Skip to content

AI data market

Enterprise context product investment: three tests at the tenant wall

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Enterprise context already sells inside software: Microsoft priced Microsoft 365 Copilot at $30 per user per month in 2023, and Salesforce agreed in May 2025 to buy Informatica for about $8 billion to strengthen its data foundation for agents. That context serves the customer's own agents and stays in its tenant. It becomes a product of its own only when packaged as a reusable task, licensed to leave and priced by an outside buyer.

Value drivers

Enterprise context product

This page covers enterprise context product investment: three tests at the tenant wall.Illustrative

Key takeaways

  • Inside the tenant, enterprise context is already priced: per user for Microsoft 365 Copilot, per action for Agentforce, and through acquisitions such as Informatica and data.world.
  • Vendors such as Microsoft promise that in-app context will not train foundation models, so outside model developers must license context from the companies that own it.
  • Context becomes a standalone product only when it is packaged as a reusable task, licensed to leave the tenant and priced by an outside buyer.
  • Constraints in force, interaction history and verified outcomes are the packet rows outsiders cannot rebuild, and the likely source of pricing power.
  • Outside-the-tenant demand rests on one buyer's stated plan; independent buyers, repeat purchases and holding prices would show a category forming.

When does enterprise context become a product of its own?

When three things hold at once: the context can be packaged as a unit that makes sense without the software that produced it, its owner can license it to leave the company's tenant, and someone other than the software vendor pays for it. Enterprise context here means the state, history, rules and outcomes around a piece of work: the account as it stood, the actions staff could take, the policy version that applied and what happened next.

Today nearly all of it is monetized inside software, where a vendor prices access to a customer's own context for that customer's users and agents. That market is real. The investment question tested here is narrower and earlier: whether reconstructed context can also be sold outside the tenant, to teams that train and evaluate agents, with its own packaging, rights and price.

Public evidence so far points one way: the inside market is proven, the outside market has one prominent stated buyer and almost no public prices, and the deciding variable is rights rather than technology. The sections that follow set out that evidence, an agent-task input map, a value chain, the strongest alternative and the risks.

Inside the tenant, context already has a price

Context is already sold inside software, at published prices. Microsoft announced in September 2023 that Microsoft 365 Copilot would reach enterprise customers on November 1 at $30 per user per month; its privacy documentation describes Copilot working with data reached through Microsoft Graph. Salesforce introduced Flex Credits for Agentforce in May 2025 at $500 per 100,000 credits, with one agent action using 20 credits, or $0.10.

Platform owners are also paying to own the layers that organize context. On May 27, 2025 Salesforce signed a definitive agreement to acquire Informatica for about $8 billion in equity value, net of its existing stake, at $25 a share in cash, and said the deal would strengthen the data foundation needed for agentic AI. In the same month ServiceNow agreed to acquire data.world, saying its catalog and governance platform would let customers enrich data with meaning, context and relationships for AI agents; terms were not disclosed.

Each of these prices access to context, or the machinery around it, for the company that owns the context. None sells one company's context to another. The comparison places them beside the product in question on four criteria:

Inside the tenant, context already has a price
OfferUnit of saleWhose context, used by whomLeaves the tenant for outside model training?Public price evidence
Microsoft 365 CopilotPer user per monthThe customer's own data, for its own staffNo: Microsoft says prompts, responses and Graph data do not train foundation models$30 per user per month from November 1, 2023
Salesforce AgentforcePer agent action, through Flex CreditsThe customer's CRM and service data, for its own agentsNot addressed in the pricing announcement$0.10 per action at launch, May 2025
Informatica (Salesforce agreement)Acquisition of the companyTools that integrate and govern a customer's dataNot applicable: a data management layerAbout $8 billion equity value, announced May 27, 2025
data.world (ServiceNow agreement)Acquisition of the companyA catalog and semantic layer over a customer's dataNot applicable: a metadata layerTerms not disclosed, May 2025
Licensed task packetsDataset licenseOne company's reconstructed work, for outside model developersYes, when the owner licenses itOne buyer's stated plan and price range; no transaction data

The tenant wall: why in-app context does not turn into training data

Because vendors such as Microsoft promise customers that it will not. Microsoft's Copilot privacy page says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, including the ones Copilot uses. A promise like that reassures customers, and it also leaves the vendor in no position to resell the context it serves.

Call that boundary the tenant wall. Inside it, vendors compete on how well they organize and serve a customer's context, and they are paid per seat, per action or through acquisitions of the supporting layers. Outside it, a model developer that wants enterprise context to train or test agents generally cannot buy it from the software vendor. It has to obtain it from the company that owns the records, under a license that company signs.

That makes rights the gate. A company's authority to license its records turns on its contracts with customers, employees and vendors and on privacy review, and having the access to export records from a software tool settles none of it. Much of the value of a context product lies in getting through the wall cleanly, with a record of who approved what.

