Deal economics
Workflow diversity data defensibility: count decisions, not logos
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Workflow diversity makes a data product defensible only when it means many distinct decisions that buyers' tasks need, spread so that a rival would have to sign many suppliers to match them, under rights consistent enough to combine. A supplier-logo count measures none of this. Count decision families, compute the smallest supplier set that would replicate the coverage, and check whether buyers return for breadth.
Value drivers
Workflow diversity data
Key takeaways
- Thirty suppliers running the same workflow add volume, not coverage; the unit that matters is the distinct decision family.
- The replacement set, the fewest suppliers a rival must sign to match most of the coverage, turns breadth into a number a rival's plan can be tested against.
- Training research rewards curated variety: the LIMA authors chose 1,000 examples for quality and diversity, and raters preferred the result to a reproduction trained on 52,000.
- The strongest objection is substitution: most buyers need one workflow and can source it directly from one large supplier.
- Breadth defends a product only where the rights behind each decision family can be combined under one scope a buyer accepts.
Can workflow diversity create a defensible data product?
Yes, but only one kind of diversity does, and supplier counts do not measure it. Workflow diversity defends a data product when it covers many distinct business decisions that buyers' tasks require, when those decisions are spread across suppliers so that no small group of them could be signed to reproduce the coverage, and when the rights behind each decision family can be combined under one scope.
A workflow is a repeatable sequence of business work, such as claims intake, technician dispatch or vendor invoice approval. A decision is a choice point inside it, with inputs, options, an action and an eventual outcome: approve or deny the claim, send the senior technician or the nearest one, pay the invoice or hold it for a mismatch. A decision family groups decisions that follow the same rule structure, so a model trained on one member would be expected to handle the others.
The investor question has two halves: is breadth measured in decision families hard to copy, and do buyers pay for it? The rest of this article turns both halves into tests, starting with the number most pitch decks show instead.
Why supplier logos are the wrong denominator
A logo count measures how many companies signed, not how many different decisions the product can teach or test. Thirty residential HVAC contractors running the same dispatch software produce one decision family thirty times over, and a rival who signs the two largest of them matches most of what the logos imply.
Repetition is a documented problem in training data. The authors of Deduplicating Training Data Makes Language Models Better report that one 50-word sequence appears 60,000 times in the C4 dataset, and found that models trained on deduplicated data reached similar or better perplexity while memorizing less. Near-identical workflows from near-identical suppliers are the business-records version of the same problem: more rows, little new information.
Curated variety, by contrast, has measurable value. The LIMA team fine-tuned a 65-billion-parameter model on 1,000 examples, 750 selected from community question-and-answer sites for quality and diversity and 250 written by hand for task variety, and human raters preferred its answers to those of a same-size model trained on 52,000 Alpaca examples. For a data product, the lesson is to count what is different, not what is plentiful.
What counts as one distinct decision family?
A decision family is a set of choices that share a trigger, the inputs that matter, the options available and the kind of outcome that judges them. Two companies' refund approvals belong to the same family if the same rule structure decides them, even when their software, wording and volumes differ.
Set the granularity from the buyer's side, not the supplier's. Buyers increasingly organize evaluation by occupation and task: OpenAI's GDPval covers 44 occupations across the top nine sectors contributing to U.S. GDP, with tasks built from real work products by professionals averaging 14 years of experience. Anthropic's guidance on agent evaluations calls 20 to 50 simple tasks drawn from real failures a great start, a reminder that evaluation buyers value distinct, real cases over volume. A taxonomy that lines up with the task lists buyers already use is one they can check.
Three rules keep the count honest. Count a family once, however many suppliers contribute to it. Count it only when it holds enough outcome-linked decisions to train or test something; a family represented by a handful of records is a gap, not coverage. And keep exceptions, such as a disputed invoice or an escalated claim, as their own families when the rule structure changes, because that is where buyers' models fail.
The replacement set: how many suppliers must a rival sign to match the coverage?
The replacement set is the smallest group of suppliers whose combined records would reproduce a chosen share of a product's decision families. It converts breadth into a question a rival's business plan has to answer: how many separate negotiations stand between a newcomer and equivalent coverage?
Compute it greedily. List each supplier's covered families, pick the supplier that covers the most, then keep adding whichever supplier covers the most families not yet covered, recording the cumulative share after each step. The resulting replacement curve rises steeply for a narrow portfolio and slowly for a broad one. Report the replacement set at a stated threshold, such as 80% of families, beside the curve itself.
Pair it with redundancy. For each family, compute the effective number of contributing suppliers, the inverse of the sum of squared supplier shares, so that one dominant supplier counts as roughly one. A family with a single effective supplier is fragile: if that supplier leaves, the coverage leaves with it. The same concentration arithmetic, applied to customers inside a dataset, is set out in the concentration guide linked from this page.
