As reported alongside their product announcements, GitHub shipped spending controls for its AI development products on July 1, 2026, including automatic model selection that routes tasks to the cheapest capable model, credit pools, per-user spending caps, and cost-center budget allocation. Anthropic shipped Claude Enterprise administrative spend controls on July 2, 2026, including model-level entitlements, an analytics dashboard, and configurable spend-threshold alerts. Both vendors shipped these features in response to enterprise-buyer feedback on cost overruns.

As reported by TechTimes, Uber burned through its entire 2026 AI budget in four months and imposed a US$1,500 monthly cap on individual employee AI spend. Tesla told employees AI spending would be capped at US$200 per week starting July 6, 2026. Microsoft began cancelling internal Claude Code licenses across a division ahead of its June 30 fiscal-year close. Survey figures reported by the Snowflake FinOps blog put roughly 79 percent of enterprises experiencing AI cost overruns in the past twelve months and roughly 98 percent of FinOps teams now managing AI spend, up from roughly 31 percent two years earlier.

Because per-task cost is decoupled from the deliverable (a small change against a large codebase can cost more than a large change against a well-scoped module), retry loops silently multiply spend when an agent iterates through failed attempts, model-routing decisions between frontier, mid-tier and small models are consequential per task, and opaque markups by some partners compound the variance. Traditional hours-based estimation does not capture any of these four.

One, model routing to the cheapest capable model with a written per-task-category policy. Two, per-project spend caps with threshold alerts and hard stops that require explicit re-authorisation to lift. Three, cost-center and per-team allocation of AI spend in the monthly reporting pack. Four, evals and guardrails against wasteful retries and runaway agent loops, with hard retry caps and runaway-loop detection. Five, transparent pass-through of model costs itemised per tier and per project, with any markup capped and defined in the contract. Six, FinOps discipline baked into the delivery contract as a named exhibit with reporting cadence, escalation paths and remedies.

Model costs should be passed through at cost, itemised per model tier, per project and per period, with the underlying provider invoices available on request. If the partner insists on a marked-up service line, the markup percentage should be defined and capped in the contract and the underlying cost should still be itemised so the markup is auditable. Opaque bundled AI line items on an invoice are the pattern to refuse at negotiation.

Call IT Dev operates software development, dedicated development team and AI automation engagements from Morocco with nearshore EU-time-zone delivery, aligned with CNDP Law 09-08 and GDPR obligations. The engagement applies the six-check framework end-to-end: a documented model-routing policy per task category, per-project spend caps with threshold alerts and hard stops, cost-center allocation in the monthly reporting pack, evals on retry counts and token spend per merged change with hard retry caps and runaway-loop detection, transparent pass-through of model costs itemised per tier and per project, and a named FinOps exhibit in the master service agreement. Delivery pods combine senior engineers, an AI tooling lead responsible for the routing and eval layer, and a delivery manager who reports the AI cost line alongside classical delivery metrics.