Production recovery

Stuck with AI another vendor could not ship?

About a third of our AI engagements start exactly there.

Recover the build →
Case studies Book a 30-minute discovery call

AI project recovery: assess, retain and repair with evidence

Recover an AI project by reproducing the intended task and its failures, checking source and operating boundaries, then comparing what to retain, repair or replace. Preserve the evidence behind each component decision. Estimate bounded work, verify transfer and involve the receiving owner before expanding exposure or declaring the recovery complete.

Kanika Mathur
By Kanika Mathur, Service Delivery Lead
Reviewed by Arjun SareenPublished Sep 16, 2026Updated Nov 16, 202614 min read
Recovery
A laptop showing an out-of-focus terminal with red highlighted lines, a printed architecture diagram with red circled boxes, blank sticky notes on a dark wall, a glass of coffee and a desk lamp, no people
Key takeaways

The short version

  • Reproduce the blocker. A symptom does not establish its source or solution.
  • Decide by component. Retain useful work and compare bounded alternatives.
  • Verify the transfer. Access, rights and operating capability need separate evidence.
  • Estimate actual work. Separate assessment, repair, replacement and recurring costs.
  • Keep expansion conditional. A calendar or cheaper call path cannot settle acceptance.

Recognize a stalled project without guessing the cause

AI project recovery starts by identifying the intended task, the evidence already available and the decision that remains blocked. A favorable demonstration can coexist with weak access, evaluation, integration or operation. Investigate those gaps before adding scope or replacing a model, and contain an active security or operating incident through the appropriate response process.

Consider a SaaS support assistant that drafts internal replies from product documentation and a permitted ticket. The stalled build produces plausible text, but reviewers cannot consistently locate its evidence and its internal draft store sometimes records a draft twice. Recovery should first establish the supported product/source boundary and reproduce those failures. The proposed release remains draft-only; it does not autonomously send replies, close tickets or change accounts.

Assess
Reproduce the task, failure and operating boundary.
Recovery review
Retain
Preserve components supported by usable evidence.
Component decision
Measure
Compare accepted work, complete cost and recovery.
Verification
  1. Untested exposure: the demonstration has not covered the intended users, source variation, concurrent work or dependency failures.
  2. Unclear acceptance: people disagree about useful, complete or unacceptable behavior.
  3. Weak evaluation: important failures cannot be reproduced against a versioned candidate and clear judgment.
  4. Incomplete economics: the team cannot connect requests, calls, retries and retained review with the cost of an accepted task.
  5. Moving commitments: scope, dependencies or dates change without an explanation of the evidence and decision.
  6. Fragile transfer: usable access, rights or operating knowledge depend on an unexplained supplier arrangement.
  7. Unusable detection: responders lack the evidence or authority needed to investigate a live failure.

Deloitte's technology-value survey discusses expertise and other transformation barriers.4 Such association is context for checking capability, not a diagnosis that every stalled AI build has a skills problem. The AI project failure guide examines underlying causes; a recovery review needs the evidence from this system.

Review ten areas and explain what each gap means

Assess the task, evaluation, sources, implementation, cost, operation, authority, accountability, release and transfer. Use interviews to understand the work and artifacts to verify important claims. Record the consequence, uncertainty, owner and next action for each finding rather than treating a color score as an objective recovery verdict.

Begin with authorized access and a clear investigation boundary. Request the relevant repository, configuration, provider and hosting information, existing evaluations, invoices and suitably limited logs. If access is unavailable, distinguish a contractual restriction, missing permission and a technical recovery problem. Some useful assessment may continue, while a blocked fact remains explicitly unresolved.

