Each driver compounds the one before it
1. History: how many years, without gaps
Buyers pay for time. A dataset covering seven continuous years of operations is worth far more than a bigger one covering eighteen months, because the value is in the long arcs: how decisions played out, how customers behaved across cycles, how the business changed. Continuity matters too: a business that switched systems five times with nothing migrated has less than it thinks. Rule of thumb: every additional year of continuous history compounds the price; gaps discount it.
2. Coverage: how many systems, and whether they connect
A pile of emails is a commodity. Emails plus the chat threads plus the project tickets plus the code plus the CRM records about the same events is something else entirely: a connected record of how work happened. AI buyers specifically want that cross-system connective tissue, because it is what teaches models how organisations actually operate. Eight or more systems with overlapping history is the strong zone.
3. Rights: whether you can actually sell it
This is the driver that kills deals outright. You need the authority to license what you hold: employment terms that put work product with the company, customer terms that do not forbid it, and third-party content filtered out. Clean chain of title can double effective value because the buyer is not pricing in legal risk. Messy rights do not lower the price; they usually end the conversation. The good news: rights can often be repaired before going to market, and de-identification removes most privacy obstacles entirely.
4. Quality: real work, by real people, with outcomes
Buyers can tell the difference between an organisation's genuine operating record and a graveyard of auto-generated notifications. What lifts quality: real discussion and decision-making, documents that changed over time, visible outcomes (the deal closed, the feature shipped, the customer churned). What sinks it: bot noise, imported archives with no context, and data that is mostly duplicates of public content.
5. Rarity: what a model cannot learn anywhere else
The public internet is already in every model. Value concentrates in what is not: private operational reality, niche industry expertise, non-US perspectives, and languages of work underrepresented in existing corpora. An unglamorous logistics firm in Brisbane can outprice a generic tech startup here, because nothing like its data exists in any training set. Rarity is also why APAC supply carries a premium right now: buyers are actively balancing corpora dominated by US sources.
So what are the numbers?
Public comparables run from tens of thousands to hundreds of millions of dollars depending on scale and rarity: see what data actually sells for. For a single SMB or scale-up, honest answers range from "not much yet, and here is what would change that" to sums that are material against the company's annual profit. Anyone who quotes you a figure without looking at the five drivers above is guessing.
The assessment exists to replace guessing: a fixed-scope review of your estate against these five drivers, ending in a defensible range and a recommendation, including "leave it in the ground" when that is the truth.
General information, not financial or legal advice. Every dataset and every deal is different.