AI data market
Software data access restrictions: four scored AI market scenarios
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Software data access restrictions are narrowing the live route to business data, not the licensed one. Slack's 2025 API terms and HubSpot's May 2026 developer terms bar outside apps from training models on API data, while Slack still documents export tools for workspace owners. For the AI market, that shifts value toward supplier-authorized historical licensing, unless vendors start limiting customer exports. Four scenarios, scored against dated signposts, test it.
License terms
Software data access
Key takeaways
- The restrictions published so far target outside apps reading data live: Slack and HubSpot bar training on API data, and Slack cut history-reading limits for unlisted commercial apps in May 2025.
- A customer's own export is a separate route with separate gates; Slack documents owner exports, though private-channel exports need an application and approval.
- If vendors close live copying but leave customer exports alone, supplier-authorized historical licensing becomes the main lawful route for outside developers to obtain business history for training.
- The scenario that would hurt licensed history most is an export squeeze; none of the sources reviewed shows one, and the EU Data Act pushes toward portability.
- Every call in the scorecard resolves from a public document on a stated date, so readers can score it themselves.
Do vendor restrictions shrink AI data supply or reroute it?
Mostly they reroute it. The restrictions published so far limit what outside apps and agents may do with data they read live through a vendor's interfaces; none of them stops a customer from exporting its own records and deciding, under its own contracts, whether to license them. The investor's question is which of those two routes a business actually depends on.
The two routes need separate names because they fail for different reasons. Live agent access means a third-party app or agent reading or acting on records inside a vendor's system in real time, under the vendor's developer terms, rate limits and approval programs. Supplier-authorized historical licensing means the record owner exporting a bounded set of past records, reviewing rights, preparing the data and licensing it under a signed agreement. The vendor governs the first route; the second runs mainly through the customer agreement and the company's own obligations to its customers and staff.
Useful evidence on software data access restrictions in the AI market is therefore narrow and checkable: dated changes to developer terms, rate limits and export tools, the scope each change covers, and whether any change reaches the customer's own export. The sections below collect that evidence, turn it into four scenarios and attach to each a forecast that resolves from a public document.
What vendors have restricted so far, and which route each rule hits
Five dated actions show one pattern: each limits live or third-party access, and none withdraws a customer's own export.
Two features of the pattern matter for valuation. Every restriction binds the developer of an app or agent, not the record owner. And vendors are opening sanctioned real-time routes while closing copy-and-train routes: Slack's Data Access API lets an app read context without keeping it, and HubSpot exempts partners in good standing. Access is being tiered, not ended.
The table lists each action with the route it hits and what it leaves open, as the vendor's own documents describe it.
| Date | Vendor action | What it restricts | Route hit | Left open |
|---|---|---|---|---|
| 2025, current terms | Slack API terms | Apps offered outside the developer's organization may not train an LLM on API data or bulk export message and file data without an additional agreement | Live, third-party | Owner exports through Slack's own export tools |
| May 29, 2025 | Slack rate limit change | History methods moved from Tier 3 to Tier 1, at most 15 objects per call, for new installs of commercial apps not approved for the Marketplace | Live, third-party | Internal customer-built apps keep 50+ calls a minute and 1,000 objects |
| Current docs, checked October 2026 | Slack Data Access API | Apps may read Slack context for answers but must not store or copy it or use it for training | Live, sanctioned | Real-time answers inside the customer's workspace |
| October 31, 2025 | Amazon letter to Perplexity | Says shopping agents should operate openly and respect a provider's decision whether to take part; demands Comet's agents identify themselves | Live, agentic | Agents that identify themselves where a provider allows them |
| May 4, 2026 | HubSpot developer terms update | Bars using API-accessed data to train, fine-tune or improve AI models; says the terms cover agentic features and MCP-based access | Live, third-party | Single-customer use cases; technology partners in good standing are exempt |
Why live agent access and licensed history fail for different reasons
Live access fails when the vendor says no; licensed history fails when the record owner cannot say yes. That difference decides which businesses a restriction actually hurts.
Live access is a pipe the vendor operates, and open standards make the pipe cheaper to build without changing who controls it. The Model Context Protocol specification describes MCP as an open protocol for connecting LLM applications to external data sources and tools, and its authorization section lets clients call restricted servers on behalf of resource owners through flows based on OAuth 2.1. A standard plug lowers integration cost; the server operator still decides who connects, how fast and under which terms, and HubSpot's May 2026 statement that its terms cover MCP-based access puts that control in writing.
