The method
Most agent projects fail on ownership, on the definition of correct, and on the people whose work changes. Not on the model. This page is the whole method, so you can run it yourself or hold us to it.
Shadow mode · week 4
Team and agent, side by sideTest set closed · Corrections join it weekly
The model is rarely the problem. Rollouts stall on four things, and each one has a rule.
Failure 1: No business owner
One person who runs the workflow today owns the agent after launch, with the authority to change the policy it follows.
Failure 2: No definition of success
Before any build, two of your people grade the same 50 closed files blind. Where they differ is the human baseline and the agent’s miss budget.
Failure 3: No guardrails
Code keeps eligibility, limits, authority and the audit trail. The model reads the evidence, weighs conflicts, applies the written policy and writes down its reasoning.
Failure 4: Change management as an afterthought
The people whose work changes grade the test set, watch shadow mode and correct the agent. Their corrections become the test set.
Week 1
One workflow your team already runs, with its documents, systems and written rules, one owner on your side, and one number that decides whether it worked.
Weeks 1 to 2
Fifty closed files with known outcomes, graded blind by two of your people. We write the agent without looking at them, run it, and report false positives and false negatives against the human baseline.
Weeks 3 to 8
About 100 new files run beside your team with no answers given. Your team compares weekly. The test set stays closed while we tune, so the agent is never fitted to it.
Weeks 9 to 10
The agent finishes the routine cases alone. Every batch is stamped with the policy version and can be undone. Exceptions go to a person with the facts gathered.
After go-live
Every finished case is checked for eight weeks, then a sample. Every correction joins the test set. The error rate is reported to the owner weekly.
Two of your people grade the same fifty files blind. Where they agree is the answer key. Where they differ is the human baseline, and the agent is allowed to miss that much and no more. Only then does the agent run, once, against the key.
50-file test · blind grading
Two graders, then the agentWhere graders differ is the miss budget
01 · Your systems
Read only. Nothing here is changed directly.
02 · Agent runtime
Policy in code
Eligibility, limits, authority and the audit trail. Deterministic, versioned, tested.
Model
Reads the evidence, weighs conflicts, applies the written policy, writes down its reasoning.
Test set
Your 50 graded files plus every correction since. Run before each policy change.
Runs in your tenant or ours. ISO/IEC 27001 controls. No training on your data.
03 · What comes back
Batch write-back
Validated rows, stamped with the policy version, one-click undo.
Exception to a person
The facts gathered, the reasoning attached, the decision left to your team.
Error watch
Every row checked for eight weeks, then a sample. Weekly report to the owner.
The four rules and this method were presented to technology executives at an OpenAI event in San Francisco.
October 22, 2026
OpenAI, San Francisco
Tarun Upaday, Routespring’s CEO, presented at an OpenAI event for technology leaders in San Francisco. The session covered why agent rollouts stall in operations: no business owner, no definition of correct, no guardrails, and change management left to the end. The four rules on this page are the answer.
Read the four rulesFifty closed cases from the workflow you want to change, under NDA. Within two weeks you get the agent’s error rate against your own team and, for recovery workflows, the dollars it would have found. No charge.
