Skip to content

Deal economics

The AI data valuation milestones that hold up when a committee checks

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

AI data valuation milestones should be completed events that an outside party could have refused: revenue recognized after a buyer accepted a delivery, a second accepted delivery at similar cost, licenses covering the exact records and uses, and a base no single buyer or supplier can erase. Letters of intent, records held, supplier sign-ups and headline contract value tell a story but cannot carry a number.

Value drivers

AI data valuation milestones

This page covers the AI data valuation milestones that hold up when a committee checks.Illustrative

Key takeaways

  • A milestone supports value when it turns an assumption into an observed input, and IFRS 13 ranks observed inputs above unobservable ones.
  • Accepted delivery is the hinge event: IFRS 15 treats customer acceptance as a sign that control has passed, and a record of similar accepted deliveries can support that conclusion.
  • Repeatability means a second buyer or a renewal for the same dataset category at a known cost, not a larger first order.
  • Rights hold only when signed licenses cover the exact records, uses, sublicensing and term; public dataset metadata omitted licenses more than 70 percent of the time in one audit.
  • Contingency and concentration multiply: in the illustrative grid, 100 units of signed value with half awaiting acceptance and 60 percent from one buyer leaves 20 units under the joint downside.

What counts as a valuation milestone for an AI data company?

A valuation milestone is a completed, dated event that someone outside the company could have refused and that changes an input to the valuation: the revenue base, how long it lasts or how risky it is. For an AI data company, five events qualify: revenue recognized after a buyer accepted a delivery, a repeat accepted delivery, licensing rights that survive a document check, a buyer and supplier base no single party can erase, and prices actually paid by unrelated buyers.

Valuation practice already leans on milestones for young companies. The International Private Equity and Venture Capital Valuation Guidelines (December 2022 edition) say many seed and early-stage investments are valued with a milestone approach or scenario analysis, using milestones agreed when the investment is made. Their examples are generic: revenue growth, profitability expectations, cash burn and covenant compliance on the financial side; phases of development, testing cycles and patents on the technical side; customer surveys, testing phases, market introduction and market share on the commercial side.

A data licensing company needs a sharper list, because several of those generic markers can be met without one buyer accepting one dataset. This memo sorts candidate milestones into those a valuation can lean on and those it cannot, prices the cost of leaning on the wrong ones with an illustrative grid, and closes with a risk register a committee can use as written. It estimates no company's value and quotes no valuation multiple.

Why observed transactions outrank forecasts in a data company valuation

Because fair value accounting ranks inputs by how observable they are, and almost nothing about an early data company is observable except its own completed transactions. IFRS 13 gives the highest priority to unadjusted quoted prices for identical assets in active markets and the lowest to unobservable inputs, which it allows only to the extent that relevant observable inputs are unavailable. A measurement takes the level of the least observable input that is significant to it.

Private data companies have no quoted price, and data licensing terms are often confidential, so disclosed comparable deals are scarce. The forecast inside a valuation model, how many datasets at what price, accepted how often and at what cost, is therefore an unobservable input by construction. Each completed transaction converts part of that assumption into something observed: a price someone paid, an acceptance someone signed, a renewal someone chose.

Two other public frameworks point the same way. The US regulation on valuing private company stock for deferred compensation, 26 CFR 1.409A-1, names recent arm's length transactions involving the stock among the factors a reasonable method weighs, beside asset values, discounted future cash flows and comparable companies, and judges reasonableness on the facts at the valuation date. IPEV's guidelines say the price of a recent funding round is not a standalone valuation technique; it calibrates a method whose inputs must be updated at each later measurement date. A round priced on a story does not become evidence by being repeated.

