AI coding tools are now a funded line item
LeadDev's September 2026 article on AI spend reports two adoption figures from its own survey. "LeadDev's AI Impact Report 2026 found that 90% of organizations using AI tools now pay for AI models, moving them beyond experimentation and into funded, sanctioned use." And: "Around two-thirds also pay for both autonomous coding agents, such as Claude Code and Devin, and AI-powered code completion built into traditional development environments."
A funded tool comes up for renewal. For many engineering organizations the next renewal is the first one where the AI coding budget is large enough for finance to ask what it bought, and where the bill covers two product categories with different cost structures.
The AI engineering return on investment (ROI) guide already lists six diagnostic questions in its section "What to check before renewing an AI coding budget". This guide covers the step before those questions: the procurement and evidence plan that lets a chief technology officer (CTO) answer them with data on the day the contract comes up, rather than with the vendor's slide.
Why most renewals run on the wrong evidence
Most organizations bought their first AI coding seats as a trial and expanded them without recording where delivery stood beforehand. At renewal, two kinds of evidence are on the table. The vendor brings usage data from its own admin dashboard. The engineering team brings impressions. Neither shows whether the organization ships more working software per dollar than it did before the tools arrived.
That gap matters because generation and delivery move at different rates. Coding activity can climb much faster than releases, a pattern covered in AI throughput is not engineering throughput. If you judge a renewal on activity alone, it can look like a success even when delivery has not moved.
Step 1: Set a baseline before rollout
A baseline is a snapshot of delivery and cost taken before the tool changes anything. Take it before a new rollout, and again before any expansion to new teams or to a higher-priced plan, such as one that adds agent features. If the tools are already everywhere, your version control and continuous integration (CI) history still holds the pre-rollout record: pull it now, before log retention deletes it.
Record at least these, for the trailing quarter, per team:
- Delivery flow: merged pull requests (PRs), lead time from first commit to production, and deployment frequency.
- Review load: median time to first review, review rounds per PR, and how many reviews each senior engineer handles in a week.
- Quality: change failure rate, reverts, and the share of merged changes rewritten within 30 days.
- Allocation: the share of engineering time spent on new features. LeadDev's article recommends tracking this "innovation ratio" over time.
- Cost: seat licences, metered usage, CI minutes, and any review tooling billed separately.
Keep one or two teams on a later rollout date if you can. A comparison group helps separate the tool's effect from a reorganisation, a hiring freeze or a quieter quarter, which before-and-after data alone often cannot do.
Step 2: Decide what evidence to collect over the contract term
Decide the evidence plan when you sign, not when you renew. The vendor's data and your own systems answer different questions, so assign each item a source and an owner.
| Evidence | Source system | Cadence |
|---|---|---|
| Delivery flow against baseline | Version control and deploy pipeline | Monthly |
| Review load per senior engineer | Code review tool | Monthly |
| Rework and reverts | Version control, incident tracker | Monthly |
| Total cost per merged change | Vendor billing plus CI billing | Monthly |
| Architectural corrections in review | Review comments, sampled by senior reviewers | Quarterly |
| Innovation ratio | Issue tracker or time allocation | Quarterly |
| Usage and adoption | Vendor dashboard or API | Monthly |
Two contract terms make this plan workable. First, ask for usage data through an export or API at the user and team level, with retention that covers the full term. Some vendor figures exist only on screen, outside both the API and the newline-delimited JSON (NDJSON) bulk exports. GitHub's documentation says the ROI estimates in its Copilot impact dashboard "are available only in the dashboard and are not included in the Copilot usage metrics API or NDJSON exports." Second, ask for billing data at the same team granularity as your delivery data, so you can divide one by the other.
"Total cost per merged change" is the number that ties the plan together. It counts every cost the tool causes in a period, divided by all changes merged to the main branch in that period, not only AI-assisted ones, so the tool's effect on the rest of the team's work stays in the result. It falls when the same spend ships more changes. Under usage-based billing it also rises when agents generate work that is abandoned, because those sessions are billed; under a flat seat price, abandoned work shows up only as fewer merged changes for the same spend.
Step 3: Separate metrics that measure use from metrics that measure value
Vendor dashboards report what they count. The trouble starts when a renewal deck presents use as value. Microsoft's EngThrive framework, published in May 2026 by Tim Bozarth and David Liu, treats productivity as a system of speed, ease and quality, with developer thriving as a guardrail. Its authors put the distinction in one line: "Activity metrics are not outcome metrics."
