AI data market
Enterprise workflows as an AI training thesis: 3 things manuals omit
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
The enterprise workflows AI training thesis says AI that learns how companies actually operate sees three things process descriptions leave out: the real order of steps, the exceptions people resolve by hand, and the feedback showing whether a decision held. The public BPI Challenge 2019 purchasing log splits 251,734 cases across four matching flows. The thesis is investable only if those signals beat simulated companies and larger general models on held-out real cases.
Record timeline
Enterprise workflows AI
Key takeaways
- Public descriptions show the designed process; operating records add sequence variants, exception branches and outcome feedback.
- The BPI Challenge 2019 log from one Dutch coatings and paints company records 1,595,923 events across 42 activities by 627 users, 20 of them batch processes.
- Agents fail where those signals matter: τ-bench found pass^8 below 25 percent in retail, and CRMArena-Pro reports success falling from about 58 percent single-turn to about 35 percent multi-turn.
- Simulated companies are the credible substitute, so the thesis needs real logs to beat them on held-out real cases.
- A large stock of operating records can coexist with few buyers; count written workflow requests before counting logs.
What changes when AI learns from how a company actually operates?
The model sees the work as it happened rather than as it was described. Manuals, standard operating procedures and software documentation describe a designed path. Operating records add three things those descriptions usually omit: the real order of steps and its variants, the exceptions people resolved by hand, and the feedback that showed afterwards whether a decision held.
That difference is the core of an enterprise workflows AI training thesis. If agents fail mainly because they never saw sequencing, exceptions and feedback, then companies that hold those records, and businesses that can license them, control an input no amount of public text supplies. If agents fail for other reasons, such as tool access or reasoning limits that larger models fix, the records are interesting history rather than a scarce asset.
The decision for an investor is which explanation the evidence supports today and what would change the answer. This article starts with one public event log, builds a falsifiable thesis map with evidence columns, checks it against agent benchmarks and the strongest objections, and ends with questions to put to any seller of operating records.
What one purchase-to-pay log shows that a process description does not
It shows one textbook process splitting into several real ones, with people and machines handing work back and forth. The BPI Challenge 2019 log, published on 4TU.ResearchData, comes from purchase order handling at a multinational coatings and paints company in the Netherlands. It holds 76,349 purchase documents with 251,734 line items, each item treated as a case, and 1,595,923 events covering 42 activities performed by 627 users, of whom 607 are people and 20 are batch processes.
The dataset's own description splits the work into four flows: three-way matching with the invoice after goods receipt, three-way matching with the invoice before goods receipt, two-way matching with no goods receipt, and consignment, where invoicing is handled in a separate process. In the second flow an invoice can be entered early but stays blocked until the goods arrive, and either a user or a batch job running at regular intervals releases it. The description adds that the real complexity goes beyond these four categories.
A procurement manual would describe one path: order, receive, invoice, match, pay. The log records which path each item took, where it waited, who or what moved it and where it deviated. Process mining exists to measure exactly this gap. The IEEE Task Force on Process Mining's manifesto frames the field's aim as discovering, monitoring and improving real processes rather than assumed ones, and it describes conformance checking, which compares a process model with an event log to locate deviations.
The thesis map: sequence, exceptions, feedback, authority and timing
Each row names one signal operating records add, what a public description of the same work shows, the evidence today and the result that would falsify the claim that the signal matters for training. A row supports the thesis only if records beat both descriptions and simulation on held-out real cases, which no public study has yet shown for business workflows.
The status column is an editorial judgment from the cited evidence as of 2026-10-26, not a measurement. Observed means the signal demonstrably exists in records; untested means nobody has published whether training on it helps. The figure contrasts the documented purchasing path with the four logged flows.
| Signal in the records | What a description shows | Evidence today | Would falsify it | Status |
|---|---|---|---|---|
| Sequence variants | One ordered path | BPI Challenge 2019: four matching flows inside one purchasing process | Agents trained on the documented path match log-trained agents on held-out cases | Observed; training value untested |
| Exception branches | Rules plus a general escalation step | τ-bench, 2024: pass^8 below 25 percent on rule-bound retail service tasks | Exception cases add no lift over clean cases of equal number | Failures observed; cause open |
| Outcome feedback | The intended result | Lightman and colleagues, 2023: step-level feedback beat outcome-only feedback on MATH | Step events from business records fail to raise held-out success | Shown in math, not in business work |
| Authority and handoffs | Roles on an org chart | BPI Challenge 2019: 627 users, 20 of them batch processes; user or batch releases blocked invoices | Knowing who or what acted adds nothing to agent accuracy | Observed; training value untested |
| Timing and waiting | Deadlines and service levels | BPI Challenge 2019: early invoices held until goods arrive, then released | Time-aware tasks score the same as time-blind versions | Observed; training value untested |
Where today's agents break, and why it points at operating records
They break on consistency, rules and multi-turn work, which is where sequence and exception records carry information. In the τ-bench study of 2024, a gpt-4o function-calling agent completed fewer than half of simulated retail and airline service tasks, and its chance of succeeding on all eight repeated attempts at a retail task fell below 25 percent. The authors concluded that agents need better ways to act consistently and follow rules reliably.
