Free AI Coding Is Not Free: Route It by Blast Radius With a 6-Field Ledger

A free model endpoint can push the invoice line close to $0 and still make delivery slower. The constraint is rarely token price; it is the review queue, the rerun rate, the data-handling work, and the chance that a plausible change reaches a high-blast-radius path before anyone prices the downside.

Use this when a team says: “The tool is free, so why not let everyone try it?” Treat free capacity as a routing problem, not a seat-count problem.

Disclosure: This article was prepared as part of MonkeyCode's product outreach. MonkeyCode's free model access and free server option can be a low-friction place to rehearse the workflow below; I am not assuming quotas, permanence, hardware, model names, or benchmark results. Validate current limits before depending on them.

The unit that matters: accepted work, not generated work

Define one pilot unit as an accepted task: a change or analysis that a named owner would merge, ship, or act on after review.

For each candidate task, record six fields before routing it to AI assistance:

Field	Definition	Example
S	Minutes saved if the first answer is useful	45
V	Verification minutes: tests, review, security checks	20
r	Probability the output causes rework or rollback	0.25
R	Minutes to recover when r happens	70
H	Data-handling minutes: redaction, access setup, audit notes	10
B	Blast radius: 1 local docs, 5 production/security path	4

Decision rule:

NetMinutes = S - V - H - (r * R)
Allow AI route only if:
  hard gates pass
  and NetMinutes > MinimumUsefulMinutes
  and B <= MaxBlastRadiusForThisPilot

Worked example: S=45, V=20, H=10, r=0.25, R=70 gives 45 - 20 - 10 - 17.5 = -2.5. The task looked “free,” but expected value is negative before morale cost. If verification drops to V=12 and recovery to R=45, NetMinutes becomes 45 - 12 - 10 - 11.25 = 11.75; now it may clear a 10-minute threshold.

Break-even for rework probability:

r* = (S - V - H - MinimumUsefulMinutes) / R

With the first example and a 10-minute threshold, r* = (45 - 20 - 10 - 10) / 70 = 0.071. If more than ~7% of these tasks need recovery, the route fails even when the model costs nothing.

A 3-route pilot map

Route by blast radius and reversibility, not by enthusiasm.

Route	Task shape	AI policy	Hard gate	Exit if
Green	Reversible, local, low data sensitivity: test naming, draft migration notes, refactor suggestions behind CI	AI-first allowed	No secrets; branch only; CI passes	r > r* for 2 weekly samples
Amber	Shared modules, internal tooling, moderate customer impact	AI proposes, human disposes	Owner reviews diff; rollback path named; no external actions	Review wait exceeds saved minutes for 5 tasks
Red	Production incident path, auth/payments, regulated data, irreversible ops	No autonomous use	Manual only; AI may summarize post-incident after approval	Any boundary failure, leak, or unowned action

This ledger is a conversation tool, not objective truth. Its job is to force the argument into variables a team can measure.

Practical workflow

1. Pick 20 real tasks from the last sprint; do not use demo tasks.
2. Score S, V, H, r, R, and B with the engineer and reviewer who would actually touch the work.
3. Run the same prompt pattern across at least two environments if procurement allows, one of which can be a free-access or free-server setup to reduce invoice friction.
4. Record accepted work, rework events, review latency, and any boundary violation in one shared sheet.
5. Set an expiry date, owner, and archive rule before the pilot starts. Example: owner is the platform lead; expiry is 30 days; raw prompts are deleted after 14 unless legal requires retention.
6. Decide with thresholds: expand Green only, renegotiate Amber, stop Red.

Hard gates that should not be waived: no production credentials, no unexplained network/tool actions, no unreviewed generated migrations, no personal or regulated data in prompts, no “temporary” access without an expiry owner.

Sensitivity questions that reverse the decision

Ask these before expanding:

• If rework probability doubles, which routes still clear the threshold?
• If review capacity drops by one person, does Amber become negative?
• If the free tier changes terms tomorrow, what is the migration cost for prompts, logs, and evaluation data?
• If generated output becomes 20% better but verification stays flat, which route improves most?
• Which variable would make you stop: boundary failure, review queue, data-handling time, or rollback minutes?

Who should not use this approach

Do not use a free-access rehearsal as your decision basis if you cannot isolate secrets, cannot measure rework, lack CI or rollback, have regulated prompts that require vendor review first, or need guaranteed capacity and support. Also avoid it for high-severity incident response until your boundary controls are already boring.

Free capacity is useful for learning your own routing table. It is not evidence that a route belongs in the delivery path. If you want one low-friction place to rehearse the ledger before a paid commitment, MonkeyCode's free model access and free server option are worth checking against the gates above—then measure your own tasks, not the demo.