Licensed history runs through the owner, though not entirely free of the vendor. Slack's help center documents exports for workspace owners and admins: public-channel exports by default on Business+, exports of private channels and direct messages only after an owner applies and is approved, and an application-based self-serve export of all conversations for Enterprise org owners. The vendor still shapes how much history leaves and in what form; the decision to license what leaves sits with the company, its customer agreements, its notices and the people who wrote the records.
That is why a training ban on API data is not a ban on licensing. It removes the easy third-party route to bulk history and puts the record owner, and its rights review, back in the delivery path. Whether a given owner may license a given export is a separate legal question that counsel answers on that company's own contracts.
Four scenarios for software data access restrictions
Four scenarios cover the plausible paths: open plumbing, toll road, walled garden and export squeeze. They differ on two questions, how far vendors close live third-party access and whether they also narrow the customer's own export.
The sources reviewed for this article fit the toll road and the walled garden. Slack's lower rate tier for apps outside its Marketplace and HubSpot's exemption for partners in good standing are toll-road mechanics; the training bans at Slack and HubSpot, Slack's no-copy real-time API and Amazon's demand that agents identify themselves are walled-garden mechanics. None of those sources narrows a customer's own export, which makes the export squeeze a risk to watch rather than the current reading.
Each row of the table states what vendors do, what happens to each route and who gains; the figure plots the same grid with the signs seen by October 2026.
| Scenario | What vendors do | Live agent access | Licensed history | Who gains |
|---|---|---|---|---|
| Open plumbing | Publish standard agent interfaces with no training limits | Cheap and broad | Competes with direct pulls for training use | Agent builders; rights review and preparation become the scarce work |
| Toll road | Gate access through approval programs, rate tiers and fees | Priced and slower for unlisted apps | Gains relative value for bulk and training use | Vendors with marketplaces; owners who can export |
| Walled garden | Bar training on API data, require agents to identify, offer read-only real-time APIs | Closed to outside copying | Becomes the main lawful route to history for outside developers | Record owners with clean rights; vendors' own AI features |
| Export squeeze | Also limit customers' bulk exports or AI use of exported data | Closed | Narrowed at the source | Vendors as gatekeepers; history licensing loses |
Which business models gain and which lose under each scenario
Businesses that copy data out of other companies' systems lose under the walled garden; businesses that hold the record owner's signature gain, until an export squeeze. Sorting a target by the route its supply travels is the quickest exposure test an investor can run.
Four business models behave differently across the grid. The direction for each is an editorial reading of the scenario table, not a measured result.
- Agent apps built on third-party APIs: strongest under open plumbing, squeezed by the toll road, and limited to read-only, no-copy answers under the walled garden, which caps what they can learn from customer history.
- Data products that aggregate records pulled through vendor APIs: exposed in every scenario except open plumbing, because rate tiers, fees and training bans like Slack's and HubSpot's make the bulk-copy route they depend on slower, costlier or closed.
- Licensing businesses that work from the record owner's own export: gain relative value under the toll road and the walled garden, and lose most under an export squeeze, which is why the scorecard tracks customer exports separately.
- System-of-record vendors with their own AI features: gain in the walled garden and the export squeeze, since they keep live access to their customers' interactions while outside developers cannot copy them.
The public forecast scorecard: five calls a reader can check
Five calls, each resolving from a public document on a stated date, will show which scenario is unfolding, and any reader can score them without trusting this article.
A new training ban at another system-of-record vendor, or a second platform demanding that agents identify themselves, would strengthen the walled-garden reading. A vendor withdrawing customers' own bulk export would move weight to the export squeeze, the single result that would most weaken the case for supplier-authorized history. The two regulatory calls test whether rule makers push vendors toward portability or let them charge for access: the Commission says the Data Act, in application since 12 September 2025, requires open interfaces and machine-readable export from software-as-a-service providers when customers switch.
