TL;DR (aka future GT, read this first next time): Two months into running sales ops, I realized the most trust-sensitive document in the company — the comp plan — still lives in a spreadsheet the whole industry knows is broken. 86–90% of spreadsheets contain an error, 22% of reps dispute their pay every year, and reps burn 25–50% of their month re-checking whether you'll pay them right. Meanwhile Stripe just spent $7 billion buying OpenRouter — a toll booth, not an AI company — because it wanted the reconciliation layer: the boring place where truth lives and nobody can bluff. Now the comp vendors are shipping agents that will happily calculate your reps' pay from whatever formula you hand them. The move: version the comp plan as code, build a deterministic oracle, and make every agent reconcile against it. Commission is a pure function — an agent's answer is either right or provably wrong. So I'm versioning the rules before the agents arrive.
The setup: nobody warned me about the spreadsheet
Two months ago I took a job running sales operations. Fifteen years in marketing and GTM, and I thought I knew what the gig was: the CRM, the forecast, the dashboards, the workflow that breaks at 11 PM and needs me.
Nobody warned me about the spreadsheet.
Every sales team has one. It's where the comp plan actually lives — not the polished PDF HR sends around, but the real one: the accelerators, the clawbacks, the territory splits, the ramp schedules, and the one cell nobody will admit they edited. It's the document where a single formula mistake costs someone their rent. And it's the document nobody wants to be the person who breaks.
Here's the part that should scare you, and it isn't even my spreadsheet.
Ray Panko at the University of Hawaii spent years auditing real-world spreadsheets. His number: 86 to 90% of them contain at least one error. The average cell error rate is 3.2%, and 82% of the mistakes are logic errors — the kind you cannot catch by eyeballing. WorldatWork puts the consequence in plain terms: 22% of sales reps file at least one commission dispute every year. Aberdeen found that reps spend 25 to 50% of their month on "shadow accounting" — building a private spreadsheet to check whether the company is paying them right.
Read that last one again. Your most expensive people are spending up to half their month manually verifying whether you're going to pay them correctly. That's not an ops problem. That's a trust problem wearing a spreadsheet costume.
There's a line from the comp people at Sales Cookie that is the whole thing in one sentence: "Reps don't quit over a percentage point. They quit when they no longer believe the company can pay them what they earned, on time, without a fight."
And the bill for that quit? DePaul's numbers: $114,957 to $150,000 to replace one B2B rep, and 6.2 months to fill the seat. So the cheapest retention lever in the building — make the rep's number and the company's number the same number — is sitting in a file that's 86–90% likely to be wrong somewhere.
Meanwhile, the market just wrote a check
A few weeks ago, Stripe bought OpenRouter for north of $7 billion.
OpenRouter owns no models and trains none. It routes requests across 400+ models and charges roughly a 5.5% markup on the pass-through. Stripe's payment take rate is about 0.36%. OpenRouter's is about 5.5%. The line everyone quoted: "Stripe just bought an interchange fee sitting on top of the fastest growing category of transaction volume on the internet."
Strip the jargon and Stripe didn't buy an AI company. It bought the reconciliation layer — the boring bit that answers who actually served the token, what did it cost, and who bills it. The layer where truth lives and nobody can bluff, because the answer is either right or wrong and you can prove which.
The market just paid frontier-lab money for the boring truth layer.
And the vendors are coming for mine
The comp-tool vendors figured out the same thing I did, only faster.
CaptivateIQ just shipped a portfolio of AI agents built for the comp and planning lifecycle — a Comp Ops Agent that catches payout anomalies, a Comp Builder Agent that debugs formulas, a Rev Planning Agent that maps accounts to territories. Salesforce's Spiff announced an MCP server shipping in H2 2026 so you can wire your comp data into Slack, Claude, Teams, or any MCP-enabled tool.
Read that slowly: the comp plan — the most trust-sensitive, most political, most "if this is wrong someone quits" artifact in a company — is being opened up to arbitrary AI agents as a surface.
Here's what the vendors won't say at the booth. Their agent will happily calculate your reps' commission from whatever formula you hand it. Hand it the spreadsheet, and it will automate the spreadsheet's errors at machine speed. An agent doesn't fix your source of truth. It amplifies it. Point a fast machine at a wrong rule and you get wrong paychecks, faster, with more confidence.
Salesforce's Spiff blog literally opens with "Don't DIY build your comp management tool with AI" — and they're right about SOX and audit trails and security. But notice what that argument does: it nudges their formula builder into being the source of truth by default. The fight isn't over who calculates. It's over who owns the rules the calculator enforces.
The one place in GTM where a machine can't bluff
Here's the insight that reframed my whole job.
Commission is a pure function. It's (the plan) × (the closed-won data). Hand an agent the plan and the deal records, and its output is either mathematically correct or it isn't — you can compute the right answer yourself, diff the two, and get a provable bug instead of a vibes disagreement.
Everywhere else in GTM, "did the agent do its job right" is a judgment call. Was the prospecting email any good? Did it stage the deal correctly? Those have vibes. The comp layer doesn't. It's the one place where "verify the agent" is a deterministic oracle — a failing test case — instead of a subjective review.
Which means two things at once:
- It's the cheapest place in the GTM stack to build trust in — or against — an agent. A diff against the source of truth is proof, not persuasion.
- It's the most catastrophic place to get it wrong. The machine can't bluff on your commission. It's either right or provably wrong. And the person who owns the proof owns the layer.
So I'm going to version it.
Comp plan as code
The plan becomes a file — comp-plan.yaml, or SQL, or whatever your team already version-controls. Every accelerator, clawback, ramp schedule, territory split, and cap is a rule, not a paragraph. A plan change is a commit. A rep's commission is a test case. A mid-quarter territory change is a merge conflict.
Then you build the oracle: given (a closed-won deal) → commission, seeded with the ugly edge cases — ramp reps, mid-quarter splits, accelerator tier-breaks, capped deals, split credits, prorations. Each one is an assertion.
And when the comp agent shows up for its demo this quarter — and it will, the vendors are racing in — I don't argue with the slides. I hand it the plan and the test set and diff its output against the oracle. Wrong means it doesn't ship. Right means it ships as a read-only view over my rules.
The buy/no-buy decision becomes a test result instead of a sales call.
And here's the payoff that matters to the org: it kills shadow accounting. The rep's private spreadsheet and the company's number become the same number, derived from the same committed rules, with git log as the audit trail. The rep stops re-verifying because there's nothing left to re-verify. That's the difference between "trust us" and "here's the diff."
The boring truth layer is the moat
Stripe spent $7 billion to own the reconciliation layer of the AI economy. CaptivateIQ and Spiff are racing to own the reconciliation layer of the revenue economy — comp, territory, quota, forecast. Both are the layer where truth lives and where a machine can't bluff.
I run sales ops. Comp, quota, and territory are part of what I own now. I sit on one corner of that layer — and it's the one corner where the check is deterministic.
I'm not going to let a vendor's formula builder become the source of truth for my reps' pay just because it got there first.
The market paid $7 billion for the AI version of this proof. I can own the revenue version for the price of a git init.
So here's my question for the sales ops and RevOps folks reading this: what's the one document in your stack you'd be scared to put in version control?
That's the one to version first.