Ten review areas; findings need consequence, evidence and an owner
Audit pointEvidence to ask forQuestion raised by a gap
1. Intended taskAcceptance, baseline and responsible decisionNo agreement on useful behavior
2. EvaluationCases, versions, judgments and failuresOnly a favorable demonstration
3. SourcesPermission, lineage, freshness and coverageUnclear rights or missing evidence
4. ImplementationBuild, dependencies, prompts and configurationCannot reproduce the relevant path
5. CostCalls, retries, infrastructure and retained workUnknown unit or incomplete accounting
6. OperationRequest time, failures and load evidenceOnly a narrow happy-path run
7. AuthorityRoles, tools, limits and denied casesConsequential access without justification
8. AccountabilityRisk decisions and incident responsibilityNo one can accept or reduce exposure
9. Release/recoveryVersioned artifacts, observation and restoreUntested partial failure or reconciliation
10. TransferRights, accounts, exports and usable knowledgeDependence cannot be explained or transferred
A laptop showing a blurred file tree of folder icons beside a monitor with a blurred dashboard of coloured panels and a line graph, a spiral notebook and pen on a printed checklist with empty tick boxes, under a desk lamp, no people
Inspect the implementation and its operating evidence alongside the people who can explain the task and accept the next decision.

Connect the review areas to reproducible decisions

For the support assistant, retain examples of an unsupported answer, an unavailable source and a duplicate destination write. Reproduce the request path with its relevant versions, then separate source selection, generation, review and integration outcomes. A symptom visible in the final response may originate earlier in the workflow.

  • Task and evaluation: define the accepted result and meaningful failure cases. Anthropic describes agent evaluations using outputs, environment outcomes, trials and explicit graders.6 Choose coverage from the task rather than importing a fixed number of cases.
  • Data and lifecycle: Google's MLOps guidance connects offline development with deployment and operation.5 Inspect the obligations relevant to this application without requiring a training pipeline it does not own.
  • Threat and authority: OWASP identifies disclosure, injection, excessive agency and unbounded consumption among LLM risks.7 Test actual source and tool paths, including denied and repeated operations.
  • Accountability: NIST AI RMF treats risk tolerance as contextual.8 Determine who can accept the proposed exposure, limit it and respond when intended behavior is not met.
  • Transfer: identify the access, rights, formats and knowledge the receiving owner needs. Account names alone cannot establish legal ownership or operational independence.

Do not let a summary rating hide a consequential gap. An active disclosure or unauthorized action may require containment before an ordinary repair plan. A missing performance observation may instead justify a bounded test. Escalate according to consequence and authority, with appropriate security, legal or business involvement where needed.

The evaluation guide develops task cases and judgments, while governance cases illustrate distinct authority and response questions. Use them to sharpen the local review without assuming a framework reference settles the findings.

Choose what to retain, repair or replace by component

Retain components supported by the required evidence, repair bounded deficiencies and compare replacement where the current approach cannot reasonably meet the task or operating conditions. A mixed outcome is often plausible. Missing information calls for investigation; it does not automatically justify a rewrite of the whole project.

Component-level hypotheses to verify before choosing a recovery path
ComponentKeep whenFix whenRebuild whenFirst test
Task and valueUseful task and acceptance are supportedClarify a specific criterion or baselineReconsider the task if value is unsupportedObserve the actual user workflow
DataPermitted source coverage meets the taskRepair extraction, access or freshnessTest an alternative when needed sources cannot be usedTrace a representative source record
Model/promptRelevant cases meet acceptanceTarget a demonstrated behavior or context gapCompare another candidate if evidence warrantsRun equivalent cases and preserve failures
ServingRequired behavior fits operating constraintsRepair a measured bottleneck or failurePrototype an incompatible requirement elsewhereExercise load and dependency failures
EvaluationCases and judgments support decisionsImprove coverage or reproducibilityBuild missing evaluation capabilityReproduce an important failed case
AuthorityPermissions match intended operationsReduce unjustified access and add checksRedesign a fundamentally unsafe action pathTest allowed, denied and repeated actions
ImplementationReceiving owner can build and maintainDocument and repair bounded opaque areasCompare replacement with migration obligationsBuild and investigate from clean access

For the support assistant, the existing document importer may be usable while source citation needs repair and destination-write handling needs redesign. Verify each hypothesis. Test the current model with corrected evidence before concluding it is the constraint, and compare alternatives on the same accepted task rather than a favorable new demonstration.

Keep the product question open. If staff do not need generated replies but do need better document lookup, a narrower feature may be the useful recovery. Unusable source rights or unsupported business value can require rescoping or stopping, not simply more infrastructure.

