In the pilot, not one first draft survived unchanged.
The tool I built pulled from call recordings, support cases, and the CRM to propose a health score for an account. Then it asked the CSM to review the score and the reasoning behind it. Every time, the CSM knew something the AI didn’t, and it changed the score, the reasoning, the scope of the account plan that came next, or all of them.
Had those drafts gone straight to the CRM, they would have arrived looking precise, with a number, a rationale, and a plan. Actions would have been taken against a plan that was not correct resulting in, at best, no change, and, at worst, real damage being done. The review step is what caught it, and it left a record of who decided what.
The Reviewer Matters More Than the Model
I was at the Pavilion GTM2026 Summit in New York last week, and the word I heard most was judgment. It deserves more than a passing mention, because a health score looks like a calculation but is really a judgment about a relationship. Is the champion still engaged? Does a quiet quarter mean stable or checked out? Data can carry traces of those answers. It can’t make the call.
Few of the people I talked to had AI doing something that clearly moved revenue, margin, or EBITDA, and most were still working out where it earns its place. A lot of the worry came back to the same thing. AI can’t make the calls that take real judgment, and there’s pressure to let it make them anyway.
The other half of that risk is easier to miss. Putting AI in front of a person who lacks judgment is its own problem. That person can’t tell a sound health score from a shaky one, and a confident-sounding model makes approving easy.
Companies tend to cover both risks with one sentence: A human is in the loop. That sentence says nothing about whether the human has the judgment to catch a bad answer, or whether anything records what they decided.
A reviewer who changes nothing has told you something, and it isn’t always that the AI was right.
A Decision With a Name on It
In the tool, the CSM reviewed the score, challenged it with their own reasoning, and made the final decision inside the AI. By the end, the AI and the CSM shared an understanding of the score and why it was set. The tool then interviewed the CSM, drafted a Customer Success Plan, and ran the same process again: Review, challenge, decide. Only then did it update the CRM with the final score and plan.
That sequence matters for accountability. The decision became a position a specific person had taken, with their reasoning attached to it, instead of a number the system produced. When a renewal goes sideways six months later, you can ask who decided and why, and there’s an actual answer.
It also exposes weak judgment. When reasoning has to be put into words, a reviewer who doesn’t understand the account can’t hide behind a click.
This is exactly the kind of gap our AI Revenue Risk Assessment is built to surface, before a confident score nobody owns drives a renewal decision: see how it works →
Training Both Sides
I’d expect the first pass on almost any account to look like the pilot did. A few passes later, I’d expect two things to change. The CSM learns what information to capture in the CRM and other data sources so the AI gets closer to the real answer, and the AI learns how that CSM interprets each input when setting a score.
I won’t say the AI will never get it right. I do expect it takes time and a team willing to train it.
That also changes what you should measure. If I were tracking whether this works, I’d watch how often a CSM changes the AI’s draft, and why. A high rate early on is healthy. A rate that stays high means the data feeding the AI is missing something. A rate that falls to zero quickly deserves suspicion, because it can mean people stopped looking.
That last part needs an owner, too. If I were running this, someone would read the overrides on a regular cadence and decide which point to missing data worth capturing and which are one CSM’s view of one account. Training the AI isn’t a one-time setup.
Each correction is also a small data-quality report. When a CSM keeps fixing the same input, someone has finally noticed that input is wrong or incomplete.
Where This Leaves the ROI Question
Every post in this series has pointed at some version of the same blind spot. This one is about who owns the call.
Ask the person who last approved an AI-generated score or recommendation what they changed. If the answer is nothing, ask whether that’s because the AI was right or because nobody expected them to check. Then ask them to explain the decision in writing. If they can’t, nobody is accountable for it.
This is the sixth post in an eight-post series on why most B2B SaaS companies can’t answer the AI ROI question, and what closes the gap. If you want a scored, five-minute read on where your own organization stands, the AI Revenue Risk Assessment gives you the diagnostic: take the assessment →
Andrea Mulligan is a B2B SaaS executive and advisor with 30 years of experience building Customer Success, Professional Services, and GTM organizations. She works with PE-backed and growth-stage companies on CS transformation, revenue retention strategy, and post-sale model design. Start a conversation →