Devin Gets Hierarchical Budgets: AI Coding Agent Spend Now Follows the Org Chart

Cognition has given enterprise customers of its Devin coding agent a new way to control what the agent costs: hierarchical budgets that split one company-wide allowance into department, project and team pools, each with its own owner and its own limit. The feature went live on November 2, 2026 for customers on a Devin Enterprise plan, and it comes with a hard edge that developers will notice. If a budget runs out while Devin is in the middle of a task, the session stops.

The launch is a sign of where AI coding tools are heading. Agents that run for minutes or hours, open pull requests and investigate bugs on their own are billed by consumption, not by seat. That makes them harder to forecast than a fixed licence, and finance teams are starting to treat agent spend the way they learned to treat cloud bills a decade ago.

Key facts at a glance

• Launched November 2, 2026 for Devin Enterprise customers, under Enterprise Settings, then Usage Policies, then Budgets.
• One root budget, child budgets up to three levels deep. Each budget sets a monthly limit in Agent Compute Units (ACUs), Cognition's billing unit that combines model and cloud infrastructure cost.
• Temporary bumps last until the end of the current monthly window, and budget owners can grant them to the teams beneath them.
• When a budget is exhausted, the running session stops. Engineers get warnings first and can request more ACUs from inside the product.
• Builds on October's Efficiency Score, a 0 to 100 rating that can automatically raise limits for engineers whose usage is graded Healthy or Satisfactory.

Why Cognition built it

Before this release, Devin's spending controls were fairly flat: organisation-level limits, per-user usage tiers and separate settings for the desktop app. That works for one team running a pilot. It breaks down when a large company signs a big ACU commitment and needs to divide it among dozens or hundreds of groups.

Lukas Burger, who works across go-to-market and engineering at Cognition, said the feature came directly from the company's forward-deployed engineers, who work on-site with customers. "That's fine for one team. It doesn't hold up when a bank is dividing a large ACU commitment across hundreds of groups," he said of the older controls. Some customers had started building their own internal tools just to share capacity among business-line leaders, a gap Cognition says it designed budgets to close.

"Devin stops belonging to one team," Burger said. "A pilot might sit with a single platform group. A broad rollout means many departments, each with its own priorities and its own budget owner."

How the budget tree works

Every enterprise has a single Account budget at the root. Child budgets sit beneath it, up to three levels deep. A budget can contain one Devin organisation or several, but each organisation belongs to exactly one budget.

According to Cognition, each budget carries five things:

• A limit: the ACUs the budget and everything under it can use in a monthly window.
• An optional temporary limit: a bump that expires at the end of the current window.
• A parent: whose administrator controls the child's limit.
• Its own user limits: usage tiers and per-person overrides that apply only inside that budget.
• Its own people: who can draw from it, adjust child limits and pass permissions on.

The company's own worked example shows how delegation plays out. A central IT administrator sets each department's total. An engineering head can then manage every budget under engineering, for instance giving a payments team a temporary 5,000 ACU increase, taking it from 10,000 to 15,000 for a busy launch week, while the whole engineering tree stays inside its combined 30,000 ACU limit. Department heads cannot see other departments' budgets, so if engineering needs more, the request goes back up to central IT.

A second pattern is the project bucket. Cognition describes a security hardening push with its own 5,000 ACU budget and a more generous default allowance of 1,000 ACUs per engineer, separate from everyday developer budgets set at 200 ACUs per person. The point is attribution: spend in the project budget can be tied directly to that initiative and audited on its own.

Warnings, requests and the Efficiency Score

The hard stop is softened by several layers that sit in front of it. Devin's usage policy documentation says members get a notification at around two-thirds of their limit, then persistent banners as they approach and reach it. From there they can request more ACUs, with a reason, without leaving the product. Requested increases are always temporary and expire at the next monthly reset.

