Every operations team we’ve ever Scored has at least one. Sometimes a dozen. The spreadsheet that started as a list of three things, then twelve things, then a tab per client, then a colour-coding system only one person understands, then a macro, then a "please do not touch" warning in the filename. By the time we see it, it’s running a meaningful slice of the business and nobody’s quite sure who would be left holding the bag if it broke.
We open it first. Before the CRM. Before the ERP. Before the project management tool. Because the spreadsheet is where the truth lives.
The spreadsheet is honest in ways that the official systems aren’t. The CRM has the names and the deal stages. The spreadsheet has the actual notes about which client is wobbling, which invoice is two weeks late, which staff member can’t be put on which account. The ERP has the line items. The spreadsheet has the markup logic that the line items get massaged through before anyone presents them. The project tool has the tickets. The spreadsheet has the running list of things that aren’t tickets but are definitely going to become problems.
For an AI project, this matters more than people expect. If we’re going to deploy an agent that handles invoicing, we need the truth about how invoicing actually works. Not the policy. The truth. The truth is in the spreadsheet, which is why we read it before we read the policy.
The pattern under almost every load-bearing spreadsheet we’ve seen is the same: an institutional fact that the official systems can’t represent yet, captured by someone who’s smart enough to need it and doesn’t have the access or budget to put it somewhere proper. That fact is a clue. It tells you what the next iteration of the system needs to model. Sometimes the answer is “build a database for this.” Sometimes the answer is “this fact shouldn’t exist; we should change the policy that creates it.” Sometimes the answer is “this spreadsheet is fine; bless it, version-control it, and stop pretending it isn’t infrastructure.”
There’s a particular failure mode we’ve seen too often: the AI vendor who looks at the official systems, designs a beautiful agent against them, and then deploys to a team whose actual workflow runs through the unofficial one. The agent works perfectly on data that doesn’t reflect what the team is doing. The team ignores the agent and goes back to the spreadsheet. Six months later the project is quietly shelved.
A Score report from us almost always includes a “spreadsheets to surface” section. It’s not glamorous. It frequently surfaces a number of mild embarrassments. It is the cheapest piece of work we do and disproportionately the most useful.
If you want to know what’s actually going on in your operations, ask your team this question: “If you could only keep one file open all day, which one would it be?” The answer is the spreadsheet that’s pretending it’s not a database. That’s where we start.