  • Compare future paths: retain useful evidence and account for migration, transition and operation. Sunk effort cannot establish which option now meets the need.
  • Change one hypothesis clearly: distinguish prompt, retrieval, model and integration changes where feasible, recording concurrent changes that affect interpretation.
  • Replace within a boundary: a new component still needs acceptance, source permission and recovery evidence; it is not exempt because the old one failed.

The PoC-to-production guide expands those connected release responsibilities. Preserve the old system's useful task knowledge without copying its unsupported commitments into the replacement scope.

Use a review calendar with conditional checkpoints

A 30-60-90 day example can organize assessment, selected repairs and a supported release or handover. It is a planning calendar, not an expected recovery duration. Make the checkpoints capable of extending, narrowing or stopping work when access, feasibility, operating risk or value remains unresolved.

In this example, the first period establishes the draft-only task, source boundary and reproducible citation/duplicate-write failures. It also handles any active exposure and creates a baseline. Later work targets justified repairs and verifies them before the receiving team accepts a limited operating scope. Discovery may change the calendar.

Illustrative review calendar; extend, narrow or stop when the evidence requires
PhaseGoalMain workDeliverableExit gate
Days 1 to 30Understand and containAccess, task baseline, gaps and urgent responseEvidence map and bounded optionsContinue, investigate, rescope or stop
Days 31 to 60Verify selected repairsTargeted implementation, evaluation and recoveryResults against the original questionAccept supported scope or revise
Days 61 to 90Observe and transferLimited release where justified; paired operationOperating evidence and receiving-owner recordOwner accepts the tested boundary

Set meaningful checkpoint artifacts: accepted task criteria, a reproduced failure, a tested repair and an operator completing recovery. A deadline alone should not convert missing evidence into approval. Review the current commitment with the people who can accept its business and operating consequences.

  • Establish the comparison: preserve actual baseline cases, complete retained effort and the versions tested. A recovery may improve one task while leaving another unsupported.
  • Limit exposure: use a staged path appropriate to the consequence, with detection and a capable response. Do not assume every data or external effect can be undone automatically.
  • Transfer while learning: involve the receiving owner in the investigation and verification so handover does not begin only after the implementation team leaves.

The AI release checklist helps preserve evidence for the next exposure decision. Reassess when new data, users or actions enter scope; the original pilot cannot stand in for those tests.

Transfer usable access, rights and operating knowledge

A workable handover lets the receiving team build, operate, change and investigate the supported system under the applicable rights and contracts. Inventory code, configuration, data, derived artifacts, accounts, licenses, evaluations and recovery knowledge. Verify the transfer through real tasks rather than assuming every asset must or can be assigned without conditions.

OMB M-25-22 addresses portability, documentation and vendor-lock-in protections in US federal AI acquisition.10 It offers useful procurement context, but its federal scope is not a law governing every private takeover or a guarantee that all third-party models and code transfer. Review the actual agreement and licenses with the responsible people.

  • Implementation: identify repositories, build/deployment configuration, prompts, dependencies and release artifacts needed for the supported path.
  • Accounts: verify administrators, collaborator access, credential handling and the sequence for transfer or replacement. Preserve evidence needed for incident investigation.
  • Data: agree usable formats, source rights, deletion obligations and treatment of derived indexes or outputs. Test an export and import where transfer is required.
  • Models and components: identify provider terms, usable weights where applicable and licenses. NIST's Generative AI Profile addresses distinctive risks and management actions involving such dependencies.9
  • Evaluation: transfer permitted cases, judgments, configuration history and known limits so the next team can reproduce important findings.
  • Operation: provide runbooks and practice build, release, failure investigation and recovery with the receiving owner.

Plan a factual transition with the outgoing supplier

Use the existing contract, access obligations and relevant business/legal advice to arrange the transition. Agree the evidence, people, dates and commercial terms needed for knowledge transfer. Keep requests specific: a missing configuration, an unusable export or an unreproduced failure is easier to resolve than a general accusation that the system does not work.

Coordinate credential changes and data movement so the transition does not destroy needed access or create a new exposure. Record unresolved rights or dependencies and their effect on the plan. A cooperative handover can still need independent verification of the receiving team's capability.