The calls in the table are editorial judgments made on 2026-10-26, not measured probabilities, and each reading cites only the sources listed in the methods note.
| Call | Resolves by | Scores true if | Reading on 2026-10-26 |
|---|---|---|---|
| Another system-of-record vendor bars training on API data | 2027-10-31 | A CRM, help desk, chat or ERP vendor publishes developer terms with a training ban like Slack's or HubSpot's | Two dated cases: Slack in 2025, HubSpot in May 2026 |
| No vendor named here removes customers' own bulk export | 2027-10-31 | Slack and HubSpot keep a documented owner export and add no ban on AI use of exported data | Holds for Slack, whose export help still documents owner exports with approval gates |
| Agent identification becomes a stated access condition | 2027-10-31 | A second large platform publishes a rule or letter requiring AI agents to identify themselves | One case: Amazon's October 2025 letter |
| EU switching charges end on schedule | 2027-01-12 | The Data Act's ban on switching and data egress charges applies from that date as scheduled | Scheduled; transitional charges are allowed until then |
| US open-banking access fees get settled | 2027-10-31 | The CFPB publishes a proposed or final rule deciding whether fees may offset data providers' costs | Open: reconsideration notice issued in August 2025 |
Could better models, synthetic data or concentrated buyers make access irrelevant?
They could shrink it, and an investor should weigh all three before backing the thesis. Each is a reason the value of business history might fall whatever vendors decide.
Better models and synthetic data come first. Epoch AI's 2024 study sized the effective stock of quality public human text at roughly 300 trillion tokens, a supply it projected models would fully use between 2026 and 2032 on current trends; its later outlook, What will AI look like in 2030?, judges a data bottleneck hard to rule out but surmountable, with synthetic data among the ways around it. If synthetic workflows pass the same acceptance tests as real support, sales or finance histories, the premium for licensed history falls and vendor restrictions matter less.
Concentrated buyers come second. Vendors that close outside training routes can keep their customers' interactions for their own AI features or for partners they choose; HubSpot's exemption for technology partners in good standing is a vendor deciding who may use API data. If a handful of model developers and the vendors themselves are the only buyers of business history, sellers of licensed history face pricing pressure even in the walled-garden scenario.
One observation separates these explanations from the access thesis. If outside developers keep asking record owners for licensed history after the restrictions, access was the constraint; if those requests fade as synthetic data improves or buying consolidates, it was not.
What could make the access thesis wrong
Five developments would weaken the case that restrictions on live access raise the value of supplier-authorized history. Each has a public marker an investor can watch, and a business should be able to name the one that would cut its supply deepest; one that cannot has not stress-tested where its records come from.
- Vendors extend restrictions to customer exports. An export squeeze narrows history at the source, so watch export documentation and customer agreements, not only developer terms.
- Terms change faster than diligence. Slack's May 2025 terms applied to new apps at once and to older apps from June 30, 2025, according to its changelog; a thesis built on one version of a vendor's terms can age in weeks.
- Regulators settle access by fee. The CFPB's August 2025 reconsideration of its personal financial data rights rule asks how fees may offset data providers' costs; if paid access becomes normal, live access stays open to whoever pays and history loses its relative edge.
- Synthetic data closes the gap. Epoch AI's 2030 outlook treats a data bottleneck as surmountable; if that holds for business workflows, licensed history becomes a niche input.
- Owners cannot license what they export. Customer contracts, privacy notices and third-party content can block licensing even when export is easy, which caps supply whatever vendors do.
Diligence questions on supply routes, from vendor terms to exports
Ask where supply comes from, which route it travels and what happens to it if the scorecard turns. The same questions fit any company selling business history, SourceX included.
SourceX is one application of the licensed-history route. Companies license chosen records through it for AI use, while SourceX carries the qualification, rights review, contract, data preparation, delivery and payment steps. Nothing is delivered until the record owner has approved the deal's scope and terms and signed, so authority comes from the owner's signature rather than from a vendor's API.
SourceX has not published supply, deal or commercial metrics; it gives company and commercial metrics to investors directly instead. Only verified accredited investors can take part in its Rule 506(c) raise; investor materials are available on request through its investors page, where a reader can bring any question below, such as how a gated export is counted.
- What share of supply arrives as customer-controlled exports rather than live API pulls, and how is that share measured?
- For each source system, which version of the vendor's developer terms and export documentation did you review, and on what date?
- Which call in the scorecard, if it resolved against you, would remove the most supply?
- How many datasets were delayed or dropped because a vendor gated an export, and how is a gated export defined?
- Do any customer agreements behind your supply limit AI use of exported records, and who checked them?