Business software shows the same shape. CRMArena-Pro, built on 25 interconnected Salesforce objects spanning service, sales and configure-price-quote schemas, reports that leading agents succeed in about 58 percent of single-turn scenarios and about 35 percent of multi-turn ones. TheAgentCompany, a simulated software firm first posted as a benchmark in December 2024, finds simpler tasks often solved while long-horizon tasks remain beyond current systems.
Measurement has the same blind spot. OpenAI describes GDPval as limited to one-shot evaluations, so it does not capture work where a model must build context over time or improve a deliverable across several drafts. A benchmark that grades a single deliverable cannot reward the sequencing and feedback this thesis is about, which is one reason multi-step evaluation tasks rebuilt from real records could find buyers.
None of this proves operating records are the fix. It shows the failures sit where records are informative and descriptions are thin, which makes the thesis testable rather than true.
Is feedback in business records a training signal or only history?
It can be a training signal when it attaches to individual steps and can be checked; it is only history when it arrives late and mixed with unrelated causes. The strongest public evidence that step-level feedback helps comes from mathematics. In Let's Verify Step by Step (2023), Lightman and colleagues found that rewarding each correct reasoning step significantly outperformed rewarding only the final answer; their process-supervised model solved 78 percent of a representative subset of MATH test problems, and the released dataset holds about 800,000 step-level human labels.
Business records contain step-level signals of a rougher kind: an approval refused and resubmitted, an invoice blocked and released, a ticket reopened, a credit memo that reverses a price. Whether those signals behave like the step labels in mathematics is an open question. They are not written as grades, they often carry business reasons unrelated to correctness, and the final outcome can arrive weeks after the step.
That gives investors a clean experiment to ask for. Take one workflow and build three training sets for the same held-out cases: documentation only, outcomes only, and outcomes plus logged step events. Compare agent success over repeated trials. If the logged steps add nothing, the feedback row of the thesis map fails, and with it much of the case for paying a premium for full event histories.
The strongest objection: plenty of logs, few buyers and simulated companies
The objection has two parts, and both have evidence. First, a large stock of operating data can coexist with weak purchasing demand. Reddit's fiscal 2025 annual report traces substantially all of the contract value tied to its licensing revenue to two partners, and it cautions that those arrangements might not renew, or might renew on worse terms. Holding a lot of human-generated records has not, so far, meant selling them to many buyers.
At a glance
The strongest objection: plenty of logs, few buyers and simulated companies
Second, simulated companies are a credible substitute for real logs. CRMArena placed a simulated organization, 16 interconnected objects such as accounts, orders, cases and knowledge articles, inside a real Salesforce environment, and TheAgentCompany built an entire simulated software company. If environments like these teach agents the sequencing and exceptions of real work, the premium on real logs shrinks toward the cost of building a better simulator.
Stated demand only partly answers the objection. OpenAI's data partnerships page, published in 2023, asks for large datasets that are not easily available online and singles out data expressing human intention, such as long-form conversations rather than disconnected snippets. That is a stated interest, not a purchase volume. The thesis survives only if three counts move together:
- Buyers: distinct developers that issued written requests naming a workflow and an acceptance test, counted per year.
- Supply: the share of reviewed logs that keep case identifiers, timestamps, actors and outcomes after de-identification and rights review.
- Substitute gap: the lift from training on real logs over a simulated environment of equal cost, measured on held-out real cases.
Signals that the enterprise workflows AI training thesis is wrong
Four signals would show the thesis is wrong, and each can be observed without private data. Track them at every major model release and benchmark update.
- Larger general models close multi-turn gaps with no new data sources. If later CRMArena-Pro or τ-bench results show multi-turn success climbing to single-turn levels while training inputs stay public, sequence records are less scarce than the thesis assumes.
- Simulators match real logs. A published comparison in which agents trained in a simulated company score as well on real held-out cases as agents trained on real logs would remove most of the premium.
- Buyers stay concentrated. A market in which disclosures like Reddit's keep naming two partners is one where suppliers take prices rather than set them.
- Records lose their structure in preparation. If de-identification and rights review routinely strip actors, timestamps or outcomes, the supply that reaches buyers is description-grade, whatever its source.
What to ask a company selling operating records, SourceX included
Ask to see the thesis map filled in with the company's own evidence, row by row, and accept no row without a denominator; SourceX should be asked for the same.