Select recovery help by the problem it can verify

Ask a prospective team to explain its assessment method, the lead's role, relevant permitted evidence and how it will distinguish repair from replacement. Confirm what access it needs, what the first bounded finding will be and who operates the result. A supplier category or claimed project-success percentage cannot establish fit for this inherited system.

Assess your own retained capacity as well. Domain decisions, source permission and acceptance may remain with the buyer even when a partner implements the repair. Use the development partner evaluation guide to compare responsibilities and evidence rather than assuming a change of vendor resolves the underlying constraint.

Estimate recovery from the findings and the supported task

Price assessment and recovery from the actual system boundary, access, uncertainties and work required to verify the result. Separate alternative repair and replacement scopes from optional integration and ongoing operation. Avoid mapping an inherited project to a generic price range before knowing what can be retained and what needs evidence.

For the draft-only support assistant, assume one application, one product documentation set, two user roles and one model-provider path. The assessment budget is 40 hours: 8 for access/inventory, 12 for reproducing task failures, 12 for source and integration inspection and 8 for bounded options. The hourly rate below is a chosen USD 100 for transparent arithmetic, not a supplier quote or market benchmark.

Illustrative work packages for a narrowly scoped support-draft recovery
Work packageAssumed effort and costBoundary or dependency
Assessment40 h × USD 100 = USD 4,000One app, source set and model-provider path
Bounded repair160 h × USD 100 = USD 16,000Targeted access, evaluation and release work
Alternative replacement368 h × USD 100 = USD 36,800Same limited drafting task; migration included
Optional integration validation56 h × USD 100 = USD 5,600New external draft attachment; internal-store repair separate
Ongoing support allowance12 h/month × USD 100 = USD 1,200/monthChosen capacity; usage and availability terms separate

The repair option's 160 hours budgets 40 for source/citation work, 40 for access and duplicate handling, 50 for regression/recovery and 30 for release and transfer. For the replacement, reuse the assessment’s task criteria, permission map and reproduced failures; its separate 40 hours are not charged again inside this option. Allocate 48 hours to a source importer, 64 to retrieval/citation logic, 48 to the draft/review interface, 32 to role enforcement, 72 to evaluation and correction, 40 to data transition, 40 to release/recovery and 24 to receiving-team transfer: 368 additional hours. The transition assumes at most 100 current documents and 500 internal drafts, stable record identifiers and two role mappings. Include a trial import, record/source reconciliation and a rollback rehearsal; historical ticket migration is excluded.

Optional integration adds one new ticket-system draft attachment path, beyond the existing internal draft store. Allocate 12 hours to the destination contract and authentication, 16 to its adapter, 12 to state/idempotency handling, 8 to denied/timeout/retry tests, 4 to buyer acceptance and 4 to operator handover: 56 hours. Repair already fixes duplicate handling in the internal store; this optional line validates a different external destination and its confirmation/read-back path. It cannot send replies, close tickets or change accounts. Reuse the assessment’s source/role findings but inspect the new destination contract separately. These are chosen task estimates that actual interfaces and failures may change. The monthly allowance reserves chosen support capacity; it does not promise continuous coverage or include every incident. Do not add repair and replacement as if both are required. The separately priced assessment informs either option; its accepted findings are reused, with new implementation and verification allocated above.

Hosting, provider usage, tools, specialist independent review, tax and unpriced internal participation are excluded. Estimate elapsed time separately from hours, using available productive capacity and external dependencies. The AI engagement pricing guide develops scope and contract questions without supplying a transferable quote for this recovery.

Investigate consumption without weakening the accepted task

Trace the actual call path before selecting a cost lever. One request can trigger extraction, retrieval, generation, validation and retries. The illustration holds requests at 100 and varies chosen calls per request; real consumption also depends on model, input/output lengths, cache behavior and unfinished work.

