AI data market
Business exception data investment: need is proven, budgets are not
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Business exception data passes the need tests for an AI infrastructure category but not yet the purchase tests. Models learn rare facts poorly (Kandpal et al., 2023), leading agents completed about 35% of multi-turn business tasks in CRMArena-Pro (2025), and Gartner's 2029 forecast covers common service issues, not the rare ones. No verified public source shows buyers budgeting for exception data as its own line, so treat the category as a hypothesis.
Value drivers
Business exception data
Key takeaways
- A category needs its own budget, a repeatable unit, a checkable quality standard, scarce supply and room outside incumbents; exception data has evidence for some of these, not all.
- Need is well supported: rare facts are learned poorly, business agents struggle in multi-turn work, and a regulator has documented chatbots failing when a problem falls outside their scope.
- The sellable unit is a verified exception episode, not an exception record: a frozen starting state, allowed actions, constraints and authority, a checkable outcome and refused look-alikes.
- Value would accrue in the middle of the stack, linking and grading, which sits between system-of-record vendors and model developers who could each absorb it.
- The purchase test is unproven: no verified public source here shows buyers paying for exception data as a distinct line, so ask for written demand first.
What would make business exceptions an AI infrastructure category?
A category exists when buyers budget for something on its own, suppliers can deliver it in repeatable units, and quality can be checked without trusting the seller. Business exception data, the records of cases where a company departed from its default process on purpose, has a strong case on need and an unproven one on purchase.
The distinction matters because features get bundled and categories get funded. If exception handling becomes a capability that software vendors or model developers build from data they already control, an independent exception data business is a feature with a short life. If buyers need verified exception episodes they cannot produce themselves, a layer forms between the companies that hold the records and the labs that train on them.
This article scores the category on six conditions, maps what an agent needs from an exception, and locates where value would accrue. It cites no market size, because no verified public source measures one for exception data, and it treats every forecast as a forecast.
Why routine data underrepresents the cases agents fail on
Exceptions are rare by construction, and models learn rare things poorly. Kandpal and colleagues (ICML 2023) found strong correlational and causal relationships between a model's accuracy on a fact question and the number of pretraining documents relevant to it, and estimated that models would have to grow by many orders of magnitude to answer competitively on questions with little support in their training data.
Business work concentrates judgment in the same tail. O*NET's task list for customer service representatives (occupation 43-4051.00) includes resolving complaints by exchanging merchandise, refunding money or adjusting bills, and referring unresolved grievances to other departments for investigation. Those adjustments and referrals are the exceptions: the cases a routine script does not cover.
Forecasts point the same way once read closely. Gartner predicted in March 2025 that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. The forecast is about common issues; the uncommon remainder is where exceptions live and where training examples are thinnest.
The cost of missing that remainder is on the public record. The Consumer Financial Protection Bureau's June 2023 report on chatbots in consumer finance described repetitive 'doom loops' and linked them to customer problems that fall outside what the chatbot can handle. A routine-only agent does not fail quietly; it fails on the customers with the most at stake.
The agent-task input map: what an exception episode must contain
An agent trained or tested on an exception needs four things from the case and one from its look-alikes: the starting state, the actions it could have taken, the constraints and authority in force, an outcome a grader can check, and similar cases where the exception was refused. Records supply some of these directly; the rest has to be built around them.
Current agents show why the full set matters. Salesforce researchers' CRMArena-Pro benchmark (2025) spans 19 expert-validated tasks across sales, service and configure-price-quote work; leading agents reached about 58% success on single-turn tasks, falling to about 35% in multi-turn settings, and the authors reported a significant gap between current capabilities and enterprise needs. An exception adds a departure from policy on top of that multi-turn difficulty.
The map uses an illustrative warranty exception, a repair approved outside the warranty period for a long-standing customer, to show where each input usually sits and what has to be added before the case can train or test an agent.
| Input | Illustrative warranty exception | Usually in the records? | What must be built around it |
|---|---|---|---|
| Starting state | Claim filed 40 days after warranty end by a nine-year customer | Yes, across CRM and service tickets | A frozen snapshot as of the decision time |
| Allowed actions | Deny, approve the repair, offer a discount or escalate | Partly; only the action taken is logged | The full action space from policy and tools |
| Constraints and authority | Policy version, goodwill budget, approver's limit | Rarely as structured fields | Dated policy text and an authority table |
| Checkable outcome | Repair completed; service contract renewed | Often in a separate system | A grading rule tied to the outcome record |
| Refused look-alikes | Similar late claims that were denied | Yes, if denials are retained | Matched cases so the boundary is visible |
Six conditions for a category, scored on public evidence
On the evidence verified for this article, exception data meets three of six category conditions, partly meets a fourth and has not shown the last two. The need conditions pass; the conditions that turn need into a budget line do not yet.
