Allocation policy for AI coding spend
Copilot, Claude Code, Cursor, and Codex are bought side by side and wasted side by side. Spillway moves each builder's limit with demand, under your policy: headroom left in the pool goes to the builders about to hit the wall, and anything past the pool stays capped per person.
Apply, and we will email you to set up a conversation about your organization. We only have 3 spots left for design partners.
Two failures hit one organization in the same month
Budgets, spend limits, and paid-usage policies are all native ceilings, and each one decides when someone stops. Deciding who gets the remaining headroom is the job Spillway does.
On one side of the organization, an allowance resets untouched. That is spend you already made, expiring on schedule, and the invoice never shows it because unused volume has no line item.
32.3% of GitHub licenses sit unused, according to Zylo research
On the other side, your highest-output builder is blocked mid-cycle, waiting for a reset that arrives on the vendor's clock. The other exit is paid overage. GitHub turns it on by default and stopped falling back to cheaper models on June 1. So the wall forces a binary: dead time, or surprise spend.
72% of engineering leaders and practitioners surveyed had an unexpected AI cost spike or bill in the past year, according to Harness
How it works
01
Spillway watches measured, unused allocation across your organization and flows it into a short-lived overflow pool before the reset takes it.
Figure 2 shows the credit spillway across one billing cycle.
02
As a builder approaches the wall, the policy walks the marginal demand ladder: headroom in the overflow pool first, then metered overage for that builder only, under a per-user cap, with queuing as the last resort.
Figure 3 shows the marginal demand ladder: the pool first, then capped overage, then the queue.
03
Only measured, unused volume feeds the pool. Per-user draws are capped, and baselines can differ by role, so heavy-need users get headroom while every draw stays bounded.
Figure 4 compares a flat cap with role baselines and capped draws.
Field data
32.3%
of GitHub licenses sit unused, according to Zylo research
44%
is the drop in included Copilot Enterprise credits per seat after the June through August promotion ended, according to GitHub
28 days
is all GitHub's usage dashboard shows, according to GitHub documentation
17%
is the cut in Claude Code weekly limits from September 14, compared with the May through September promotion, according to Anthropic
Connections
Each connection reads usage metadata and sets per-user limits through the vendor's own administration API. For Claude Code, Cursor, and Codex, those APIs are Enterprise-only. Spillway never touches repositories, prompt contents, or code.
What we touch
No repository access
Reads usage metadata only. It never sees prompt contents, code, or repositories.
Every write is logged
A full audit ledger records who got refilled, when, how much, and under which rule.
Lag-respecting policies
Vendor data lags by hours to days, so policies pace their changes around that lag.
One click to revoke
You're back on vendor-native ceilings. Override policies as needed.
Least-privilege tokens
Asks for the narrowest scope each vendor offers for reading usage and writing per-user limits.
Seats and plans untouched
It never adds or removes seats, and it never changes a plan or a payment method.
No lock-in
Revoking access returns every account to vendor-native ceilings. Spillway keeps no residual access, and the audit ledger stays with you as your record.
Apply, and we will email you to set up a conversation about your organization. We only have 3 spots left for design partners.
Questions, answered
Spillway turns the credits you already bought into headroom for the builders who need them. Unused allocation flows into a pool before the reset forfeits it, builders approaching their wall get refilled automatically, and every draw stays under the caps your policy sets, so the budget holds while demand still moves. It applies the same per-user limits you could set by hand, logs each draw with the rule that authorized it, and works across every vendor from a single policy engine.
Spillway connects to your vendors’ administration APIs, reads usage metadata, and writes per-user limits. When measured, unused allocation is about to expire, it flows into a short-lived overflow pool. Your policy then refills users approaching their wall: small draws happen automatically, and large ones go through a visible request board.
GitHub Copilot is connected in design partner early access, and the free Copilot usage report is available today as a read-only GitHub App. Claude Code, Cursor, OpenAI Codex, and Figma are evaluating demand, and demand decides the roadmap.
No, nobody can hoard it. Per-user draw caps and role-based baselines are enforced by the policy engine. Draws are bounded per cycle, the pool itself is short-lived, and every draw is logged with the rule that authorized it.
The policy moves each builder through two steps: metered overage for that builder only, under a per-user cap your policy sets, and then a queue for anything past that cap. Without a policy, GitHub’s defaults leave two outcomes at the edge of the pool: overage for everyone, because paid usage is on by default, or, with paid usage off, a stop for everyone except code completions. Since June 1, GitHub has had no automatic fallback to a cheaper model.
Refills happen through the vendor’s own administration API. When your policy fires, Spillway writes a new per-user limit for that user. It is the same control you could set by hand, applied automatically on the schedule you defined, with an audit trail.
A cap decides when a user stops. Deciding who gets the remaining headroom takes a policy that moves those caps during the month. Fixed budgets leave both failures in place, with credits expiring on one side while builders are blocked on the other. Spillway keeps your ceilings and routes the headroom. We wrote up what each control does and where it stops.
Revoking a change reverts to vendor-native ceilings. Spillway keeps no residual access and leaves no stranded configuration, and the audit ledger remains as your record.
Resources
These research notes on each vendor's billing mechanics are written for administrators. Each note covers what changed, when it changed, and what you can do about it.
Copilot · Billing
GitHub Copilot premium requests and AI credits, explained
What changed on September 1, 2026, and what to check before the October invoice.
Copilot · Budget
GitHub Copilot budget controls: working with per-user caps
How each control works, how far it reaches, and what a cap actually moves.
Copilot · Billing
Out of Copilot credits mid-month: what a developer can do
What the credit wall is, what you can see about your own usage, and what to ask for before the reset.
Claude Code · Billing
Claude Code usage limits and credits on Team and Enterprise plans, explained
How the session and weekly limits work, and what changed on September 14, 2026.
Claude Code · Budget
Claude Code spend limits: working with per-member caps
The seat allowance, usage credits, spend limits and the Enterprise APIs, one at a time.
Cursor · Billing
Cursor usage limits, weekly included usage and prepaid credit on Teams and Enterprise, explained
The weekly pool per seat, the team's prepaid balance, the Token Rate, and the move to The X Subscription.
Codex · Billing
OpenAI Codex usage limits, credits and plans on Business and Enterprise, explained
The five-hour and weekly windows, workspace credits per million tokens, and what changed on April 2 and August 24, 2026.
Copilot · Tool
Copilot credits calculator
Size your monthly pool, the credits that expire at the reset, and the cost of drawing past it.
The clock is ticking
Be one of the first to put allocation policy on your AI credits as a design partner. Join the waitlist to be notified of product and feature availability.
Apply, and we will email you to set up a conversation about your organization. We only have 3 spots left for design partners.