Call-count sensitivity before comparing serving cost
Assume 100 requests, with one, two, four or five model calls per request. Counts show a chosen workflow scenario; they do not imply equal token use, price, quality or task success.
Illustrative request-to-call arithmetic Each path is an assumed call count per request multiplied by 100 requests. Real cost also depends on model, tokens, cache behavior, retries and retained work. Fewer calls do not establish adequate quality. 0200400 Four-call path Five-call path Two-call path One-call path 400500200100
Chosen request and call inputs, not provider prices
PathAssumptionResult
Four-call path100 requests × 4400 calls
Five-call path100 requests × 5500 calls
Two-call path100 requests × 2200 calls
One-call path100 requests × 1100 calls
Illustrative sensitivity only. Measure actual call paths and accepted outcomes before changing the system; token volumes and cost per call need their own inputs.
  • Caching: Anthropic distinguishes write lifetimes and model-specific read pricing.11 Check eligibility, expiry, hit rate and the actual model's rates. Do not assume every reused-looking token is billed at one universal discount.
  • Asynchronous work: OpenAI Batch supports eligible work at 50% lower costs with a stated 24-hour turnaround.12 Confirm the deadline and failure/retry handling fit; that pricing difference is not a 50% reduction in total application spend.
  • Routing and model choice: compare candidates on the same cases and complete costs, including retained review. Fewer calls or a smaller model can change the result and need verification.
  • Limits: test usage controls and their effect when reached. OWASP identifies unbounded consumption as a risk,7 but a token cap alone cannot establish a whole-system budget ceiling or safe continued service.

Keep useful quality, authority and operating evidence beside the cost comparison. The AI cost optimization guide examines those tradeoffs. A favorable provider rate is only one input to the cost of an accepted, supported task.

A recovery decision needs a reproducible task, an accepting owner and evidence that the proposed change addresses the observed failure.

Planning recommendation

Resourcifi's AI integration team can discuss an inherited assistant using its task evidence, access constraints and known failures. Bring those materials, current usage and the proposed receiving owner to an AI project recovery discussion so the next assessment or repair has a concrete boundary.

Frequently asked

AI project recovery questions

Where should AI project recovery begin?
Define the intended task, supported users and authority, then reproduce important failures with relevant versions and evidence. Check access, sources, evaluation, integration, operation and transfer constraints. Contain an active incident through the appropriate response process. A structured assessment should identify the next bounded investigation or repair without assuming a new model or vendor will resolve the actual cause.
How do you decide whether to repair or rebuild?
Compare components against the supported task and operating requirements. Retain evidence-backed work, repair a bounded deficiency and test replacement where the current approach cannot reasonably meet the need. Missing information may call for investigation rather than a rewrite. Include migration, compatibility, verification and transfer in the comparison, and keep stopping or narrowing the product open when value or usable sources remain unsupported.
Does a 30-60-90 day plan promise recovery in ninety days?
Use it as an illustrative assessment, repair and review calendar. Access, feasibility, source rights, verification and receiving-team capacity can change the timing or scope. Each checkpoint should allow investigation, revision, narrowing or stopping. Estimate effort separately from elapsed dependencies and do not treat a calendar date or unrelated company delivery metric as evidence that this recovery is complete.
What is needed for a usable vendor handover?
Inventory the code, configuration, accounts, permitted data, derived artifacts, model/component terms, evaluation history and operating knowledge needed for the supported system. Confirm applicable rights and transfer obligations rather than assuming all third-party assets can be assigned. Exercise build, release, investigation and recovery with the receiving owner. Account administration and legal ownership answer related but different questions.
How should recovery cost be estimated?
Use findings to define separate assessment, alternative repair or replacement, optional integration and recurring capacity. The example allocates 40 assessment hours, 160 repair hours or 368 additional replacement hours, reusing accepted assessment findings. Optional 56-hour work adds a new external ticket-draft attachment path, distinct from repair of duplicate handling in the internal store. Hypothetical USD 100/hour produces USD 4,000, USD 16,000, USD 36,800 and USD 5,600 respectively. Compare alternatives, exclusions, buyer participation and actual productive capacity separately from elapsed time.
Can caching or batch processing prove the recovery will save money?
Measure actual eligibility, model rates, cache writes/reads, expiry and hit behavior, or asynchronous deadlines and retry handling. OpenAI Batch’s eligible-work pricing difference does not imply the same reduction in total spend. Compare accepted task quality, complete consumption and retained support work. A cheaper token or fewer calls can alter behavior and needs evaluation before it supports a cost decision.
Kanika Mathur

Kanika Mathur

Author · Service Delivery Lead, Resourcifi

Kanika Mathur is Service Delivery Lead at Resourcifi. She leads the engineering pods that scope, build and run client software, from mobile apps to AI systems, and she reviews the process and figures in our engineering guides for accuracy.