Methods note: how the scenarios and calls were built
The two-route distinction, the four scenarios and the five calls are editorial frameworks built on 2026-10-26. Evidence comes from Slack's API terms, rate-limit changelog, Data Access API documentation and export help page; HubSpot's May 2026 developer changelog; Amazon's October 2025 letter to Perplexity; the Model Context Protocol specification; the European Commission's Data Act explainer; the CFPB's 2025 reconsideration page; and two Epoch AI publications. Vendors appear because their public documents state a dated restriction, not because they are representative of the market.
Not measured: how much business data moves through live APIs compared with exports, how often vendors approve export applications, what buyers pay for either route, and any SourceX supply figure. Each needs records that are not public. Vendor documents change without notice, so every reading is as of the date shown.
Frequently asked questions
Do API training bans stop a company licensing its own data?
Not by themselves. The bans Slack and HubSpot published bind developers of apps that read data through the vendors' interfaces. A company that exports its own records with the vendor's export tools faces a different question, whether its customer agreements, notices and contracts allow licensing, which counsel answers on the company's own documents.
What is the difference between real-time agent access and historical data licensing?
Real-time access lets an app or agent read or act on live records inside a vendor's system, under the vendor's terms and rate limits. Historical licensing is the record owner exporting a bounded set of past records, preparing them and licensing them under a signed agreement. The vendor controls the first directly; the owner controls the second within its own obligations.
Why are software vendors restricting AI agents?
Their own documents give three reasons. Slack said its history methods in unvetted apps could exfiltrate large amounts of sensitive conversational data, HubSpot exempts technology partners in good standing from its training ban, and Amazon's letter to Perplexity says providers should decide whether agents may operate on their services at all.
Does MCP make enterprise data more accessible to AI?
It makes connections cheaper to build, not permission easier to get. MCP standardizes how AI applications connect to data sources and tools, and its authorization rules let a server require credentials on behalf of resource owners. The server operator still decides who connects and on what terms, and HubSpot says its developer terms apply to MCP-based access.
Does SourceX publish how much of its supply comes from exports?
No. SourceX has not published supply or commercial metrics and shares company and commercial metrics with investors directly. For questions about export-based supply, request investor materials through sourcex.si/investors; any offering is made exclusively to verified accredited investors.
Sources
- Slack's API terms bar providers of apps offered outside their organization from using API data to train a large language model and from bulk exporting message and file data unless an additional agreement allows it. Source
- Slack's May 29, 2025 changelog moved conversations.history and conversations.replies from Tier 3 to Tier 1 with a 15-object limit for commercially distributed apps not approved for the Marketplace, kept internal customer-built apps at 50+ requests a minute and 1,000 objects, and applied the updated API terms to apps created before May 29 from June 30, 2025. Source
- Slack's Data Access API documentation says developers must not store or copy data retrieved from the API and may not use it for training. Source
- Slack's help center says Business+ owners and admins can export public-channel data, private channels and DMs need an owner application and approval, and Enterprise org owners can apply for a self-serve export of all conversations. Source
- HubSpot's May 4, 2026 changelog says its Developer Terms now bar using API-accessed data to train, fine-tune or improve AI or machine learning models, with a carve-out for single-customer use cases and an exemption for technology partners in good standing, and that the terms cover agentic features and MCP-based access. Source
- Amazon published its October 31, 2025 cease-and-desist letter to Perplexity, saying third-party shopping agents should operate openly and respect service providers' decisions whether to participate, and demanding that Comet's agents identify themselves. Source
- The Model Context Protocol specification describes MCP as an open protocol that enables integration between LLM applications and external data sources and tools. Source
- The MCP authorization specification lets clients make requests to restricted MCP servers on behalf of resource owners, based on OAuth 2.1 and related standards. Source
- The European Commission says the Data Act applies from 12 September 2025, requires open interfaces and machine-readable export from PaaS and SaaS providers when customers switch, and ends switching and egress charges from 12 January 2027. Source
- The CFPB's reconsideration of the personal financial data rights rule, opened with an August 2025 advance notice, seeks comment on issues including how fees may offset data providers' costs. Source
- Epoch AI estimated the effective stock of quality public human text at about 300 trillion tokens and projected full use between 2026 and 2032 on current trends. Source
- Epoch AI's 2030 outlook finds a data bottleneck difficult to fully rule out but surmountable, with synthetic data among the routes around it. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.