Ironclad Rethinks AI Token Spend ROI

Ironclad will not chase more tokens and is measuring AI token spend against a new proxy it calls trusted throughput, according to AI Engineer.

Mingsheng Hong, VP of Engineering for AI at Ironclad, laid out the framework at AI Engineer after the contract lifecycle company pushed past broad adoption over the last two quarters.

Ironclad builds AI for legal contracting, where lawyers test output on known contracts before they trust it on redlines and anomalies. That same trust ladder now shapes how it judges engineering output.

Why a dashboard is not a leaderboard

Hong cited a voluntary dashboard at Amazon (NASDAQ:AMZN) where engineers competed to top a token usage board, a pattern he said also appeared at Meta Platforms (NASDAQ:META), plus a sensational claim of a company burning $500 million on cloud in a month.

His fix is to keep per-team and per-person cost tracking but to treat it as a smoke detector. Low-use pockets signal an adoption gap to investigate, bursts trigger a check on legitimacy, and teams are compared only in context because platform and UI teams extract value differently.

What trusted throughput actually measures

Ironclad evolved from lines of code to open pull requests to merged pull requests, then added a complexity weight using a prompt that asks one or two models to score each merged PR by t-shirt size.

The qualitative layer has three buckets. Objective gates include test coverage, security checks and canarying, subjective review covers clarity, maintainability and architecture fit, and customer validation looks at rollbacks, fires and tickets for bugs or friction.

Where the bottleneck moved

Generation is now abundant, so pressure has shifted to review and continuous integration. Hong warned against the workaround of shipping large PRs to avoid hour-long CI runs, which thins reviewer attention and hurts quality.

Ironclad puts AI review first as a filter for style and missing tests before human reviewers apply judgment on design and security. For CI, it tracks time from ready-to-submit to submitted and retry counts for flaky tests, and staffs platform work to fix the infrastructure instead of having engineers or agents babysit merges.

How Ironclad controls cost without austerity

The goal is not to minimize spend but to raise return per token. The pragmatic framework sets budgets, quotas and anomaly alerts alongside regular human review, with a learning loop that feeds findings back into institutional best practices.

Team-level practices include capping agentic loops that auto-fix tests, structuring prompts with stable system prefixes first to benefit from prompt caching, and pruning or compacting context in long chats, including with tools like Claude Code and Codex that now do it automatically. StartupHub.ai data shows ROI raised $10M in a Series A in 2026, a useful anchor as private legal-tech peers try to prove that token efficiency scales to real throughput.

Build versus buy is explicit. Ironclad buys non-differentiating IDE and CI infrastructure and builds its internal playbook of prompts for bug fixes, UI features and refactors, while still evaluating whether to wrap Claude Code in a cloud builder agent or buy a vendor alternative.

Token budgets are now an engineering management problem, not a vendor billing detail. Ironclad is betting the teams that instrument review and CI will ship more trusted code per dollar than those that just cap prompts.