SourceX applies this thesis to licensing; it does not prove it. Companies pick the business records they are willing to license, and SourceX turns them into AI datasets, managing sourcing, qualification, rights review, contracting, preparation of the data, delivery and payment. It holds no inventory for AI buyers; each dataset is sourced from U.S. businesses once a buyer requests it. The owner of the records signs only after approving the scope and terms, and nothing is delivered before that signature.
Buyer counts, signal retention rates and lift results for SourceX are unpublished; company and commercial metrics are something it discusses with investors directly. Its Rule 506(c) raise accepts accredited investors only, and the investors page is where materials are requested. Questions for any seller of operating records:
Checklist · 5 items
- For one workflow you have delivered, which of the five signals survived preparation, measured as a share of cases?
- Can a buyer see the exception branches and their frequencies before purchase, and how were they counted?
- Has any buyer compared training on your logs with a simulated environment or documentation alone, on held-out real cases?
- How many distinct buyers issued written workflow requests in the past year, and how many came back?
- Which preparation step most often destroys sequence or actor information, and what do you do about it?
0 of 5 ticked. Tick each item as you go.
Methods note: sources, dates and gaps in the operating-records map
The five-signal thesis map, its status labels and the proposed experiments are editorial frameworks written on 2026-10-26. Evidence comes from the BPI Challenge 2019 dataset description on 4TU.ResearchData, the IEEE Task Force on Process Mining manifesto, the τ-bench paper (2024), CRMArena and CRMArena-Pro, TheAgentCompany, OpenAI's GDPval and data partnerships pages, Let's Verify Step by Step (2023) and Reddit's fiscal 2025 annual report.
Not measured: whether training on real operating logs beats documentation or simulation for any business workflow, how often the five signals survive de-identification, and buyer counts for any private supplier. The BPI Challenge log covers one company's purchasing process; it is used to show what records contain, not how common its patterns are.
Frequently asked questions
What is an enterprise workflows AI training thesis?
It is the claim that AI models, especially agents, improve most when they learn from records of how companies actually ran their work: the order of steps, the exceptions and the outcomes. It implies that companies holding such records control a scarce input. It can be tested by comparing training on real records with documentation and simulation on held-out real cases.
Why are process manuals not enough to train AI agents?
Manuals describe the designed path, while real work often follows several paths with waits, overrides and reversals. The public BPI Challenge 2019 purchasing log splits one process into four matching flows and notes further complexity beyond them. An agent trained only on the manual has no examples of those branches.
What does a process mining event log contain?
It records what happened to each case as a sequence of events naming the activity and, usually, when it happened and who or what performed it. The BPI Challenge 2019 log, for example, ties 1,595,923 events to 251,734 cases and 627 users. Process mining tools use such logs to discover real process flows and check them against a designed model.
Can simulated companies replace real operating records for AI training?
Possibly, and that is the main substitute risk for this thesis. Benchmarks such as CRMArena and TheAgentCompany already build simulated organizations for agents to work in. Whether training in them transfers to real cases as well as training on real logs does is an open, testable question.
Does SourceX publish evidence for this thesis?
No. SourceX has not published buyer counts, retention of workflow signals or lift results; it shares company and commercial metrics directly with investors. The investor materials request is at sourcex.si/investors; participation in any offering is reserved for verified accredited investors.
Sources
- The BPI Challenge 2019 log, from purchase order handling at a Dutch multinational coatings and paints company, holds 76,349 purchase documents, 251,734 items (cases) and 1,595,923 events across 42 activities by 627 users (607 human, 20 batch), in four matching flows. Source
- The IEEE Task Force on Process Mining manifesto describes discovering, monitoring and improving real rather than assumed processes from event logs, including conformance checking of a model against a log. Source
- OpenAI notes GDPval is limited to one-shot evaluations and does not capture work that requires building context or improving through multiple drafts. Source
- τ-bench (2024) found a gpt-4o function-calling agent succeeded on fewer than 50 percent of tasks, with pass^8 below 25 percent in retail, and called for agents that act consistently and follow rules reliably. Source
- CRMArena built a simulated organization of 16 interconnected objects, such as account, order, knowledge article and case, loaded into a real Salesforce CRM environment. Source
- CRMArena-Pro spans 25 interconnected Salesforce objects across Service, Sales and CPQ schemas; leading agents reach about 58 percent success single-turn and about 35 percent multi-turn. Source
- TheAgentCompany (first posted December 2024) is a simulated software company benchmark in which simpler tasks are often solved but long-horizon tasks remain beyond current systems. Source
- Let's Verify Step by Step (2023): process supervision significantly outperformed outcome supervision on MATH; the process-supervised model solved 78 percent of a representative MATH test subset; PRM800K holds 800,000 step-level labels. Source
- OpenAI's data partnerships page (2023) seeks large datasets not easily accessible online, especially data expressing human intention such as long-form conversations rather than disconnected snippets. Source
- Reddit's fiscal 2025 Form 10-K says substantially all contract value associated with its licensing revenue comes from two partners, with renewal risk. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.