The lean-on ledger: five completed events and the input each one moves

Each row names a completed event, the valuation input it can move, the document that proves it and the counterfeit that often stands in for it. A pre-revenue data startup will have few of these rows filled, which is fine as long as the memo prices the empty ones as assumptions rather than achievements. Read across a row before accepting a milestone in a memo:

The lean-on ledger: five completed events and the input each one moves
Completed eventValuation input it can moveEvidence that holdsCounterfeit versionWhy the counterfeit fails
Revenue recognized after acceptanceRevenue baseThe buyer's written acceptance against a stated test, the invoice, cash received and the policy that recognized the revenueSigned contract value, or invoices still inside an acceptance or refund windowThe buyer can still reject, so the amount remains contingent
Repeat accepted deliveryHow long revenue lastsA second unrelated buyer accepting the same dataset category, or a renewal, at a cost per dataset within the first delivery's rangeA larger first order, or a second order from the same pilot teamSize shows appetite, not that the process repeats for someone else
Rights that holdRisk applied to every forecast yearSigned licenses naming the records, permitted AI uses, sublicensing and term, checked against the source systems' own termsSupplier sign-ups, data-room uploads or a license field in metadataUpload access and metadata are not licensing authority
Spread buyers and suppliersRisk and durabilityQuarterly tables of revenue from the largest buyer and supply from the largest supplier, as sharesA logo slide or a count of conversationsLogos carry no denominators and no revenue shares
Executed arm's-length pricesPrice assumption in the modelExecuted agreements with unrelated buyers showing price per dataset or per record classA rate card, list price or quoted rangeA price nobody has paid is still an assumption

Five signals that fill a deck but cannot carry a number

These signals reflect real activity, but none is an event an outside party could refuse at a stated price, so none should move a valuation input on its own:

  • Letters of intent and memoranda of understanding. They record interest without a price, an acceptance test or an obligation; carry them as pipeline, never as revenue base.
  • Records or terabytes held. Volume without signed rights and a buyer's acceptance test is an inventory of unknown value, and for a company that sources on request it may not be an inventory at all.
  • Supplier applications and sign-ups. A company that applied has not yet granted rights to specific records; count signed licenses by record set instead.
  • Headline contract value. Payments that wait on acceptance, later milestones or refund windows are contingent, and the grid further down shows how fast contingency compounds with concentration.
  • Generic commercial markers. Customer surveys, testing phases and market introduction appear among IPEV's examples, yet a data company can pass all three before anyone who paid has accepted a dataset.

Why the second accepted delivery is worth more than the first

The first accepted delivery proves a dataset could be made once; the second proves the company can make it again, for someone else, at a known cost, which is the assumption every later year of a valuation rests on.

Revenue accounting already treats a track record as evidence. IFRS 15 recognizes revenue when a performance obligation is satisfied by transferring control of the promised good or service, and lists customer acceptance among the indicators that control has passed. Its application guidance on acceptance clauses (paragraphs B83 to B86) explains that such clauses can let a customer cancel or require remedial work when a delivery misses agreed specifications, and that experience with similar contracts can support a conclusion that a delivery meets them. One accepted dataset gives a company no such experience; several accepted deliveries of one category give it the start of one.

A repeatability test that a committee can check has four parts: the same dataset category; a different, unrelated buyer, or a renewal by the same buyer after the first license ran; a cost per accepted dataset within the range of the first delivery; and no step that only a founder performed. Two deliveries pass a weak version of this test. A valuation that extrapolates several years from them still rests mostly on assumption, and the memo should say so in those words.

Rights that hold: the milestone most often assumed and least often checked

A licensing right supports value only when the signed agreement covers the exact records delivered, the AI uses the buyer intends, any sublicensing, and a term the revenue model does not outrun. Checking those four against the agreements themselves, not against a summary slide, is the cheapest diligence step and the one most often skipped.

Public dataset metadata shows why summaries fail. The Data Provenance Initiative audited 1,858 datasets across 44 widely used finetuning collections and reported license omission above 70 percent, and error rates above 50 percent, on widely used dataset hosting sites (arXiv 2310.16787, 2023). Private business records carry the same risk in another form: a supplier can own its records and still be bound by customer contracts, platform terms or privacy notices that limit outside use.

Rights also have to survive the company's own structure. Ask who signed for each supplier and with what authority, whether each license passes to an acquirer of the data company, and whether exclusivity promised to one buyer conflicts with a license sold to another. Each failure turns recognized revenue into a dispute rather than a smaller number, which is why rights belong in the risk input of every forecast year rather than in a one-time checklist.

