The agent-runtime billing problem: why per-seat pricing is cracking as buy-side agents consume research APIs at machine speed

Per-seat licensing has been the default commercial shape for buy-side research vendors for close to two decades. A Bloomberg terminal is a seat. A transcript library login is a seat. An expert-network membership is a seat. The arithmetic has always been simple: the vendor sells a number of logins, the client distributes those logins to named humans, and the query load per seat sits inside a predictable band shaped by how fast an analyst can read, click, and type.

That band is no longer the binding constraint. The consumer of the API is increasingly a machine orchestrated by a human, not a human directly, and the orchestration layer can issue queries at a rate that has nothing to do with how fast an analyst can read. Our read is that per-seat pricing on buy-side research vendors is in the early innings of a structural reset, and the vendors that handle the transition well will convert what looks like a pricing crisis into the most durable revenue line they have ever had.

The assumption that is actually breaking

Per-seat pricing is not failing because seats are the wrong unit of measurement in some abstract sense. It is failing because two assumptions that have held for twenty years are coming apart at the same time.

The first assumption is that one seat equals one human. A hedge fund that buys forty seats of a transcript library is saying something specific: forty analysts will log in, read transcripts, and make investment decisions downstream. The vendor's capacity planning, infrastructure cost, and sales motion are all built around that mapping. When one of those forty seats is actually a Rogo or Hebbia agent orchestrated by one of the forty humans, the mapping is already broken, even if the login credential looks identical to the vendor on the other end.

The second assumption is that query load per seat sits inside a human-speed band. An analyst reading transcripts and running searches generates somewhere between dozens and low hundreds of queries per day, with heavy clustering around earnings season. The band is wide but it is bounded by human reading speed, meeting load, and the simple ceiling of how many hours a person works. Agent runtimes do not have that ceiling. A single analyst running a research workflow through an orchestration layer can comfortably generate query volumes that would require a team of fifty humans to produce by hand, and the agent does not stop at 6pm or take a break during earnings week.

Both assumptions break at the same time, and they compound. A seat that was supposed to be one human generating a human-speed query load is now a mixed human-plus-agent entity generating a machine-speed query load. The vendor's revenue per seat is unchanged. The vendor's infrastructure cost per seat has multiplied. The math stops working somewhere between quietly and loudly, depending on how aggressively the client's agents are configured.

Why MCP accelerated what was already coming

Agent-driven consumption of research APIs is not a new idea. Quants have been pulling market data programmatically for decades, and the sell side has had systematic access to transcript corpora for years. What is new is how little friction now sits between a general-purpose agent and a vendor's data.

Model Context Protocol has moved from an Anthropic specification to enterprise wiring in under a year. The practical effect is that an agent runtime no longer needs a bespoke integration to call an expert-network transcript endpoint or a market-data feed. A vendor that exposes an MCP server is reachable by any MCP-aware agent, which in practice now means most of the serious orchestration layers in the buy-side stack. The protocol does not create agent-driven consumption, but it removes the integration tax that used to keep machine query load concentrated in the hands of in-house quant teams. The buy-side analyst at a long-only shop can now point an off-the-shelf agent at a vendor API with roughly the effort it used to take to configure an Outlook plugin.

That shift matters because it moves agent consumption from a quant-desk anomaly into the mainline fundamental workflow. The analyst who used to click through transcripts manually is now the same analyst running a Hebbia Max agent with persistent memory and a growing set of data connectors, or a Rogo agent sweeping across filings, transcripts, and expert-network summaries in parallel. The query load profile on the vendor side changes shape even though the client roster, and the seat count, has not changed at all.

The four commercial shapes emerging

Our read of the public signals is that vendors are converging on four distinct commercial shapes, and most will land on a hybrid of two or three rather than a single pure form.

Metered API pricing is the cleanest alignment of revenue with cost. The vendor charges per query, per token returned, or per unit of compute consumed on their side. This is the model that cloud infrastructure and foundation-model vendors have run on for years, and it is the easiest to defend when a single client's agent load spikes 100x in a quarter. The downside is forecast volatility: metered revenue is harder to model into a four-quarter plan, and the vendor loses the predictability that per-seat subscriptions give its own CFO and board. Metered pricing also puts friction back into the client workflow, because every agent query is now a line item that a procurement team can see and question.