The agent-task input map: what a context packet must carry

A packet is sellable when it carries everything an agent needs to attempt a task and everything a grader needs to judge the attempt, and a raw export rarely does.

Public benchmarks show how builders package these elements when they cannot license them. SOP-Bench pairs human expert-authored procedures with mock tools for reproducibility and ground-truth outputs, and its README says its tasks were created through human-AI collaboration. That is context rebuilt rather than licensed, and it defines what a licensed packet has to beat: it must carry the constraints, interactions and outcomes that a rebuilt task cannot.

The map ties each requirement to where it lives in a business, whether an outside party could rebuild it without the company, and which right must clear before it leaves; the last two columns are editorial judgments:

The agent-task input map: what a context packet must carry
Task inputWhat it isWhere it usually livesRebuildable without the company?Right to clear before it leaves
Initial stateThe account, order or case as it stood when work beganCRM, ERP or ticket history with field changesPartly, with simulated recordsCustomer data terms and privacy review
Permitted actionsThe tools and steps staff could use, with their limitsRole permissions, admin settings, runbooksRoughly, from product documentationUsually light; may expose internal setup
Constraints in forceThe policy version, contract term or approval rule that appliedPolicy files, contract systems, approval logsOnly the parts that were publishedConfidential terms and third-party contracts
Interaction historyMessages, calls and handoffs during the taskHelp desk, email, chat, call recordingsNo, beyond synthetic dialogueCustomer and employee privacy, consent
Verified outcomeWhat happened afterward and how it was confirmedBilling, returns, renewals, audit trailsNoFinancial confidentiality
Baseline failureWhere a current agent or a new hire went wrong on the same taskProduced by testing, not exportedYes, by whoever runs the testNone beyond the packet's own license

The value chain: where margin could sit on each side of the wall

Inside the tenant, margin accrues to whoever owns the system of record and the layers that organize it, which is why the largest application vendors are buying integration and catalog companies and pricing agents per action. Outside the wall the scarce steps are different: reconstructing a task across systems, clearing rights, removing personal data while keeping the task coherent, and attaching a verified outcome.

The figure draws the chain in two halves. Inside, value flows from records through integration and context layers to the customer's own agents, paid for by seat, by action or by acquisition. At the wall, value depends on a signed license. Outside, it flows through reconstruction and rights clearing to model developers who train and evaluate agents. The diligence question is which link a company owns, and whether that link stays scarce if the platforms reach outward.

The outside chain leans on skilled labor where the inside chain leans on software. That may make it harder to copy and is certainly harder to scale, so a company here needs a repeatable way to turn one supplier's approval into many packets without redoing the rights work each time.

A value chain split by a vertical tenant wall: inside, records flow to integration, a context layer and the customer's own agents priced per seat, per action or by acquisition; a license gate in the wall leads outside to reconstruction, rights clearing and model developers.
Where enterprise context is priced today and where a licensed product would sit. Editorial framework; no market sizes implied.

Is there a buyer outside the tenant?

There is at least one stated buyer, which is not yet a market. micro1 announced in October 2026 a plan to invest $1 billion in a year of acquiring and licensing businesses' operational data, de-identified, to build reinforcement learning environments that reflect real business operations. Its data partnerships page says it licenses qualifying trajectories for $1 million to $2 million a year, with the company's approval. Both are the company's own statements of intent and terms, not evidence of money spent or of any particular company's data being bought.

The signals that would turn a stated plan into a market are easy to name and not yet public: several independent buyers, repeat purchases of the same kind of packet, prices that hold when a second supplier appears, and evaluation-only licenses bought alongside training licenses. An investor who sees those four early is watching a category form; one announcement shows only that the question is live.

Willingness to pay will also follow whichever input the buyer lacks. A lab that can simulate initial states and actions cheaply will pay mainly for constraints, interactions and verified outcomes, the three rows of the input map that cannot be rebuilt from outside.

The strongest alternative: environments instead of record collections

The best case against the thesis is that agents learn more from interactive environments than from records, and environments can be built without anyone's archive. A record collection may lack the live system an agent acts in, the chance to try an action and see the result, and the outcome check that turns an attempt into a score.

The objection is partly right. A packet is a snapshot and cannot answer an action nobody took. Its value holds where an environment builder needs what simulation cannot supply: the real mix of cases, the constraints actually in force and outcomes that took weeks to appear. That yields a direct diligence test: ask to see a buyer using the packets to build or calibrate an environment, and ask what the buyer could not have produced from documentation alone.

If the answer is nothing, the packets are a commodity that simulation will undercut. If the answer names constraints and outcomes, the product has a defensible core.

