes, support openers who self-select. The vast drifting middle produces almost no direct signal.
An LLM trained on ticket transcripts and survey responses learns the distribution of people who produce ticket transcripts and survey responses. The silent accounts are absent from the training data and absent from the inference data.
This is worse than a biased human, because a biased human occasionally remembers they have blind spots. A model trained on biased data has no such humility. It generalizes from the loud minority as if the loud minority were the population, and it does this in the tone of a confident advisor.
Duncan Watts made this point in Everything is Obvious in 2011: observational studies of public behavior miss the people who don't participate. CS analytics are observational studies of customer behavior. The people who don't participate are the people whose relationships are drifting.
Integration is upstream of intelligence
The companies getting real value from AI in their customer-facing teams have one thing in common that nobody wants to put on a conference slide: they did the boring work first.
They piped Zendesk ticket sentiment into the same system that reads Stripe payment status, HubSpot activity logs, Gainsight usage data, and Jira bug reports tagged to the account. They reconciled the customer ID across all five systems, which sounds trivial and takes four months. They built the dataset that a model could learn from, and then they put a model on top.
The companies that skipped this step and bought an AI copilot as a shortcut are now paying seat licenses for a tool that generates confident briefings on the same partial data their humans were already seeing. The briefings are prettier and the blind spots are identical.
W. Edwards Deming rebuilt Japanese manufacturing in the 1950s. He kept saying the same thing to every executive who asked him about quality: 94% of problems are in the system, 6% are in the people. Swapping the people (or the tools) without fixing the system changes the lighting while the terrain stays the same.
The map gets sharper, the ground stays the same
A dashboard is a map. A dashboard with an AI narrative layer bolted on is a sharper map. The ground underneath, which is the actual relationship with the actual customer, is still whatever it is.
If the signals that describe the ground are fragmented, the AI is drawing a detailed picture of a fragment. You can ask the AI to summarize, rank, prioritize, and draft follow-ups, and it will do all of those things well, on the fragment.
The customers who churned out of the fragment will still churn, and you still won't have seen it coming. The AI will produce a clean post-mortem that explains what the data showed, which wasn't much.
Where the work is
Before your next AI procurement cycle, ask your vendor what the model sees. If the answer is "whatever is in your current stack," and your current stack is five systems that don't share a customer record, the model sees what your humans see. You already know that isn't enough.
The companies doing interesting work in this space are going to spend 2026 pulling signal out of systems that were never meant to share it: LinkedIn stakeholder changes, Slack activity in shared customer channels, Zoom call transcripts, invoice aging, product telemetry at the feature level, email engagement from CSM outreach tools. Welding those into a single customer object is tedious, political, and unglamorous.
But it's the only thing that makes AI in customer success worth what it costs.