Statuses are editorial judgments on the cited evidence as of 2026-10-27. A status of not shown is not proof of absence: purchase programs can exist without public documentation, which is precisely why an investor should ask for written demand rather than infer it.
Two kinds of evidence would move the open rows. For the budget condition, look for written requests that name exception episodes as the deliverable, with a task, a volume and an acceptance test, from more than one buyer, and for repeat orders after the first delivery. For the incumbent condition, look for application vendors licensing outside exception data rather than relying only on their own customers' records. A single large contract would prove a customer, not a category.
| Condition | What would pass | Public evidence, dated | Status |
|---|---|---|---|
| Scarce in routine data | Rare cases learned poorly from typical corpora | Kandpal et al., ICML 2023: accuracy tracks relevant document counts | Supported |
| Agents fall short today | Leading agents fail multi-step business work | CRMArena-Pro, 2025: about 35% multi-turn success | Supported |
| Errors are costly | Documented harm when automation misses the exception | CFPB, June 2023: doom loops when issues exceed chatbot scope | Supported |
| A checkable unit exists | Episodes with outcomes a program or rule can verify | Tulu 3, 2024: verifiable-reward RL named as a method; outcome records vary | Partial |
| Buyers budget for it separately | Purchases or programs that name exception data | No verified public source found for this article | Not shown |
| Room outside incumbents | Independent supply not bundled by app vendors | Gartner, August 2025: 40% of enterprise apps with task-specific agents by end of 2026 | Not shown; at risk |
Where value would accrue: the exception data stack
If a category forms, it forms in the middle of the stack, where raw exception records become linked, de-identified episodes with checkable outcomes. The layers below and above belong to companies that could absorb that work.
At the bottom, software vendors control export paths and API terms. Above them, operating companies own the records and the rights to license them, and only they can say who held authority on a given day. The middle layers, preparation and grading, are where independent businesses can specialize, because the work repeats across companies even when the records do not. Above those, model developers train and evaluate, and application vendors deploy agents, often inside the same software that produced the records.
That overlap between top and bottom is the structural risk. Gartner predicted in August 2025 that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. A vendor whose software both records the exception and runs the agent can try to keep the loop inside its own platform, subject to its customers' terms.
| Layer | Who controls it | Exception-specific scarcity | Absorption risk |
|---|---|---|---|
| Systems of record | Software vendors | Export paths and API terms | High: vendors also ship agents |
| Records and rights | Operating companies | Authority and decision context | Low: cannot be produced elsewhere |
| Linked, de-identified episodes | Data preparers | Cross-system linking and privacy review | Medium: labs may build in-house |
| Outcome checks and graders | Environment builders | Outcome records and grading rules | Medium: labs build graders too |
| Training and evaluation | Model developers | Sealed exception tests | Not applicable: this is the buyer |
| Agents in production | App vendors and deployers | Live exception traffic | High: they own the data loop |
Could a record collection be the wrong asset for agents?
Yes, if the records lack the action space or the outcome. A pile of resolved exception tickets shows what someone did, not what else they could have done or whether it worked, and an agent trained to imitate the logged action learns the choice without the boundary.
Post-training practice raises the bar. Tulu 3 (Ai2, 2024) named reinforcement learning with verifiable rewards as one of its three post-training methods, beside supervised fine-tuning and preference optimization, and the method's name states its requirement: a reward that depends on an answer someone can verify. For exceptions, the valuable unit is therefore an episode whose outcome a program or a documented rule can check, which many record collections cannot support without the outcome system and a grader.
This is the strongest alternative to the category thesis: environment builders could simulate exception workflows from policy documents and skip licensed records altogether. Whether simulated exceptions reproduce the boundary a real team enforces is a testable question, and until a matched comparison on sealed real cases exists, both sides are arguing from priors.
What could make the exception category thesis wrong
Five risks could keep exception data a feature or shrink its market, and each has a public marker an investor can watch.
- Incumbents keep the loop. If application vendors embed agents in the systems that hold exception records, as Gartner's August 2025 forecast anticipates, they can tune agents without an outside supplier, within their customer terms.
- Exceptions stay with people. Gartner predicted in January 2026 that by 2030 generative AI's cost per resolution in customer service will exceed $3, more than offshore human agents cost; if automating the hard tail stays uneconomic, companies route exceptions to staff and buy less data to automate them.