What could make the enterprise context thesis wrong

Five developments would weaken the thesis, and each can be watched:

At a glance

What could make the enterprise context thesis wrong

  • Platforms reach outward. If vendors that already price context inside the tenant begin licensing consented, de-identified context to model developers, independent suppliers lose their place at the wall. Current public commitments point the other way, as Microsoft's no-training statement shows.
  • Rights stay blocked. If customer contracts, employee agreements or privacy reviews stop most companies from licensing records, supply never reaches scale, however strong demand becomes.
  • Simulation catches up. If generated environments come to supply realistic constraints and outcomes, the hardest rows of the input map lose their scarcity; benchmark builders already rely on mock tools and human-AI generated tasks.
  • Demand stays concentrated. If one or two buyers account for most purchases, prices and terms follow their roadmaps; a single announced plan is a concentration signal as much as a demand signal.
  • Packets do not travel. If each buyer needs a custom reconstruction, unit costs stay high and the business behaves like services work.

Diligence questions for a context supplier, SourceX included

Ask for evidence that each of the three tests is passed, not a description of the market. SourceX works at this boundary: when a business licenses selected operational records to AI developers, it qualifies the company, reviews its rights, handles the contract, prepares and delivers the records and manages payment, and the business signs off on scope and terms before SourceX receives any data. It is raising from accredited investors under Rule 506(c) and shares its company and commercial metrics with investors directly through its investor page; this article reports none of them.

The questions below apply to SourceX as much as to any other company in the space:

  • Packaging: show one complete packet with initial state, permitted actions, constraints, interactions and a verified outcome. Which parts were reconstructed and which exported as they were?
  • Rights: who signed for the records in that packet, on which documents, and what do the supplier's customer and employee agreements allow?
  • Willingness to pay: how many distinct buyers issued written requests, for training or for evaluation, and how many bought again?
  • Repeatability: how long did a second packet from the same supplier take compared with the first, and why?
  • Substitution: what did a buyer say it could not have built from documentation or simulation?
  • Concentration: what share of requests came from the largest buyer, and how would terms change if that buyer left?

Methods note: how the comparison and the input map were built

Sources: announcements and documentation from Microsoft, Salesforce, ServiceNow and micro1, and the SOP-Bench repository, each checked on its own domain on October 27, 2026; dates and figures appear as those sources state them. The four comparison criteria (unit of sale, whose context is used and by whom, whether it may leave the tenant for outside model training, and public price evidence) were chosen to separate in-tenant offers from licensable products.

Not measured: market size, buyer counts, transaction prices, whether announced acquisitions have closed, or contract terms beyond what the cited pages say. The input map's rebuildability and rights columns are the author's judgments, not survey results. No SourceX revenue, pipeline, transaction or cost data is used or implied.

Frequently asked questions

What is enterprise context in AI?

It is the state, history, rules and outcomes around a piece of work: what the record looked like, which actions were allowed, which policy applied and what happened next. Agents need it to act correctly, and graders need it to judge them.

Is enterprise context a separate market from copilots and agent platforms?

Not yet, on public evidence. Copilots and agent platforms sell access to a customer's own context inside its tenant. A separate market would sell reconstructed context to outside model developers under license, and public prices for that remain scarce.

Why don't software vendors sell their customers' context to AI labs?

Because offers such as Microsoft's promise the opposite: Microsoft says Copilot prompts, responses and Graph data are not used to train foundation models. Licensing context to outsiders needs the data owner's own agreement.

What evidence would show the enterprise context category is real?

Several independent buyers, repeat purchases of the same packet type, prices that survive a second supplier, and evaluation-only licenses bought alongside training licenses. A single buyer's announcement is a start, not proof.

How should investors diligence a company selling enterprise context?

Ask to see a complete packet, the signed rights behind it, the number of distinct buyers and repeat purchases, the time needed to produce a second packet and what a buyer could not have rebuilt itself. Treat any answer without documents as unverified.

Sources

  • Microsoft announced in September 2023 that Microsoft 365 Copilot would be available to enterprise customers from November 1, 2023 at $30 per user per month. Source
  • Microsoft says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, including those used by Copilot. Source
  • Salesforce's May 2025 Flex Credits pricing for Agentforce: $500 per 100,000 credits, with one agent action consuming 20 credits ($0.10). Source
  • On May 27, 2025 Salesforce agreed to acquire Informatica for about $8 billion in equity value, net of its existing stake, at $25 per share in cash, citing a data foundation for agentic AI. Source
  • ServiceNow agreed in May 2025 to acquire data.world to let customers enrich data with meaning, context and relationships for AI agents; terms were not disclosed. Source
  • micro1 states a plan to spend $1 billion over 12 months acquiring and licensing de-identified enterprise operational data for reinforcement learning environments. Source
  • micro1's data partnerships page says it licenses qualifying trajectories for $1 million to $2 million a year, with the company's approval. Source
  • SOP-Bench is built from human expert-authored procedures, mock tools for reproducibility and ground-truth outputs, with tasks created through human-AI collaboration. 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