You know the sunk cost fallacy: you keep funding a bad project because of what you’ve already spent on it.
Meet its uglier sibling. The sunk data fallacy: you drag ten years of legacy data into every new system because of what it cost you to collect — never asking whether it’s fit to make decisions on.
Every platform migration in this industry begins the same way. Somebody says “obviously we bring the historical data over.” Nobody asks the follow-up question. And that reflex, more than any feature gap, is what quietly caps how much an AI-era platform can ever do for you.
The Half of the Quote Nobody Finishes
In 2006, mathematician Clive Humby coined the most overused line in tech: data is the new oil. You’ve heard it in a hundred pitch decks.
Here’s what almost nobody quotes: the second half. The full point, expanded by Michael Palmer, was that data is like oil because it’s “valuable, but if unrefined it cannot really be used.” Crude has to be refined before it powers anything. Data has to be clean, structured, and correctly attributed before it can drive a decision.
The industry memorized the bumper sticker and skipped the point. Data isn’t valuable because you have a lot of it. It’s valuable when it’s refined. A decade of legacy CRM exports isn’t a reserve of oil. Most of it is crude — and some of it is sludge.
What’s Actually in Your Historical Data
Be honest about what a ten-year-old direct-response dataset contains.
Campaign taxonomies from three marketing hires ago, where “FB-Q3-TEST-FINAL-v2” meant something to someone who left in 2019. UTM conventions that changed four times. Affiliate attributions that were negotiated, not measured. The same customer existing as three records with two spellings and conflicting addresses. Orders whose “source” field was whatever the call center picked from a dropdown. Refunds logged in one system, chargebacks in another, and a subscription status field that stopped being trustworthy after the second replatform.
None of this made your old system useless — a ledger tolerates mess, because humans interpret ledgers. Humans squint, remember context, and mentally correct for the weird year when everything was tagged wrong.
Machines don’t squint.
Why Reconstruction Fails — Even With an Army of Devs
The standard answer is “we’ll map it during migration.” The research on how that goes is grim. Depending on which study you believe, somewhere between four in ten and eight in ten data migration projects fail outright or blow through their budgets and timelines — the most-cited figure puts it at 83%, and even the most conservative survey found nearly 40% failing with average overruns in the hundreds of thousands of dollars. Cost overruns average around 30%, schedule overruns around 41%. No credible study puts the failure rate below a third.
And that’s for moving records. Reconstructing attribution — which touchpoint gets credit for which revenue, across systems that defined those concepts differently — is the hardest version of the problem. It means rebuilding the semantics of decisions made years ago by people who no longer work for you, encoded in conventions nobody documented. You can throw a dev team at it for months. What comes out the other side isn’t truth. It’s a plausible-looking guess wearing truth’s clothing.
“Fine,” says the modern operator, “we’ll have AI clean it up.”
That makes it worse. AI doesn’t fix bad data — it accelerates it. A model handed confidently-wrong attribution can’t tell it apart from correct attribution; it just learns the wrong pattern faster and applies it everywhere. This isn’t theoretical. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that aren’t supported by AI-ready data, and found that 63% of organizations either lack the right data practices for AI or aren’t sure they have them. Half of generative AI projects die after proof of concept, with poor data quality named as a leading killer. And as AI agents flooded into companies — adoption quadrupling in two quarters — data quality concerns jumped from 56% to 82% of organizations, becoming the single biggest barrier to getting value from AI at all.
Read that trajectory carefully. The more decisions your software makes on its own, the more expensive every polluted record becomes.
The Part That Isn’t Your Fault
Hoarding the data was the right call when you made it. Storage was cheap, “keep everything” was the universal advice, and for fifteen years the worst thing old data could do was sit there. The cost model changed underneath you: the expense is no longer storing data — it’s every automated decision made on top of it. Nobody sent a memo when that flipped. You’re not behind; the ground moved.
Why RevCent Starts You Clean
RevCent does not import your historical data. That’s not a missing feature. It’s the design.
Your history stays exactly where it is — in the system that created it, or in your exports, archived and readable whenever a human wants to look something up. What it doesn’t get to do is flow into the engine that makes forward decisions.
Because from day one on RevCent, every record is born clean. Every transaction lands with its real decline code. Every sale carries attribution captured at the moment it happened — not reconstructed from a spreadsheet later. Every customer is one customer. One schema, one data layer, no inherited guesswork. When an AI agent works a declined rebill or routes a transaction, it reasons over data that has never been anything but true.
That’s what “clean data is gold” actually means in practice. It’s not hygiene for its own sake. AI makes the best decisions for you exactly to the degree the substrate underneath it is pure — and purity, once polluted, doesn’t come back. We’d rather hand your agents three months of gold than ten years of sludge.
The compounding works in your favor faster than you’d think. A subscription business runs in cycles — after your first full rebill cycle on RevCent, the engine is already operating on real retry outcomes, real decline reasons, real attribution. From that point, every month forward, your clean dataset gets more valuable while the old crude sits safely in the archive where it belongs.
The sunk data fallacy says you can’t afford to leave your old data behind. Flip it. The one thing an AI-native platform can’t afford is to bring it along.
See how RevCent works — starting from zero, and treating that as the feature it is.