- Agent projects stall. Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls; cancellations cut demand for the data that trains them.
- Simulation substitutes. If generated exception workflows pass a sealed test built from real cases, licensed exception records lose their premium for that task family.
- Rights and privacy narrow supply. Exception records carry the most sensitive details in a workflow, such as hardship claims, disputes and approvals by named staff, so a larger share may need redaction or exclusion, and usable supply can be smaller than record counts suggest.
Questions to ask an exception data supplier, SourceX included
Ask for evidence at the layer the supplier claims to own, and ask what happens to it if the layer above or below absorbs the work.
SourceX is a supplier-side example, not proof of the category. It brokers licenses for business records a company chooses to offer, carrying each deal from qualifying the records and reviewing rights through the contract, preparation, delivery and payment. Its general guide is that buyers look for two or more years of consistent history and tens of thousands of records in one system, though smaller sets can qualify when the work is rare, with every deal assessed on its own.
None of SourceX's supply, deal or commercial figures are public; the company gives them to investors directly, and its raise is open only to accredited investors under Rule 506(c). Materials can be requested at sourcex.si/investors, and every question below applies to SourceX too.
Checklist · 5 items
- What share of your exception cases include the outcome record, and how is the outcome checked?
- Do you retain refused and expired requests alongside approvals, so the boundary is visible?
- Which system-of-record terms limit the exports your supply depends on?
- How many distinct buyers have asked in writing for exception episodes, and for which tasks?
- If an application vendor shipped its own exception agent, which of your layers would it replace?
0 of 5 ticked. Tick each item as you go.
Methods note: what the six conditions rest on
The six conditions, the input map and the stack were drawn up for this article on 2026-10-27 as editorial tools. Evidence comes from three arXiv papers (Kandpal et al., CRMArena-Pro and Tulu 3), four Gartner press releases (March 2025, June 2025, August 2025 and January 2026), O*NET's occupation summary and the CFPB's June 2023 chatbot report, each verified that day through search excerpts restricted to its own domain.
Statuses are judgments on that evidence, and Gartner figures are forecasts, not measurements. Not measured: market size, prices, the number of buyers for exception data, the share of company records that are exceptions, and anything about SourceX's supply or deals.
Frequently asked questions
What is business exception data?
It is the record of cases where a company departed from its default process on purpose, such as a late warranty claim approved, a credit hold released or a fee waived. Useful exception data keeps the starting state, the rule that applied, who approved the departure, what was done and what happened afterward.
Why do AI agents need exception data if they handle routine work well?
Routine cases are plentiful in training data and exceptions are not, and models learn rare cases poorly, as Kandpal et al. (2023) showed for long-tail facts. Agents also struggle with multi-step business work: in CRMArena-Pro (2025), leading agents succeeded on about 35% of multi-turn tasks. Exceptions combine both difficulties.
Is there evidence that AI developers buy exception data as its own category?
Not in the public sources verified for this article. Need is documented, but no verified source shows buyers budgeting for exception data separately, which is why this analysis treats the category as a hypothesis and puts written demand first on the list of things to ask for.
How is exception data different from an RL environment?
Exception data records what happened in real cases; an environment lets an agent act and receive a checkable result. The two meet when exception records become episodes with a starting state, allowed actions and an outcome a grader can verify, which is the unit most likely to sell if a category forms.
Does SourceX publish its exception data supply or sales?
No. SourceX does not publish supply, deal or commercial figures and gives them to investors directly; materials can be requested at sourcex.si/investors. Any offering is limited to verified accredited investors.
Sources
- Kandpal et al. (ICML 2023) found strong correlational and causal relationships between question-answering accuracy and relevant pretraining document counts, with large scaling needed for low-support questions. Source
- O*NET's customer service representative tasks include resolving complaints through exchanges, refunds or bill adjustments and referring unresolved grievances for investigation. Source
- Gartner predicted in March 2025 that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. Source
- The CFPB's June 2023 chatbot report linked doom loops to customer issues falling outside a chatbot's limited capabilities. Source
- CRMArena-Pro (2025) reported leading agents at about 58% single-turn and about 35% multi-turn success across 19 expert-validated business tasks. Source
- Tulu 3 (Ai2, 2024) introduced reinforcement learning with verifiable rewards as one of its three post-training methods alongside SFT and DPO. Source
- Gartner predicted in August 2025 that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. Source
- Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. Source
- Gartner predicted in January 2026 that by 2030 generative AI's cost per resolution in customer service will exceed $3, above offshore human agent costs. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.