By Ilman Shazhaev, Founder and CEO of Dizzaract
Enterprises are switching AI agents on and pulling them back almost in the same breath, and the numbers behind that habit deserve a closer look.
Among 2,527 senior decision-makers surveyed across 10 countries for Sinch’s The AI Production Paradox report, 62% say their companies already run AI agents in customer communications, yet 74% have rolled back or shut down at least one live agent.
Money keeps flowing all the same, with 98% of surveyed companies raising their AI communications spend in 2026. Among organizations with the most mature guardrails, the rollback rate climbs to 81%, a sign that the teams watching most closely are also finding the most to fix.
The outlook for autonomous agents looks rockier still, since 40% of enterprises are expected to demote or decommission them by 2027 because of governance gaps that surface only after something breaks in production. The root cause, as analysts see it, is a habit of treating each agent as either completely locked down or completely trusted.
Building AI infrastructure teaches you quickly that every answer a model gives ends up as compute on somebody’s invoice. Most of the agents that companies are now pulling back trip over the same two things, namely the freedom they were given and, I’d argue, a bill nobody saw coming. Both problems grow far easier to control once inference runs on distributed infrastructure, which lets a company set access, cost and location for each workload.
Trust Settings Decide Whether an Agent Survives
Most of the agents I hear about run on capable models, and the failures trace back far more often to the trust setting, meaning what an agent may read, change, or spend. Too many teams apply one default to every agent they deploy, which leaves simple bots tangled in red tape and autonomous ones roaming with far too much reach.
When my team designed a network that runs AI inference on other people’s GPUs, full trust was never an option for us. We built it on least privilege, giving each workload the minimum access it needs, with fixed limits on GPU, memory, storage, and network use, and keeping the power to pause or revoke it at any moment.
Enterprises can apply the same discipline to agents, because a bot drafting replies in a support queue needs to read ticket history and has no business issuing refunds or editing customer records. Analysts now recommend grading agents by autonomy level, so read-only helpers face light checks and agents acting on their own face the strictest controls.
Distributed infrastructure makes those grades easier to enforce, because each agent’s workload can be assigned its own pool of hardware with its own limits. Read-only helpers can draw on shared capacity, and an agent that handles customer data can be restricted to controlled machines, with only the data it needs ever reaching the node doing the work.
Each Agent Action Carries an Inference Bill
Every time an agent reasons through a step, calls a tool, or hands work to another agent, a model runs inference and the meter keeps ticking.
Leaked financial documents first obtained by independent journalist Ed Zitron, show that OpenAI brought in about $13 billion in 2025, yet its cost of revenue, the line that tracks running its models, nearly tripled to $7.5 billion. Even after stripping out a one-off accounting charge, the documents put the company’s loss at roughly $8 billion.
If one of the best-funded AI companies finds it hard to make inference pay, an enterprise letting dozens of agents loop through multi-step tasks should brace for surprises on its cloud invoice. Inference is expected to account for more than half of all AI compute by 2030, so this cost line will only grow heavier.
Spreading inference across distributed capacity gives finance teams real levers on that bill, since smaller models can run on a single machine and larger ones can be split across several GPUs. Each agent then draws only the compute its task requires, and the work can run closer to users to keep latency low.
The industry has made real progress on how agents communicate, and the leading protocols for linking agents to tools and to each other now sit under neutral open governance. These standards make agents easier to connect, yet they say nothing about how far each agent should reach or what every exchange between agents will cost.
Any large fleet of well-connected agents with loose permissions multiplies both risks at once. Before scaling, I would want each agent to carry a named owner, a spending ceiling, and a log showing which actions it took and what each one cost.
Four Habits for Running Agents You Can Govern
Start every agent with the narrowest access its task allows, and widen that access in stages once it has proven itself on real work. Running a read-only pilot for a few weeks with clean logs gives you hard evidence for granting write access, and that evidence persuades far better than any vendor demo.
Build the pause and revoke switch before launch day, and test it the way you would test a fire alarm.
Rapid rollback and circuit breakers halt an agent once it crosses a threshold, and a team that has practiced stopping one can act within minutes when something goes wrong. On distributed infrastructure, the switch can sit at the level of each workload, so one misbehaving agent goes quiet and the rest keep running.
Use smaller, task-specific models wherever they do the job, since a model sized for one workflow costs far less to run than a frontier model answering every query. Plenty of companies reaching scale already fine-tune open-source models for narrow business tasks, which keeps the inference bill closer to the value each agent creates.
Avoid any single point of failure, whether it sits in one cloud region, one model provider, or one master agent that every other agent depends on. I believe distributed systems handle outages better because work can move to healthy nodes when one site goes dark, and the same thinking should shape how agents are arranged on top of that compute. Agents built on that footing can carry on through a regional outage without anyone redesigning the workflow behind them.
Governing Agents Before Scaling Them
Rollback figures tell me that companies are learning in production, which is costly but far from hopeless. Businesses that set access deliberately, price every action, and keep a working off switch will have the evidence to hand their agents more responsibility with each quarter.
Those that skip the groundwork will pay twice, first for compute that never turns into value, and then in engineering time. Already, 84% of AI engineering teams spend at least half their time on safety systems instead of building new features. The next agent a company launches is the cheapest place to start doing it differently, beginning with the infrastructure its inference runs on.