The curves in the table are illustrative, for the two portfolios described later in this article. They show the share of decision families a rival would replicate after signing each number of suppliers, in the order the greedy method picks them.
| Suppliers a rival signs | Portfolio A: 32 logos, one workflow | Portfolio B: 9 logos, five workflows |
|---|---|---|
| 1 | 75% of families | 21% of families |
| 2 | 100% | 36% |
| 3 | 100% | 50% |
| 5 | 100% | 71% |
| 7 | 100% | 86% |
| 9 | 100% | 100% |
Seven lines that test workflow coverage, and what would break each
Score each line from evidence the company can produce, and run its disconfirmation test before accepting it. Breadth counts toward defensibility only when the first four lines pass; the last three decide whether buyers value it enough to pay.
Two lines carry the most weight. The replacement set says whether coverage is hard to copy. Cross-workflow transfer says whether coverage matters at all: if a model trained on one deep workflow performs as well on held-out families as a model trained on the broad mix, breadth adds little a buyer cannot get from depth. The source diversity protocol linked from this page shows how to run that comparison at an equal case budget, with whole suppliers held out.
| Line | Measure and denominator | Counts toward a moat when | Disconfirmed if |
|---|---|---|---|
| Decision breadth | Families with enough outcome-linked decisions to train or test, out of the families in buyers' task lists | Coverage spans most of the families buyers name | Most families rest on a handful of records |
| Replacement set | Fewest suppliers that reproduce 80% of families, from the greedy curve | A rival needs many separate agreements | Two or three suppliers reproduce most of the coverage |
| Redundancy | Effective number of suppliers per family | Key families draw on several suppliers | Key families depend on one supplier each |
| Rights-consistent coverage | Families that survive the narrowest common license scope across their suppliers | Coverage holds under one scope a buyer can accept | Combining suppliers forces a scope too narrow for the buyer's use |
| Cross-workflow transfer | Score on held-out families and suppliers against a single-workflow model at an equal case budget | The broad mix beats depth on held-out work | One deep workflow matches the broad mix |
| Breadth demand | Share of buyer requests that need two or more families | A meaningful share of requests span families | Requests almost always name a single workflow |
| Repeat breadth buyers | Buyers who returned to license an additional family, over comparable periods | Buyers expand across families | Buyers take one family and leave |
Is breadth cheaper to buy direct? The substitution test
For many buyers it is, and that is the strongest alternative explanation for any breadth premium. A team building one claims agent needs claims decisions, not dispatch or procurement, and can often get them from one large insurer or claims administrator faster than from a broad aggregator.
Small, well-chosen datasets make direct sourcing cheaper still. The LIMA result that supports curated diversity also shows how few examples one capability can need, and Anthropic's starting point of 20 to 50 real failures points the same way for evaluation. If a buyer's task fits inside one family, a supplier-diverse product competes on price and speed, not on defensibility.
Breadth earns a premium in three situations: buyers building general business agents that cross workflows, evaluation buyers who need coverage across many occupations and tasks of the kind GDPval measures, and buyers whose models will be deployed in companies unlike any single supplier. Benchmarks such as WILDS exist because real deployments meet distribution shifts that a single training source does not show. The breadth-demand line settles which situation a company is actually in: count requests, not narratives.
Illustrative: 32 logos against 9, scored line by line
Illustrative: both portfolios are invented, and every count is an example chosen to show how the scorecard reads, not a measurement of any company. Portfolio A has 32 suppliers, all residential HVAC contractors, whose dispatch, quoting and maintenance records cover four decision families. Portfolio B has nine suppliers across claims handling, technician dispatch, procurement, collections and customer support, covering fourteen families.
On logos, A looks more than three times larger. On the scorecard, a rival replicates 75% of A's families with one supplier and all of them with two, while B needs seven suppliers to pass 80%. B's weak points are redundancy, since four of its families rest on a single supplier each, and breadth demand, which only a count of buyer requests can settle. Neither portfolio is a moat on these numbers alone; B has evidence an investor should go and test, and A does not.
| Measure | Portfolio A | Portfolio B |
|---|---|---|
| Supplier logos | 32 | 9 |
| Workflows | 1 | 5 |
| Decision families with linked outcomes | 4 | 14 |
| Replacement set at 80% of families | 2 suppliers | 7 suppliers |
| Families resting on one supplier | 0 | 4 |
| What to test next | Price and speed, not defensibility | Breadth demand and rights-consistent scope |
Where the breadth thesis breaks
Breadth fails as a moat in five predictable ways. Each can be tracked, and each belongs in a company's own risk discussion before an investor has to raise it.
- Synthetic variety. Generated tasks can multiply common decisions cheaply. The limit is the tails: a 2024 Nature paper found that models trained recursively on generated data lose the tails of the original distribution, which makes rare decision families the part of breadth hardest to synthesize and easiest to overclaim.
- Rights that do not combine. The MIT Media Lab-led audit of more than 1,800 text datasets found license information on popular hosting sites frequently missing or miscategorized. A portfolio stitched from many suppliers inherits every inconsistency in their terms, and source platforms add their own: Slack's API terms bar apps offered outside the developer's organization from using API data to train large language models.
