Most firms in financial services are running both of these scenarios right now. Sometimes in the same team, without realising it.
In the first scenario, a portfolio manager opens their preferred AI tool, pastes in a dataset they've pulled from a licensed data source, and asks it to help analyse earnings trends. Their data vendor's system sees a named user logged in. A record exists. The licence terms, as written, cover this.
In the second scenario, an AI agent runs an overnight workflow, autonomously querying multiple licensed data sources via MCP connections to build a morning research brief. The vendor's system sees programmatic queries arriving at machine frequency. There's no named user in the log. Whether the licence covers it depends on terms that were almost certainly written before AI agents existed at scale.
Same firm, same underlying data , two entirely different compliance positions.
Why the two scenarios are legally different
The distinction matters because data licences are built around a specific set of assumptions: that the data consumer is a named human, that access is initiated by a person, that the declared use case is something a person is doing, and that the query volume reflects human behaviour.
AI agents break all four assumptions simultaneously.
An AI agent is not a named user. The access it generates is not initiated by a person. It's initiated by code running on a schedule. The declared use case is autonomous signal generation or report production, which is meaningfully different from a human analyst conducting research. And the query volume can reach levels that a human user licence was never designed to accommodate.
This matters in practice, not just in theory. When a data vendor reviews your usage at renewal, they will look at query volumes, access patterns, and use case declarations. An agent running at machine frequency against a licence written for human frequency is going to stand out. The question is whether you've thought about this before that conversation happens, or during it.
The compliance question is not simply "are we licensed?" It's "are we licensed for this specific type of access, by this type of consumer, at this volume, for this declared purpose?" Those four conditions are straightforward for a human analyst. They are considerably harder to satisfy for an AI agent operating autonomously.
Data contracts were written for a different world
For the most part, data licence agreements in financial services were written for terminal access and named desktop users. Some firms negotiated API addenda as data consumption moved to programmatic interfaces, though these were still framed around human-initiated queries and application-level access controlled by IT.
AI agents via MCP are something different again. They don't fit cleanly into the "named user" or "approved application" categories that current agreements were built around. In most cases, firms have deployed agents into a gap the contract simply doesn't address: neither explicitly permitted nor explicitly prohibited.
The same question applies to Snowflake Data Share deployments. As licensed data from FactSet, MSCI, S&P, and others arrives directly into customer Snowflake accounts via Data Share, the access model is infrastructure-level , it has no awareness of the commercial licence terms the data carries. An AI agent querying a Data Share table is not, by default, operating within any entitlement framework the data vendor agreed to.
That silence is unlikely to persist. Data vendors are not unaware of what's happening. They see the usage patterns. They see query volumes that bear no relationship to headcount. And the economics of human-scale licensing look very different from agent-scale consumption. It is reasonable to expect that licence agreements will evolve to reflect this. Not immediately, and not uniformly, but the direction of travel seems fairly clear.
I want to be direct that this is a viewpoint rather than a confirmed observation. I'm not aware of a vendor who has publicly issued an AI agent addendum as a standard practice yet. But the commercial logic is there, and the firms that are thinking about it now, before the renewal conversation happens, will be better positioned than those who encounter it for the first time across the table from their vendor's commercial team.
What evolving contracts are likely to require
Based on how data licensing has evolved in previous technology transitions (from desktop to API, from application to cloud) there are some reasonable inferences about what AI agent clauses in data contracts might look like.
A requirement to declare AI agent use cases explicitly seems plausible. Not just "research and analysis" as a catch-all, but the specific agent workflow, its intended outputs, and who has oversight of it. The move from implicit to explicit use case declaration has happened before in data licensing and it tends to happen when vendors become aware that the existing declaration doesn't capture what's actually going on.
Distinct licence categories for autonomous vs human-assisted access is another possibility. The difference between a human using AI to help analyse data they have pulled themselves and an AI agent autonomously querying data without a human in the loop is significant enough that treating them under the same licence terms starts to look like a gap that benefits the consumer rather than an oversight.
Audit trail requirements as a contractual obligation is perhaps the most consequential. Several regulatory frameworks already require firms to maintain records of data used in investment decisions. It would be a natural extension for data vendors to require, as a licence condition, that firms can demonstrate compliant use on request, rather than simply assert it.
None of this is certain. But it's worth running your current agreements against these possibilities before a renewal forces the conversation.
Five questions worth answering before your next renewal
If you have data contract renewals in the next six to twelve months, these are worth working through before the meeting rather than during it.
Which AI agents in your environment are currently accessing licensed data, and under which agreements? In many firms, this inventory doesn't exist. Engineering deployed the agents; legal signed the licences; nobody mapped one to the other.
Do your current contracts contain any provision for autonomous programmatic access by AI systems? Specifically: do they define "approved application" in a way that covers an AI agent running unsupervised overnight workflows?
Can you produce a usage record by agent identity, data source, and timestamp for the past ninety days? This is the evidence question. If the answer is no, or "it would take some engineering work to compile," that's a gap worth addressing before it's requested.
Have you formally declared your AI agent use cases to your data vendors? Not informally, not assumed. In writing, as part of the licence agreement or an addendum to it.
Do your current volume parameters account for agent-level query frequency? Volume caps designed for human analysts are genuinely reachable by an agent running a continuous monitoring workflow. If you haven't modelled this, the first indication may arrive as a back-billing conversation.
If you can answer all five confidently, your compliance posture is ahead of most. If two or three are unclear, the renewal conversation may surface questions you'd rather not answer under commercial pressure.
How Entitle AI helps
This is the problem Entitle AI exists to solve: the gap between what your data licence agreements say and what your AI agents are actually doing.
Our platform maps your AI agent deployments to your actual licence terms: which agents are accessing which datasets, at what volume, under what declared use case, and whether the current licence coverage is adequate. The output is an entitlement exposure report: a clear picture of where your deployments sit relative to your contractual obligations, before a vendor or regulator asks the same question.
For firms with renewals coming up, the 30-day pilot is particularly useful before those conversations. Arriving at a renewal with a clear record of your agent access patterns, rather than trying to reconstruct it under commercial pressure, changes the dynamic of that conversation significantly.
The question worth having an answer to
Before your next vendor renewal or ExCo meeting, one question is worth being able to answer:
Can I produce, on request, a complete record of every dataset our AI agents accessed in the past ninety days, broken down by agent identity, data source, timestamp, and declared use case?
If you can answer it confidently, you're ahead of most firms in the market right now. If the answer is "probably not" or "it would take us a while to pull that together," that's the conversation worth having now, not at renewal.
I'd also genuinely welcome views on this from others who've been through recent data contract renewals or who are thinking through the same questions. The contractual landscape for AI agent data access is evolving in real time and I suspect there are perspectives here I haven't considered.
If you already know you have a gap, we're running a 30-day pilot programme right now. We instrument your existing agent workflows, no production changes required, and deliver a live entitlement exposure report at the end. The report alone is worth the 30 days.