Weeks 1–2: discovery, ownership, and access
A realistic schedule begins only after a workflow owner and system owners are available. During discovery, document the current process, establish baseline measures, identify approval points, inventory data, and begin security and vendor review. Delayed credentials and unclear ownership are common causes of apparent “AI” delays—not model performance.
The output is a signed scope with acceptance criteria, not a general list of ideas. It should name the first release, excluded edge cases, test data, integrations, reviewers, and launch decision. For insurance, include AMS and carrier-portal access owners. For A&E and construction, include project-system admins and who can approve writes to the official record.
Sourcy™ projects treat platforms as the central brain early: confirm where AI will read and write before building prompts. If the company still plans for humans to retype AI output into another system, discovery should flag that as a timeline risk and redesign the handoff.
Block calendar time for owners the same way you would for a payroll cutover. Discovery that runs in stolen half-hours produces incomplete maps and late surprises in week 5. Publish a RACI for access requests on day one: who asks, who grants, and the escalation path when a vendor portal ticket stalls.
Weeks 3–8: configure, connect, and test
A narrow workflow may be configured in several weeks; custom integrations and regulated data increase the range. Build the common path first, then exception handling, logs, permissions, and human approval. Use representative records instead of demonstration data that excludes duplicates, missing fields, unusual documents, and language variation.
Testing includes functional accuracy, security, role access, recovery from service failures, and user acceptance. Teams should know how to correct an output and where that correction is recorded. Include outage drills: what happens if the model API or a connector fails mid-transaction.
Keep a living defect log during configure-and-test. Separate content defects (wrong source text), mapping defects (wrong field), and control defects (missing approval). Mixing them into one “AI bugs” list slows triage and makes timeline slips look mysterious when they are operational.
- Week 3: approve design and data mappings.
- Weeks 4–6: build the end-to-end pilot and controls.
- Weeks 7–8: test edge cases and complete user acceptance.
Weeks 9–14: train, launch, and stabilize
Training should use each role’s real queue and include exceptions, not just the happy path. Launch to a controlled cohort, monitor daily at first, and keep a rollback or manual fallback available. Expand only when quality and operating measures meet the agreed threshold.
Projects run longer when they add systems, migrate historical data, require formal compliance review, or depend on third-party APIs. Publish these dependencies openly; compressing testing to preserve a date creates risk rather than speed. Stabilization is part of the timeline—budget at least two operating cycles of close watch after go-live.
Stabilization includes content and policy changes that arrive after go-live. Assign a weekly owner to review new exception themes and decide whether they are training gaps, rule gaps, or out-of-scope requests that belong in the next release backlog.
What stretches a timeline—and what does not
Model choice rarely dominates the calendar. Access delays, unclean master data, unclear approval authority, and changing scope do. Multi-state insurance rules, multi-office A&E standards, and multi-job contractor systems each add mapping work that must appear on the plan as tasks with owners—not as a vague “integration” line.
Parallel work helps when security review, content preparation, and connector builds can proceed independently. It hurts when three teams invent conflicting field definitions. Hold a single mapping source of truth and freeze it before week 4 unless a defect forces a change.
If leadership demands a shorter calendar, cut scope—not testing. Removing a second integration is safer than skipping edge-case review. Document the deferred scope so it does not reappear as an unspoken expectation during training.
Plan the second release while stabilizing the first
Do not start a second workflow while the first still fails its gate. Capture backlog ideas during stabilization, but fund the next release only after metrics clear. That discipline keeps the organization from mistaking activity for progress.
If you need a partner to run discovery through stabilization with clear gates, review Sourcy AI implementation services or request a quote with your target workflow and system list. Timelines stay honest when scope, access, and acceptance criteria are written before week one ends.
Hand finance and operations the same milestone list: discovery complete, pilot built, UAT signed, cohort live, gate cleared. Shared milestones prevent “the AI is late” narratives when the real blocker was credentials or an unowned approval.