Consumption tiers are the familiar SaaS compromise. The vendor sells a plan that includes up to N queries per month, with overage billed metered. This preserves the subscription shape for revenue recognition and for the client's budgeting cycle while still capturing upside when agent load pushes past the tier ceiling. The weakness is that tier design is hard: pick ceilings too low and clients perceive the move as a stealth price hike, pick them too high and the vendor absorbs the agent-load cost without compensating revenue. Consumption tiers also create a predictable gaming pattern where clients configure agents to stay just inside the tier ceiling, which is operationally fine but caps the upside the vendor was trying to capture.

Agent-seat licensing keeps the per-seat shape but adds a new seat type: a seat priced for a machine rather than a human, typically at a multiple of a human seat. This is the least disruptive move for a vendor whose sales motion, contract paper, and renewal process are all built around seat counts. The weakness is that it still assumes a predictable per-seat query load, which is the assumption that broke in the first place. An agent seat priced at 10x a human seat is a reasonable guess at the right multiple if the agent generates 10x the query load; it is a disaster if the agent generates 200x.

Hybrid floors with usage overage are the shape we expect to dominate the next 24 months. A client commits to a minimum annual spend that preserves the vendor's forecasting discipline, and anything above the floor is metered. The floor is high enough to protect base revenue and low enough that clients will clear it without negotiation friction. This is roughly the shape cloud vendors landed on with enterprise committed-use contracts, and it is the shape that balances the vendor's forecasting needs against the reality that agent query load is now the main variable driver of cost.

What the public signals are already telling us

The pricing conversation is not hypothetical. The strategic rounds and product moves of the last twelve months are already assuming programmatic access patterns rather than terminal clicks, which is a strong signal about where commercial contracts are heading.

Rogo closed a USD 160M Series D in 2026 with Barclays leading a bank strategic round. The bank-strategic signal is meaningful on its own: when a tier-one investment bank anchors a round at a buy-side-adjacent agent company, it is a vote that the agent orchestration layer will be a durable piece of the sell-side workflow too. CEO Gabe Stengel's 10,000-agent framing on the O'Shaughnessy podcast is a useful anchor for the magnitude question. If a single analyst is orchestrating thousands of agent queries in parallel across filings, transcripts, and expert-network content, the query load generated by one Rogo seat sits in a different order of magnitude than one Bloomberg terminal seat. Any vendor whose data flows through that orchestration is going to feel the shift in their infrastructure bill before they feel it in their revenue line.

Hebbia's October 2026 update added persistent memory to its Max agent and expanded Microsoft Word and data connector support. Persistent memory is the quiet feature in that release. An agent with persistent memory can keep returning to the same source APIs over the life of a research project rather than loading context from scratch on each run, which multiplies the per-seat query load against a vendor's transcript or expert-network corpus. The connector expansion extends the pattern across more vendor surfaces. Neither move is dramatic in isolation. Together they describe a workflow where the agent is the primary API consumer and the human is the orchestrator and reviewer.

OpenAI's ChatGPT for Financial Services brings a general-purpose foundation-model vendor into the same workflow lane that Rogo and Hebbia are building. The practical implication is that agent-driven research consumption is no longer a specialist buy-side-fintech story, it is a motion that any serious model vendor can enter. And Bridgewater CIO Greg Jensen's Odd Lots comments that AI already sits beside the equity analyst process is a signal from the top of the buy-side pyramid that agent orchestration is operational, not pilot.

The common thread across all of these signals is programmatic access as the default assumption. None of them scale on terminal clicks. All of them assume the client will consume the vendor's data through an orchestration layer rather than through a browser session a human is driving directly.

The counterargument: why per-seat does not die outright

It would be easy to read this trend as the end of per-seat pricing. We think that overshoots the signal.

The per-seat contract has survived two decades for reasons that are not just inertial. It gives the client a predictable budget line that a CFO can defend at renewal. It gives the vendor a forecastable revenue stream that supports a public-company valuation multiple. It aligns neatly with how enterprise procurement is structured, with contracts negotiated once a year by procurement teams that do not want to re-engage mid-year on usage overages. And it maps cleanly to compliance and MNPI controls: a named human on a seat is accountable, auditable, and surveillable in a way that a shared agent credential is not.