Each tier's administrator picks one of three approval policies: manual approval, always approve up to a ceiling, or approve based on efficiency. The last option relies on the Efficiency Score Cognition introduced on October 19. The score estimates how many productive engineering hours each session produced relative to the ACUs it consumed, weights that efficiency more heavily than raw volume, and grades each engineer as Healthy, Satisfactory, Needs Improvement or Unknown.

Engineers graded Healthy or Satisfactory can have their limits raised automatically, up to a ceiling the admin sets. Cognition stresses that auto-approval only works in the engineer's favour: requests are never denied automatically, and anything above the ceiling waits for a human. In its October announcement, Cognition said one heavy user at a company with auto-approvals hit their cap 22 times in a period and had it lifted each time, while a comparable engineer elsewhere waited four days for a manual approval.

"We designed Budgets so engineers can do their work without thinking about them," Burger said.

What it means for you

If you are an engineer using Devin at work, expect your organisation's spend to become more visible and more segmented. You may have one allowance for day-to-day work and a separate one tied to a specific project. Watch for the two-thirds warning, and request more before you start a long task rather than after a session has been cut off. Because your Efficiency Score can decide whether increases are automatic, habits such as scoping tasks tightly and not defaulting to the most expensive mode now have a direct effect on how often you get blocked.

If you manage a team or a budget, the new model lets you hand out capacity without routing every request through central IT. Each Devin session is tied to a specific budget, so spend can be traced to a team, project or business unit, and some customers already map it to internal billing codes or charge it back to departments. Cognition recommends defaults with enough headroom that productive users are not blocked unnecessarily. "We also don't want controls so tight that customers stop being productive with Devin," Burger said.

If you work in finance, this is part of a wider shift. The FinOps Foundation's 2026 State of FinOps report found that 98% of its 1,192 survey respondents now manage AI spending, up from 31% two years earlier. Burger expects the job to land with strategic finance and planning teams: "On top of their day jobs, they're now managing massive AI budgets and making sure that spend shows up as real impact on the P&L."

The other side of the argument

Not everything about the design will please developers. A session that stops when a pool empties can cost real time if it happens midway through a multi-step change, and Cognition concedes that the safeguards are mainly warnings and faster approvals. A softer fallback is still in development. "We're also working on letting teams switch to lower-cost models after a soft limit, so important work can finish," Burger said.

The Efficiency Score also invites scrutiny. It estimates productive hours, using merged pull requests and a classifier that judges whether sessions without a pull request made real progress. That is a model of value, not a direct measure of code quality or business results, and linking it to who gets more compute means engineers are, in effect, being graded by the vendor's own metric. Cognition says it validated the hours estimates against engineers' own assessments, and that the score uses recent usage so early mistakes do not count forever. Teams will still want to compare it with measures they already trust, such as merged pull requests and DORA metrics.

There is also a commercial angle. Under consumption pricing, customers who cannot see value stop paying, and Burger was open about that: "With consumption pricing, customers who can't see value stop using the product, so we've put a lot of work into cost control and measurement."

What happens next

The soft-limit option to fall back to cheaper models is the most important open item for anyone worried about interrupted sessions; Cognition has not given a date. Per-user ACU limits, which work alongside budgets, are still marked as beta and need to be switched on by a customer's account team. Limits and approved increases reset at the start of each monthly billing window, which follows each customer's contract date, so the first full cycle under the new controls will show how often teams actually hit the wall.

More broadly, expect rival coding agents that bill by usage to face the same pressure from enterprise buyers for budget trees, chargeback and approval workflows.

Takeaways

• Devin spend can now be split like a company budget: one root, child budgets up to three levels deep, each with an owner and a monthly ACU limit.
• Sessions stop when a budget is exhausted, so the two-thirds warning and in-product requests matter; a cheaper-model fallback is planned but not yet available.
• Efficiency now affects access. Engineers graded Healthy or Satisfactory can get automatic increases up to an admin-set ceiling; others wait for review.
• Project buckets make spend traceable, which helps chargeback but also means every initiative's agent costs become visible.
• AI spend is becoming a finance discipline, with 98% of surveyed FinOps practitioners now managing it.