- Concentrated demand. Reddit's Form 10-K for fiscal 2025 states that 'substantially all of the contract value associated with our licensing revenue is derived from two of our partners'; a broad product sold to two buyers is exposed to two renewals, not fourteen families.
- Model progress. As general models improve, common decision families may stop needing specialized data, leaving value in the rare ones; breadth concentrated in easy families erodes first.
- Coordination cost. Every added supplier brings its own review, preparation and delivery work, and past some point breadth costs more to maintain than buyers pay for it.
Diligence questions for a multi-workflow data company, SourceX included
Ask for the coverage map before the logo slide, and ask the company to name the evidence that would prove its breadth claim wrong. The same questions apply to SourceX.
SourceX is a place to apply the question, not evidence for an answer. Its work is turning selected business records into licensed AI datasets, with sourcing and qualification, rights review, contracts, data preparation, delivery and payment all in its remit. It keeps no catalog in stock: each dataset an AI buyer asks for is sourced from U.S. businesses on request, so for SourceX decision coverage is a property of which businesses can be qualified for a given request, which is exactly what the scorecard asks an investor to test.
No coverage, supplier, deal or commercial metrics are published; SourceX keeps company and commercial metrics for direct conversations with investors. Its capital raise runs under Rule 506(c) and targets accredited investors, who can request materials through the investor page and put any of the questions below to it.
- Which decision families do your suppliers cover, by your own taxonomy, and how does that taxonomy map to buyers' task lists?
- What is your replacement curve, and how many suppliers reproduce 80% of your families?
- Which families rest on a single supplier, and what happens to coverage if that supplier leaves?
- Under what common license scope can your families be combined, and which drop out?
- What share of buyer requests named two or more families over the last four quarters?
- Which held-out test shows that breadth improves performance over depth in one workflow?
Method, sources and limits of the coverage scorecard
The decision-family definition, the replacement set and curve, and the seven-line scorecard are editorial frameworks; they were built on 2026-10-26 for this article from the repository of the deduplication paper by Lee and colleagues (ACL 2022), the LIMA paper (2023), OpenAI's GDPval page, Anthropic's guidance on agent evaluations, the WILDS repository, the 2024 Nature paper on model collapse, the MIT Media Lab-led licensing audit, Slack's API terms and Reddit's Form 10-K for fiscal 2025. Portfolios A and B and their curves are illustrative.
Not measured: decision-family coverage, replacement sets or redundancy for SourceX or any other company; the share of buyer demand that spans families; and the price effect of breadth. The 80% threshold is a convention for comparison, not a standard. The greedy method gives a near-minimal replacement set rather than the exact minimum, which flatters the rival slightly and so errs on the conservative side for the moat claim.
Frequently asked questions
What is workflow diversity in a data product?
It is the range of distinct business decisions a data product covers, such as claim approvals, dispatch choices and invoice holds, rather than the number of companies that supplied records. Two suppliers running the same workflow add volume to one decision family; a supplier in a new workflow adds coverage.
Does having more data suppliers make a data company defensible?
Not by itself. Suppliers that repeat the same decisions add little that a rival could not match by signing the largest one or two of them. Defensibility depends on how many suppliers a rival would need to reproduce the coverage, which the replacement set measures.
How do you measure decision coverage across suppliers?
Define decision families from buyers' task lists, count the families with enough outcome-linked records to train or test, and record which suppliers contribute to each. Then compute the replacement curve and the effective number of suppliers per family, and compare coverage with the families buyers actually request.
When does breadth across workflows matter to AI buyers?
Mainly for buyers building agents that cross workflows, evaluation teams that need coverage across many tasks, and models deployed in companies unlike any single supplier. A buyer with one narrow task usually does better sourcing that workflow directly.
Does SourceX publish its decision coverage or supplier counts?
No. Coverage, supplier and commercial metrics are unpublished, and company and commercial metrics are given to investors directly. Investors who want to test these lines can request materials at sourcex.si/investors; SourceX confines any offering to verified accredited investors.
Sources
- The deduplication paper's repository reports a single 50-word sequence repeated 60,000 times in C4 and that models trained on deduplicated data reach similar or better perplexity with less memorization. Source
- LIMA fine-tuned a 65B model on 1,000 examples selected for quality and diversity or written for task variety, and raters preferred it to a 65B Alpaca reproduction trained on 52,000 examples. Source
- OpenAI's GDPval covers 44 occupations across the top nine sectors contributing to U.S. GDP, with tasks based on real work products by professionals averaging 14 years of experience. Source
- Anthropic's guidance on agent evaluations says 20 to 50 simple tasks drawn from real failures is a great start for an evaluation set. Source
- WILDS is a benchmark of in-the-wild distribution shifts spanning diverse data modalities and applications. Source
- A 2024 Nature paper found that models trained recursively on generated data lose the tails of the original data distribution. Source
- An MIT Media Lab-led audit traced more than 1,800 text datasets and found license information on popular hosting sites frequently missing or miscategorized. Source
- Slack's API terms bar apps offered outside the developer's organization from using API data to train a large language model. Source
- Reddit's Form 10-K for fiscal 2025 states that substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.