Resourcifi on LinkedIn →
Arjun Sareen

Arjun Sareen

Reviewer · Head of Business Development, Resourcifi

Arjun Sareen is Head of Business Development at Resourcifi. He scopes builds with founders and product leaders, matches each project with the right team model, and hears the same buying questions every week, which is what he writes about.

Resourcifi on LinkedIn →

Like this? Get more from Resourcifi in Google Search.

Sources

  1. deloitte.com, AI and tech investment ROI | Deloitte Insights. Transformation barriers and expertise concerns in a survey.
  2. docs.cloud.google.com, MLOps: Continuous delivery and automation pipelines in machine learning | Cloud Architecture Center | Google Cloud Documentation. ML lifecycle automation and production-operation boundaries.
  3. anthropic.com, Demystifying evals for AI agents. Agent evaluation trials, graders and environment outcomes.
  4. genai.owasp.org, LLMRisks Archive - OWASP Gen AI Security Project. LLM application threats and consumption risk.
  5. nvlpubs.nist.gov, NIST.AI.100-1.pdf. Contextual AI risk tolerance and accountability.
  6. nvlpubs.nist.gov, NIST.AI.600-1.pdf. Generative-AI risks and management actions.
  7. whitehouse.gov, M-25-22-Driving-Efficient-Acquisition-of-Artificial-Intelligence-in-Government.pdf. Federal AI acquisition portability and documentation context.
  8. platform.claude.com, Prompt caching - Claude Platform Docs. Model-specific cache reads, writes and lifetimes.
  9. developers.openai.com, Batch API | OpenAI API. Eligible asynchronous batch pricing and turnaround.
Keep reading
Related guides worth your time
All AI insights →
Agentic design patterns: choose the right control flow Strategy, architecture & ops Agentic design patterns: choose the right control flow Compare agentic design patterns by task, evidence, tools and control flow. Test retrieval, routing, agents, cost and reco... Read guide → AI cost optimization: measure tasks, compare changes and preserve quality Strategy, architecture & ops AI cost optimization: measure tasks, compare changes and preserve quality Plan AI cost optimization from task-level usage, caching, routing and retrieval, with explicit cost assumptions and tests... Read guide → AI feature validation: test value before a larger build Strategy, architecture & ops AI feature validation: test value before a larger build Use AI feature validation to compare the task baseline, implementation choices and failure cases. Record correction, pilo... Read guide → AI governance failures: cases, controls and assessment Strategy, architecture & ops AI governance failures: cases, controls and assessment Examine AI governance failures through reported cases and evidence checks for ownership, testing, agent permissions, vend... Read guide → AI in production: targets, evidence and operating decisions Strategy, architecture & ops AI in production: targets, evidence and operating decisions Plan AI in production through operating targets, evaluation, bounded release and recovery. Compare build and buy and asse... Read guide → AI model deployment challenges: from PoC to production Strategy, architecture & ops AI model deployment challenges: from PoC to production Resolve AI model deployment challenges through task evaluation, permitted sources, controlled integration and recovery. M... Read guide → Agentic RAG: step selection, retrieval patterns and evaluation Agents & RAG Agentic RAG: step selection, retrieval patterns and evaluation Understand agentic RAG through routing, corrective retrieval and evaluation, with practical tests for step selection, sou... Read guide → AI agent development cost: build effort, usage and operating work Agents & RAG AI agent development cost: build effort, usage and operating work Estimate AI agent development cost from scoped build tasks, model calls, tools and upkeep, with a worked support budget a... Read guide → AI agents for customer service: define actions and verify resolution Agents & RAG AI agents for customer service: define actions and verify resolution Build AI agents for customer service with explicit identity, knowledge, action and handoff checks. Measure verified resol... Read guide →
Bring task failures, access constraints and current usage

What can the next team retain, repair and verify?