How much of a headline figure survives contingency and concentration?

Less than either stress alone suggests, because the two multiply. In the illustrative grid below, a company reporting 100 units of signed contract value with half still awaiting acceptance keeps 50 units if the contingent half fails, 40 units if a buyer holding 60 percent of value does not return, and 20 units if both happen.

Listed companies already disclose the concentration half. IFRS 8 requires an entity to disclose when revenue from a single external customer reaches 10 percent or more of its revenue, with the total from each such customer, though not the customer's name. Asking a private data company for the same table, for buyers and for suppliers, costs nothing and replaces the grid's second axis with a real number.

The model is deliberately simple: surviving units equal (100 minus the contingent percent) times (100 minus the largest buyer's percent), divided by 100. It assumes the largest buyer's share applies equally to contingent and recognized value and that the two stresses are independent. Correlated failure, where that buyer's own policy change also blocks pending acceptances, would make the joint downside worse than the grid shows. Every input is illustrative and editable. Units of 100 signed value that survive both stresses:

How much of a headline figure survives contingency and concentration?
Largest buyer's share of value (Illustrative)0% awaiting acceptance25% awaiting acceptance50% awaiting acceptance75% awaiting acceptance
20%80604020
40%60453015
60%40302010
80%2015105
A four-by-four grid of units surviving from 100 units of signed value, with the largest buyer's share from 20 to 80 percent down the side and the share awaiting acceptance from 0 to 75 percent across the top, highlighting 20 units left when half is contingent and one buyer holds 60 percent.
Illustrative: contingency and concentration multiply. Surviving units = (100 - contingent %) x (100 - largest-buyer %) / 100.

What could make milestone-based valuation wrong?

Milestones shrink the share of a valuation that rests on assumption; they do not remove market risk, and they can be manufactured. IPEV's guidance cuts both ways: missing a milestone may indicate a decrease in value and exceeding one an increase, depending on the facts and circumstances, so a milestone accepted on weak evidence also sets up a later markdown.

Hold these six exposures next to every milestone the committee accepts. The downside scenarios are editable starting points, not estimates for any company:

What could make milestone-based valuation wrong?
ExposureHow it fools a milestoneMonitoring signalDownside scenario (editable)
Friendly or related buyersAcceptance and repeat purchase come from parties tied to the company or its investorsShare of accepted revenue from unrelated buyers; direct calls to counterpartiesExclude related revenue and rerun the grid
Timed transactionsOrders are pulled forward to land before a financingOrder and acceptance dates clustered before a round; renewals at lower pricesNext period's revenue falls by the pulled-forward amount
Correlated buyer riskOne buyer's policy change hits demand, acceptance and concentration at onceThe largest buyer's share each quarter and its published procurement or data policiesUse the joint-downside cell of the grid, or worse
Rights challengeA source system's terms, a customer contract or a ruling narrows permitted useRights review findings per dataset; open disputes and withdrawal requestsAffected datasets withdrawn and refunds paid inside acceptance windows
SubstitutionSynthetic or public data starts passing a buyer's acceptance test for the categoryBuyer acceptance tests that admit synthetic data; falling repeat rates in the categoryPrice per dataset in that category falls; repeat assumption removed
Small samplesTwo deliveries are read as a repeatable processCount of accepted deliveries per category and per buyerTreat as single-event evidence and widen the scenario range

Questions to put to any AI data company's valuation case

Ask for the document behind each answer, and apply the same list to SourceX.

SourceX is one test case for these questions, not evidence for them. It turns business records that a company selects into licensed AI datasets and runs the qualification, rights review, contracting, preparation, delivery and payment steps around them. Because it sources datasets when a buyer asks instead of holding them in stock, its valuation case would rest on completed transactions rather than on an inventory, and the term and any exclusivity of each license sit in that deal's signed agreement. SourceX has not published revenue, acceptance, concentration or rights figures, and none appear here; it provides company and commercial metrics to investors directly. Its current fundraising is a Rule 506(c) offering open only to verified accredited investors, and materials can be requested at sourcex.si/investors. Questions worth bringing to any company in this space:

  • What share of last year's revenue was recognized after written acceptance, and what share still sits inside acceptance or refund windows?
  • Which dataset categories have been accepted by more than one unrelated buyer, and what did each accepted delivery cost to produce?
  • For each delivered dataset, which signed license covers the records, the permitted uses, sublicensing and term, and who reviewed it?
  • What share of revenue came from the largest buyer, and what share of supply from the largest supplier, in each of the last four quarters?
  • Which executed prices, not quotes, support the price assumption in your model?
  • Which milestone in your last financing plan was missed, and what changed in the valuation case afterward?

