The provider has zero prior knowledge of how the client operates. This phase establishes ground truth before any technical work begins. All assumptions are documented and reviewed with the client before Phase 1 commences.
- Send pre-engagement questionnaire covering org structure, daily workflows, and tech stack
- Schedule 2-hour discovery workshop with all 7 staff
- Shadow one loan origination cycle end-to-end (observe, don't advise)
- Audit existing tools: email, CRM, loan management software, document storage
- Identify top 5 time-consuming and error-prone workflows by volume
- Document data flows, handoff points, and bottlenecks
- Complete questionnaire before kickoff call (allow 2 hrs)
- Provide read-only access to CRM, document systems, and email templates
- Assign one internal point of contact (project champion)
- Prepare list of current monthly pain points from each staff member
- Share sample loan files (redacted) to understand document types
Using the discovery findings, the provider designs the specific Claude agents and integration architecture. No code is written yet — this phase produces the blueprint that both parties approve before build begins.
- Prioritize automation candidates by ROI and feasibility
- Design agent architecture: define each agent's role, inputs, outputs, and handoffs
- Map Claude API integration points to existing systems
- Define data schemas for loan files, borrower records, and document types
- Draft security model: data handling, PII policy, API key management
- Produce agent specification document for client review and sign-off
- Estimate token usage and monthly Anthropic API cost at scale
- Review and approve the agent specification document
- Confirm which workflows are in scope vs. out of scope for Phase 2
- Provide IT credentials or support for API/integration access
- Sign off on data handling policy before development begins
First build sprint. The provider develops and internally tests the two highest-priority agents: Loan Intake and Document Review. The goal is a working demo on real (anonymized) client data by the end of the sprint.
- Set up development environment, Claude API integration, and version control
- Build Loan Intake Agent: parse application forms, extract structured data (borrower name, loan amount, property address, LTV, requested terms)
- Build Document Review Agent: ingest PDFs (bank statements, tax returns, title reports), produce structured summaries and completeness checklist
- Build lightweight internal dashboard to view agent outputs
- Write unit tests for agent accuracy on sample loan files
- Internal QA: test edge cases, missing documents, and ambiguous inputs
- Prepare demo environment with anonymized test data
- Provide 10–15 real (redacted) loan files for testing accuracy
- Designate one staff member as day-to-day tester for weekly check-ins
- Attend mid-sprint check-in (30 min) to review progress and flag scope issues early
- Claude API connected and prompts versioned in repo
- Intake Agent accurately extracts fields from ≥90% of test files
- Document Agent produces summary + missing-item flag for all major doc types
- Internal demo delivered to client at sprint end
Second build sprint. Borrower-facing communications are automated and the underwriting assistant is built. These agents directly reduce the manual writing load on staff and surface deal intelligence faster.
- Build Borrower Comms Agent: generate status update emails, document request letters, approval/decline notices, and follow-up sequences — all in the client's voice and tone
- Build Underwriting Assistant: given a structured loan record, produce a deal memo draft, flag key risk items, and surface relevant comparables or notes from similar past deals
- Integrate agents with each other: Intake Agent output feeds into Document Agent and Underwriting Assistant automatically
- Add human-in-the-loop review step: staff approves before any email is sent
- Build email preview UI within dashboard
- Provide 20–30 example emails (past sent items) to train tone-matching
- Review and give feedback on Comms Agent drafts weekly
- Provide 5–10 past deal memos for Underwriting Assistant benchmarking
- Attend weekly 45-min review calls throughout sprint
- Comms Agent produces on-brand drafts rated "usable with minor edits" by staff
- Underwriting Assistant deal memo matches client's format and hits key risk categories
- End-to-end pipeline: new application → structured record → doc summary → deal memo, no manual intervention
All 7 staff use the full system on real live deals for the first time, with the provider actively monitoring for errors, edge cases, and friction. This is the most important feedback loop in the project.
- Deploy to staging environment connected to client's live systems (read-only where possible)
- Conduct 2-hour onboarding session with all 7 staff
- Provide quick-reference guide for each agent
- Monitor agent outputs daily during UAT; log failure cases
- Triage and resolve critical bugs within 24 hours
- Collect structured feedback from each staff member via short weekly survey
- Tune prompts and logic based on real usage patterns
- All 7 staff use the system daily on real loans throughout UAT
- Log every issue, surprising output, or missed case in shared tracker
- Complete weekly feedback survey (5 min)
- Escalate any data errors or compliance concerns to provider immediately
- Project champion prioritizes bug backlog jointly with provider
- ≥90% of intake extractions rated correct by staff
- ≥80% of comms drafts used with minor or no edits
- Zero data errors or PII leakage incidents
- All 7 staff report confidence using the system independently
The system goes fully live. All agents run on production data. Staff operate independently. Provider transitions to a support and documentation role with a 2-week hypercare window post-launch.
- Migrate from staging to production environment
- Final security review and penetration check on all API integrations
- Set up monitoring, alerting, and error logging (email alerts for failures)
- Produce full technical documentation: system architecture, API key management, how to update prompts
- Record 3–5 short walkthrough videos (one per agent) for staff reference
- Conduct final all-hands walkthrough call
- 2-week hypercare: provider available same-day for any production issues
- Confirm all systems are live and stable before handoff sign-off
- Designate internal owner for monitoring dashboard
- Review and store technical documentation securely
- Sign project completion acceptance
- All 4 agents deployed and processing live loans
- Monitoring and alerting active
- Documentation delivered and reviewed
- Staff training complete for all 7 members
- Hypercare period active (2 weeks post-launch)
Included support window after full handoff. The provider monitors passively, resolves emerging issues, and delivers a final usage report with ROI metrics and optional Phase 2 expansion recommendations.
- Weekly 30-min check-in call with project champion
- Monitor error logs and usage metrics passively
- Resolve any post-launch bugs or prompt drift issues
- Deliver final ROI summary: time saved per workflow, volume processed, error rate
- Present Phase 2 expansion roadmap (optional, additional scope)
- Report any new issues within 48 hours via shared Slack/email channel
- Track staff time savings informally to support ROI reporting
- Decide on retainer or Phase 2 engagement by end of support period
| Phase | Apr W2 | Apr W3 | Apr W4 | May W1 | May W2 | May W3 | May W4 | Jun W1 | Jun W2 | Jun W3 | Jun W4 | Jul W1 | Jul W2 | Jul W3 | Jul W4 | Aug W1 |
|---|