A production-grade bespoke AI system for a $5M–$50M business can ship in 90 days if the four stages — diagnose, architect, build, deploy — are sequenced correctly and the right checkpoints are enforced. It cannot ship in 90 days if any stage is skipped or overlapped. This piece is the week-by-week framework NURO uses on every custom-build engagement, with the specific checkpoints that separate a working system from an expensive prototype.

The 90-day framework, at a glance

Weeks 1–3: Diagnose. Operational reality map, ranked leverage assessment, sequenced build plan, architecture blueprint.
Weeks 4–6: Architect. Truth Boundary spec, integration map, human-in-the-loop topology, Decision Log schema.
Weeks 7–10: Build. Agent construction, integration wiring, structured output validation, staging deployment.
Weeks 11–12: Deploy. Production rollout with monitoring, team training, and Decision Log review cadence.

Weeks 1–3: Diagnose

The first three weeks are diagnostic — no code, no vendor selection. Four deliverables:

Operational reality map. Where is the business actually losing time, money, or leverage? On-site observation and stakeholder interviews across sales, ops, and customer service. No shortcuts.
Ranked AI leverage assessment. The specific jobs an AI could compress fastest, ranked by impact and effort. Ten to twenty candidates, narrowed to a top three.
Sequenced build plan. Which agent first, which second, which third — and what to skip. Sequence matters because each system reduces the volume (and increases the value) of what the next system handles.
Architecture blueprint. The high-level shape of the eventual system — which data flows where, which integrations exist, which components ship in production.

Checkpoint — end of week 3: the operator signs off on the top-three ranked leverage points and the sequence. If sign-off is unclear, extend Diagnose by one week rather than proceed. This is the single biggest failure prevention in the entire 90 days. See Why AI Projects Fail at $5M–$50M Businesses for the failure modes this stage prevents.

Weeks 4–6: Architect

Three weeks of architectural work — still no production code, but this is where the system either becomes possible or dies quietly. Four deliverables:

Truth Boundary specification. The specific data sources the agent may reference, the specific claims it's allowed to make, the specific out-of-bound triggers that escalate to a human. Written down. Reviewed. Signed off. See AI Hallucinations in B2B for the architectural detail.
Integration map. Every system the agent touches (CRM, dispatch, calendar, payment, ticketing, ERP) is documented with authentication mode, rate limits, error handling, and locking behavior.
Human-in-the-loop topology. Which decisions the agent makes alone, which route for approval, which escalate immediately. Committed to build spec, not left for post-launch.
Decision Log schema. What gets logged for every interaction, how it's indexed, how a human operator searches it, and how it's retained. Compliance requirements drive most of this.

Checkpoint — end of week 6: operator sign-off on all four documents. This is the last cheap-to-change point.

Weeks 7–10: Build

Four weeks of production engineering:

Week 7: Agent scaffolding, retrieval-augmented generation with hard grounding against the Truth Boundary sources, structured output validation.
Week 8: Integrations — CRM, calendar, telephony (if voice), any legacy systems named in the integration map. Every integration is tested end-to-end.
Week 9: Human-in-the-loop wiring — escalation paths, approval queues, override UI. Decision Log wired to a searchable index.
Week 10: Staging deployment with a synthetic-traffic pilot. Operator team runs real scenarios against the staging system. Failure modes surfaced and fixed.

Checkpoint — end of week 10: staging passes 30 real-scenario tests without a single hallucination, missed escalation, or integration failure. If not, extend by a week. No launch on a failed staging.

Weeks 11–12: Deploy

Week 11: Production rollout in stages. Feature flags or canary routing so the first 10% of real traffic hits the agent while the operator team watches. Decision Log review daily for the first week.
Week 12: Full rollout after the canary shows clean data. Team training on the operator UI, the Decision Log, and the override workflow. Documentation delivered. Weekly-review cadence established for the next 30 days.

Checkpoint — end of week 12: the four post-launch metrics land in the target ranges (answer rate ≥ 98% for voice; escalation rate 8–15%; correction rate trending down week over week; time-to-review under one business hour).

Post-launch: the 60-day drift cadence

The 90-day framework doesn't end at deployment. The first 60 days after launch determine whether the system produces revenue durably or drifts:

Day 1–14: Daily Decision Log review with the operator team. Two hours per day for the first two weeks. Every edge case gets a rule.
Day 15–30: Weekly Decision Log review. Truth Boundary refinements based on observed traffic. Confidence-threshold tuning.
Day 31–60: Bi-weekly Decision Log review. Drift-detection dashboard scan. Correction-rate trend analysis.
Day 60+: Monthly review. Quarterly full re-baseline against the Truth Boundary set.

See How to Deploy AI Responsibly and Profitably for the full seven-part governance framework these cadences implement.

The four failure modes this framework prevents

Building before diagnosing (weeks 1–3 prevent this).
Under-specified Truth Boundaries (weeks 4–6 prevent this).
Untested staging (week 10 checkpoint prevents this).
No post-launch owner (weeks 11–12 + 60-day cadence prevent this).

How to Deploy a Bespoke AI System: A 90-Day Framework.