How this memo was built, and what it did not measure

The lean-on ledger, the five deck signals, the downside grid and the risk register were drawn up for this article on 2026-10-27. Public sources: IPEV's December 2022 valuation guidelines; IFRS 13, IFRS 15 and IFRS 8 as published on ifrs.org; 26 CFR 1.409A-1 on eCFR; and the Data Provenance Initiative audit on arXiv. Each was checked through a search restricted to its publisher's domain on 2026-10-27, so a reviewer should open each page before release. A December 2025 IPEV edition exists; its milestone examples differ slightly, and only examples present in both were used.

The grid applies one formula, surviving units = (100 - contingent percent) x (100 - largest-buyer percent) / 100, to inputs chosen so that every cell is an exact integer; there is no randomness and no rounding. It assumes the two stresses are independent and that losing the largest buyer removes its share of both contingent and recognized value. The script is kept with this article's working files for review.

Not measured: the milestones, revenue, concentration or rights position of SourceX or any other private company; how public or private markets price data companies; and whether the accounting for any specific data contract follows the IFRS 15 reading above, which an accountant should test against the contract itself.

Frequently asked questions

Which milestone matters most for an early AI data company's valuation?

A delivery that a paying buyer accepted against a written test, because it turns price, acceptance and cost from assumptions into observed facts at once. Repeat purchase, concentration and rights are all measured against that first accepted delivery. Before it exists, a valuation rests almost entirely on unobservable inputs.

Do letters of intent from AI developers support a higher valuation?

Not on their own. A letter of intent records interest without a price, an acceptance test or an obligation, so it changes no input a valuation model uses. Carry it as pipeline and ask what would convert it: a signed order with an acceptance clause and a stated price.

How should investors treat revenue that still depends on buyer acceptance?

As contingent until the buyer accepts. Under IFRS 15, acceptance clauses can delay the point at which control passes, so ask what share of signed value sits inside acceptance or refund windows. Then run that share through a downside grid together with the largest buyer's share.

Is buyer concentration disqualifying for a young data licensing company?

Not by itself; a young company can depend on a few buyers, and the question is whether the valuation prices that dependence. Ask for the largest buyer's share of revenue each quarter, in the form IFRS 8 uses for listed companies, and model the case where that buyer does not return.

Does SourceX publish its milestones or metrics?

No. SourceX has not published revenue, acceptance, concentration or rights metrics, and this article contains none. It shares company and commercial metrics with investors directly, and materials can be requested at sourcex.si/investors.

Sources

  • IPEV's guidelines say many seed and early-stage investments are valued with a milestone approach or scenario analysis, give financial, technical and commercial milestone examples, treat the price of a recent round as a calibration point rather than a standalone technique, and note that a missed milestone may indicate a decrease in value. Source
  • IFRS 13 gives the highest priority to unadjusted quoted prices in active markets and the lowest to unobservable inputs, which are used only to the extent observable inputs are unavailable. Source
  • 26 CFR 1.409A-1 lists recent arm's length transactions in the stock among the factors a reasonable valuation method weighs, beside asset values, discounted cash flows and comparable companies, judged on facts at the valuation date. Source
  • IFRS 15 recognizes revenue when control of a promised good or service transfers, treats customer acceptance as an indicator of transfer, and discusses acceptance clauses and the role of experience with similar contracts in paragraphs B83 to B86. Source
  • The Data Provenance Initiative audited 1,858 datasets in 44 collections and found license omission above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
  • IFRS 8 requires disclosure when revenue from a single external customer is 10 percent or more of an entity's revenue, with the total from each such customer, without naming it. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies