Stabilize the decision before adding more work
A stalled implementation often attracts more features, more vendors, and more spending before the failure is understood. The first rescue step is to freeze nonessential expansion, preserve evidence, establish the actual system and process baseline, and identify which decision must be made: recover, reduce, rebid, replace, transfer, defer, or terminate.
Separate symptoms from root causes
| Symptom | Questions to test |
|---|---|
| Quality is inconsistent | Is the issue source data, ground truth, task definition, retrieval, model, prompt, interface, reviewer workflow, or changing population? |
| Users do not adopt it | Does the system solve the real task, fit the workflow, reduce or add work, preserve authority, and provide usable explanations or correction? |
| Cost is rising | Which consumption, retry, review, integration, support, security, or vendor minimum is driving unit cost? |
| Governance blocks release | Which evidence, owner, impact review, access control, evaluation, monitoring, fallback, or contract term is missing? |
| Delivery is late | Are scope, dependencies, environments, decisions, data access, acceptance, and vendor responsibilities explicit? |
| The pilot cannot scale | Is the architecture operable, portable, observable, supportable, and tested on representative conditions? |
Re-estimate from remaining evidence
Do not preserve the original business case merely because money has been spent. Recalculate the cost to complete, continuing operating cost, expected value range, major risks, replacement cost, exit cost, and the value of reusable assets. Treat sunk cost separately from the forward decision.
Classify assets honestly
Reusable now
Verified requirements, clean data contracts, working integrations, test environments, client-owned evaluation cases, and documented controls.
Reusable after remediation
Code or configuration with bounded defects, missing tests, weak observability, or incomplete ownership.
Stranded
Vendor-specific artifacts without export, unsupported prototypes, contaminated evaluation sets, unsafe permissions, and undocumented one-off logic.
Unknown
Assets whose ownership, version, security, data rights, or production behavior cannot be established.
A 30/60/90-day decision structure
- First 30 days: contain risk, freeze expansion, inventory the system, verify access and data, establish baseline and estimate to complete.
- By 60 days: test root-cause hypotheses, repair only decision-critical defects, run representative evaluation, and compare recovery with replacement or exit.
- By 90 days: execute the approved recover, rebid, replace, transfer, defer, or terminate plan with acceptance and rollback evidence.
Stop conditions
Terminate or replace when critical access or safety failures remain unresolved, the business outcome is no longer material, representative evaluation cannot be created, operating cost exceeds value, the supplier prevents necessary evidence or exit, or no accountable owner accepts the continuing service.
