You know that feeling you get every once and while when an account or two feel off? Nothing's escalated and nobody's complained, they're just quieter than they were in the spring. There's data that can confirm or kill that hunch, and it's sitting in your product analytics. The problem is it rarely reaches you in a way you can act on before the renewal call, it just quietly sits there, telling a story that no one is listening to.
Let's look at closing that gap. It might only take an afternoon, an export you can already pull yourself, and a Claude capability that's already switched on in most accounts.
The question at hand is: which accounts broke their own usage pattern?
Every account has its own normal
Watching how fast usage is changing beats watching how much of it there is. Accoil's 2026 guide to health scoring draws the line with two examples: an account that logged in 30 percent less this week is noise, while an account whose primary admin has gone dark for two weeks against an eighteen-month pattern is a warning. Same file, two different questions, and only one of them finds the account while there's still time to do something.
That's also why fixed cutoffs disappoint. A rule like "flag anything under 40 percent adoption" fires every December when half your customers are on holiday, and after enough false alarms people stop reading the alerts entirely.
You may already know the baselines that would fix this. If you know one account runs reports every Monday morning, and that another only shows up around quarterly planning, then you likely have a strong starting point. None of that, by the way, is generally found in a dashboard.
Behavioural signals like falling login frequency, abandoned features, and rising support escalations tend to appear 30 to 60 days before a customer decides to leave, according to Quivly's 2026 inventory of post-sales signals. So creating a monthly export would fall into that window and leave you plenty of time to be proactive and check-in with your flagged accounts.
Letting Claude compute the numbers instead of reading them
Claude has a private computing environment inside claude.ai: a sandbox where it writes and runs code, usually Python, against a file you upload. Anthropic calls the feature code execution and file creation, and it handles CSV and TSV files up to 30MB. On Free, Pro, and Max it's on by default, and the toggle lives at Settings then Capabilities if you want to confirm it. Team and Enterprise organizations get it on by default as well, with owners able to switch it off in Organization settings.
The difference this makes is big. Without the sandbox, a model reads your forty thousand rows the way it reads a paragraph and gives you an answer that sounds right. With it, Claude writes a script, runs the math, and hands you a number you can check by asking to see the code. Telling a plausible answer from a competent one is a lot easier when there's a script sitting behind it.
The file does some of the work as well. Put real column headers in the first row, use YYYY-MM-DD dates, and export to CSV rather than XLSX so your rows aren't competing with formatting for room. Most product analytics tools already give you something close to what you need: an account identifier, a user, a date, an event or feature name, and a count. Shapes vary between tools, which is fine, because you're going to tell Claude what your columns mean anyway.
Writing the prompt that produces the account shortlist
Brief it the way you'd brief an analyst joining you on Monday: what the columns are, what you want computed, and what shape you want back.
This CSV is 18 months of product usage data. Columns: account_id,
user_email, event_date, feature_name, event_count.
Load it with pandas and compute, for each account:
- distinct active users per month for the last three months, compared
against that account's own monthly average for the prior twelve
- distinct features used per month over the same periods, same comparison
- days since the account's most recent event, and days since the most
recent event for each of its top three users by historical volume
Return a table of the accounts whose recent pattern departs furthest from
their own baseline, sorted by the size of the change rather than by
absolute usage. Show me the script you ran.
The prompt is doing something specific: it tells Claude to measure every account against that account's own history, and to show its work. What comes back is a shortlist of accounts whose behaviour actually changed, built from your data and your definition of normal, short enough to work through in an afternoon.
Once you've seen the first output, add one more step. Ask Claude to keep the working prompt in a Project alongside the export, because project files stay reachable from that same computing environment. Next month's run becomes an upload and a sentence instead of a rebuild.
Reading the shortlist against what you know
A shortlist is approachable and will help you take action. Health scoring can be misleading when the product usage carries too much of the weight, as Union Square Consulting points out: a team can log in every day and still be at risk because they never got to the capability that produces their value, and an enterprise account judged against SMB patterns looks unhealthy when it's fine.
You also know the things the file has no way of knowing. Maybe the primary user left in March. That the December dip was a holiday and the June one wasn't. That the account nobody hears from is happily self-sufficient with a power user rather than halfway out the door. The shortlist tells you where the numbers moved, and then your intimate knowledge of your accounts is going to be the deciding factor on who needs a call this week.
Which is roughly how the discipline describes health anyway. Gainsight's 2026 metrics guidance treats a health picture as a blend of usage, support history, adoption, and the CSM's own qualitative read. You're supplying the last one, and it's the part nobody can buy.
One practical note, since this is customer data. The sandbox is isolated per user, and Anthropic recommends watching the run and stopping it if Claude starts doing something you didn't ask for. On Team and Enterprise plans, owners control whether that sandbox can reach the network at all, and turning egress off is the configuration Anthropic describes as the most secure. If your export carries anything sensitive, have the five-minute conversation with whoever owns your workspace before the file goes up.
Run it once against last quarter's export. Some of what comes back you'll have flagged already. The account you hadn't is what the afternoon buys you.