The compliance point is the one that is often underweighted in this conversation. Expert-network vendors in particular have invested heavily in controls that identify which named human on which named account consumed which transcript. When an agent is the consumer, the trail does not disappear, but it does change shape. The human on the orchestration end is still the accountable party, but the mapping between that human and the specific queries the agent ran on their behalf has to be reconstructed from logs rather than observed directly. Vendors with strong compliance postures will not want to blur that line any further than they have to, and per-seat identity gives them an anchor they are reluctant to let go of.

Our view is that per-seat does not die, it narrows. It survives as the identity and compliance layer and as the base of a hybrid floor contract. What dies is the assumption that per-seat revenue scales with per-seat value delivered. The value scales with query volume, and the pricing contract will eventually scale with it too, through overage, tiers, or metered add-ons layered on top of a per-seat identity base.

A worked example: an expert-network vendor on the pricing frontier

It is worth working through what this looks like at a specific vendor type, because the right pricing shape depends on where the cost sits.

Consider an expert-network vendor whose main asset is a transcript library plus a scheduling layer for expert calls. The transcript library is the piece most exposed to agent-driven consumption. Transcripts are structured, searchable, and the natural food of a research agent that is building context on a company or sector. A hedge-fund client with forty analyst seats historically might generate on the order of ten to twenty thousand transcript queries a month across the firm. The same client, with the same forty seats, running Rogo or Hebbia or an internal MCP-connected agent across those seats, can plausibly generate ten to twenty times that volume within a few quarters of agent adoption, concentrated on a handful of the most-used seats.

The vendor's infrastructure cost scales roughly with query volume: storage is cheap, but search, retrieval, and especially any LLM-assisted summarization layer on top of the transcript are not. The vendor's revenue, under a pure per-seat contract, is unchanged. The gross margin on that client compresses, potentially sharply, and the vendor absorbs the full cost of the client's workflow transition.

The cleanest fix is a hybrid floor. The forty seats remain, repriced to include a monthly query allowance that comfortably covers the pre-agent workflow. Query volume above the allowance is metered at a per-query or per-token rate that reflects the vendor's marginal cost plus a reasonable margin. The compliance identity stays attached to the human seat. The forecastable base revenue stays attached to the floor. The agent-driven upside is captured through overage, which grows as the client's agent adoption grows. This is not an elegant contract, but it is a durable one, and it is the shape we expect most of the serious research vendors to converge on over the next 24 months.

For the scheduling and expert-call side of the same vendor, the picture is different. Expert calls are a human-scarce resource rather than a data resource, and the natural pricing shape there (per-call or credit-based) already roughly fits an agent-driven workflow. The agent can schedule and route calls faster than a human, but the ceiling is set by expert availability, not by vendor infrastructure. Per-call pricing was already metered consumption pricing in disguise; it just has not been described that way.

What to watch next

The shift from per-seat to metered-or-hybrid pricing will not happen in a single quarter, and it will not happen uniformly across the research-vendor stack. The vendors with the most agent-exposed data (transcript libraries, expert-call summaries, structured filings) will move first. The vendors whose data is already consumed programmatically (market data, pricing feeds) have mostly already moved. The vendors whose data is still consumed through a terminal UI, with Bloomberg as the obvious example, will move last and will defend the per-seat shape longest because their switching costs are the highest and their compliance moat is the deepest.

The signals we are watching are specific. Renewal cycles at the big transcript and expert-network vendors are the first place a hybrid floor shape will show up, because annual contract repricing is the natural moment to introduce an allowance plus overage structure. Earnings disclosures from the public research vendors will start to split out usage-based revenue as a distinct line, which is both an operational necessity and a signal to the market that the business model is changing. Agent-vendor contracts themselves (Rogo, Hebbia, and the orchestration layer that OpenAI is building into ChatGPT for Financial Services) will start to disclose or discuss their own pass-through economics on vendor data, which will shape how clients negotiate with the underlying vendors.

The deeper question underneath all of this is whether research-vendor revenue durability improves or worsens in the agent era. Our view is that it improves, counterintuitively. Metered and hybrid pricing aligns vendor revenue with the value delivered through client workflows rather than with the number of humans a client happens to employ, and that alignment is more defensible at renewal, more scalable as client agent adoption grows, and more resilient to the kind of seat-count cuts that have historically compressed vendor revenue during buy-side cost-cutting cycles. The transition is painful because it forces vendors to rebuild their forecasting, their sales motion, and their contract paper at the same time. The destination is a healthier business.