There’s a version of the agentic AI story that every vendor is telling right now, and it’s about capability. Agents that book the meeting, reconcile the invoice, update the CRM, file the ticket, all talking to each other across your stack without a human in the loop. The demos are genuinely impressive. The pricing pages are reassuring. And the part that almost nobody is putting on a slide is that the thing standing between an enterprise and that future isn’t model capability. It’s the toll booth that gets installed every time two systems need to exchange data.

Constellation Research put this about as plainly as anyone has. In their 2026 enterprise tech predictions, analyst Larry Dignan argued that data fees are going to be the biggest risk to scaling AI agents, and that connection fees are going to be the new cloud egress. That comparison is worth unpacking, because anyone who has run a cloud bill knows exactly how the egress story went, and it didn’t go well for the buyer.

The egress playbook, repeated

Cloud egress fees are the charges you pay to move data out of a provider’s environment. Storing data is cheap, sometimes free. Moving it somewhere else costs money, and the pricing is structured so that the more entangled you get, the more expensive leaving becomes. It’s a lock-in mechanism dressed up as a line item.

The numbers on the AI side are already meaningful. A typical enterprise AI project moves 10 gigabytes to 10 terabytes of training data, and with the data duplication that real pipelines require, the egress bill for a single model can reach $2,700, with larger projects hitting tens of thousands per training run. That same analysis notes a team moving 50 terabytes a month pays roughly $4,500 in egress alone, a recurring cost that compounds into six figures annually. None of that is compute. None of it is storage. It’s purely the cost of moving data you already own from one place to another.

Google’s Vertex AI, for example, charges $0.12 per gigabyte to move data to most destinations and $0.23 per gigabyte to some regions, which means a team exporting model artifacts daily for a multi-cloud deployment racks up charges that exist for no reason other than the data crossing a boundary. The fix the vendor recommends, predictably, is to keep everything inside their environment.

Now take that dynamic and apply it to agents, which are defined by the number of connections they make.

Why agents make this worse

A single AI agent isn’t worth much in isolation. The value comes from connection: an agent that can read your CRM, query your data warehouse, check your ticketing system, and act across all of them. The emerging standard for those connections is the Model Context Protocol, and its adoption curve has been steep. Since 2024, the MCP ecosystem has grown from three servers to over 6,800 active deployments, and in large enterprises more than 15% of employees now run at least one MCP server. The MCP Dev Summit in New York in early April drew 1,200 attendees, with Amazon and Uber presenting production deployment case studies, and the protocol now has native support across products from Anthropic, OpenAI, Microsoft, and Amazon.

That breadth of adoption is the good news. It’s also the setup for the toll problem. Every one of those connections is a place where a vendor can decide that data passing through is a billable event. And the architecture of agentic systems means connections multiply fast. An agent doesn’t make one call, it makes hundreds, chaining tool calls together to complete a single task.

The cost mechanics here are already visible to anyone running multi-server agent setups. Analysis of connecting multiple MCP servers to a single agent found that paid external APIs invoked through MCP tools accrue charges that are hard to attribute to specific agents or workflows, creating a cost-visibility problem on top of the cost itself. The same analysis found that input token usage alone dropped 58% when a setup was trimmed from a sprawling tool list down to a focused one, and 84% in a larger case. The point being that the connections themselves carry a running cost, separate from whatever the vendor decides to charge for the privilege of making them.

There’s also a commercialization wave building specifically around charging for these connections. Companies like Nevermined are now building payment layers that wrap MCP servers with paywalls, validating subscriptions and burning credits before executing each tool call, with sub-cent micropayments designed to make per-tool-call billing economically viable at scale. The technology to meter and charge for every single agent action is being built right now, and it’s being marketed to the vendors, not the buyers.

The thing enterprises keep getting wrong about ownership

The assumption baked into most enterprise AI planning is that the data belongs to the enterprise, so moving it around is free or close to it. That assumption is going to get tested. As Constellation’s analysis put it, enterprises need to keep in mind that they own the data, but in some cases the vendor may feel otherwise.

This is the part that maps almost exactly onto the early cloud era. Companies moved to the cloud for flexibility and ended up discovering that the flexibility was asymmetric. You could get in cheaply and scale fast, but getting your own data back out, or moving it to a competitor, carried a toll that grew with your dependence. The egress fee wasn’t really about the cost of bandwidth. It was about making the exit expensive enough that you didn’t take it.

Agentic AI is set up to repeat this, with a twist that makes it harder to see. With cloud egress, at least the line item was explicit and you could model it. With agent connection fees, the cost is distributed across thousands of small tool calls, attributed to no single workflow, and buried inside a consumption bill that’s already hard to forecast. The toll is the same idea, but it’s been atomized to the point where most finance teams won’t be able to find it until the invoice arrives.

What the pricing models are signaling

It helps to look at where vendor pricing is heading, because the direction tells you the intent. Constellation predicts that agentic enterprise license agreements will become the norm as CxOs push back on per-seat models, and warns that SaaS providers may sign these agreements at a loss specifically to play for the renewal, the moment when you’re completely locked in. That’s not a hypothetical. It’s the same teaser-rate strategy that’s existed in enterprise software forever, now applied to AI.

The broader pricing environment reinforces the problem. Most enterprise AI vendors don’t publish prices at all, because custom pricing lets them charge different amounts based on company size, competitive situation, and negotiation leverage. The pricing units themselves are also drifting in a direction that obscures cost: tokens become credits, credits become “intelligence units,” and each abstraction makes it a little harder to compare what you’re paying against what you’re getting. When the unit of billing is deliberately fuzzy and the connections being billed are deliberately numerous, the buyer is at a structural disadvantage.

The current industry conversation is mostly about whether AI agents work. The more useful conversation, the one that’s going to matter at renewal time, is about who controls the connective tissue between systems and what they’re allowed to charge for it.

The defensive moves worth making

The defensive moves here aren’t exotic, they’re just unglamorous and they require treating this as a procurement problem rather than a technology one.

The first thing is to map your data movement before you scale anything. Most enterprises know what their agents do but not how many times those agents move or touch data to do it, which means they can’t forecast the bill. Treating connection volume as a first-class metric, the way mature cloud teams eventually learned to treat egress, is the baseline.

The second thing is to push hard on data portability terms at contract time, not at renewal. The leverage you have is highest before you’ve built fifty workflows on top of a vendor’s connectors. Ask explicitly what it costs to move data out, what it costs to connect to a competitor’s tools, and whether per-connection or per-call fees can escalate during the contract term. The vendors that give you clean answers are telling you something, and so are the ones that don’t.

The third thing is to take ownership of the integration layer where it matters most. Open-source MCP tooling exists and is often better built than the bundled commercial equivalents, in part because the open-source projects had to earn adoption rather than bundle it. An enterprise that owns its own connectors for its highest-volume data flows is an enterprise that can’t be tolled on those flows later. This is exactly the kind of thing that’s easier with partners who understand the integration economics and not just the agent demos, because the decision about what to own versus what to rent is the whole game.

The agent capabilities are real and they’re going to keep getting better. That was never the question worth worrying about. The question worth worrying about is the one the cloud era already answered once: when a vendor controls the path your data travels, they eventually charge you for the trip. The enterprises that learned that lesson the expensive way with cloud egress have a chance to not learn it twice. Most of them probably will anyway.