GitHub's Copilot usage metrics documentation is a useful reference, because it states the limits of its own metrics. Two of its terms need defining. GitHub groups developers into cohorts by depth of Copilot use: engaged users are those in its code-first, agent-first or multi-agent cohorts, and passive users are the rest. Its adoption multiplier divides the average number of PRs merged by engaged users by the average for passive users. Sorted by what each one can support at renewal:
| Metric | What it shows | What it cannot show |
|---|---|---|
| Daily active users, adoption cohorts | Whether licences are used | Whether use improved delivery |
| Inline suggestion acceptance rate | Whether suggestions look relevant to developers | Whether accepted code survived review |
| Lines suggested, added and deleted | Volume of output; GitHub's documentation calls it "a directional view of Copilot's tangible output" | Whether the volume was needed |
| PR creation and merge counts, median time to merge | Delivery flow for AI-assisted work | Rework, reverts and incidents after merge |
| Adoption multiplier, potential ROI estimate | Engaged users merge more PRs than passive ones | Cause. GitHub's own description: "This is a measurement to understand the potential lift that becoming a more engaged user can provide." |
Use metrics still earn a place in the renewal file. Low daily active use on a seat licence is a direct cost saving: cut the seats. Rising acceptance is a reasonable leading indicator. They stop being evidence at the point where someone reads them as return. The value column has to come from your own systems in Step 2, compared against the baseline from Step 1.
Watch the PR merge count in particular. More merged PRs per engaged developer can mean more value, or it can mean the same work split into smaller pieces, or more changes that come back as rework. Read it next to reverts and the 30-day rewrite share, never alone.
Step 4: Plan for usage-based billing
Seat pricing made the AI coding budget easy to forecast and hard to judge. Usage-based billing makes it harder to forecast and easier to judge. GitHub began moving Copilot plans to usage-based billing on June 1, 2026 (annual Pro and Pro+ plans stay on premium-request pricing, with higher model multipliers, until the annual term ends). Its announcement says "Credits will be consumed based on token usage, including input, output, and cached tokens," priced at each model's published API rate, while "Code completions and Next Edit suggestions remain included in all plans and do not consume AI Credits."
A separate changelog entry says Copilot code review on private repositories also consumes GitHub Actions minutes for each review. The economics of that change are covered in GitHub Copilot usage-based billing and review economics.
For a renewal, metered billing changes four things:
- Headcount no longer predicts cost. Two teams with the same seats can spend very different amounts, depending on agent use and model choice. Budget per team, from metered history, not per seat.
- Abandoned work reaches the bill. Under a flat seat price, an abandoned agent session added nothing to the bill. Under metering it does, so cost per merged change from Step 2 reflects wasted sessions as well as shipped ones.
- Review has a price of its own. When automated review consumes CI minutes, the cost of reviewing AI-generated changes shows up in two budgets. Add the CI line to the cost of the tool, or the tool looks cheaper than it is.
- The negotiation moves. GitHub's billing post says admins can set budgets "at the enterprise, cost center, and user levels." Rates, pooled allowances, overage caps and the budget controls you get count for as much as the headline plan price.
Before you sign a metered contract, run one quarter of actual usage through the new price list. If your history cannot support that calculation because usage data was never exported, fix the export first.
The evidence no vendor dashboard can supply
One row in the Step 2 table, architectural corrections in review, needs more than a sample of comments. It covers senior engineers catching changes that compile and pass tests but break a decision the team already made. No vendor dashboard reports this, because the vendor does not hold your decisions.
Count it yourself at baseline and again before renewal: how many of the architectural decisions you rely on are checked automatically before review, and how many depend on a reviewer remembering them. Mneme, an open-source decision and control layer for agentic software development, sorts documented decisions into four tiers for this kind of count (Protected, Mneme-ready, Requires modelling and Guidance), described in its protection activation documentation. That count belongs to engineering decision governance, and it is the one line of the evidence plan that no AI coding vendor will fill in for you.
What to bring to the renewal meeting
Bring the baseline, the monthly series from Step 2, and a one-page split of the vendor's metrics into use and value. With those, the six questions in the AI engineering ROI guide have answers, and the renewal becomes a decision about which teams, which plan level and which price structure, rather than a yes or no on the whole budget.