How JieGou works with your engineering team
Four engagement phases. Architecture review (free) → 90-day lighthouse engagement → expansion pipeline → steady-state operation. Engineering-led mid-market IT teams ($50M-$1B). Operator-grade safety nets active from day 1.
§1 — The four engagement phases
Architecture review → lighthouse → expansion → steady-state. Each phase has a clean exit.
Architecture Review
Your engineering leadership evaluates our substrate against your actual stack — before any commercial conversation deepens.
What happens
- ▸ Read the /reference-architecture page (7-component decomposition, 3 trust boundaries, 10 named failure modes)
- ▸ Run the /10-layer-assessment self-audit; share scores if useful
- ▸ 30-min discovery call with the founder — no deck, no demo; we look at your existing AI stack with you
- ▸ Optional: signed mutual NDA + 1-2 deeper architecture sessions with your engineering team
- ▸ Outcome: a clear "this fits / this is the wrong shape" answer between operators
Decision point
Both sides decide whether Phase 1 makes sense. Either side can walk away cleanly — no proposal yet, no commitment, no contract obligation.
Lighthouse Engagement
Operate ONE bounded workflow end-to-end with your team — proves the substrate, proves the partnership, produces named-outcome evidence for the Phase 2 expansion decision.
What happens
- ▸ Week 1-2: Phase 0 NDA → SOW → mutual kickoff; identity + access provisioning; staging environment shared
- ▸ Week 3-8: Build + Shadow Mode operation. Your engineers integrate against our substrate; our operators run the workflow alongside your team in Shadow Mode (all AI actions reviewed by humans before customer-facing emission)
- ▸ Week 9-12: Supervised → Trusted graduation. Workflow lifts from Shadow Mode to Supervised (high-confidence actions auto-emit; low-confidence routes to humans) to Trusted on the bounded scope only.
- ▸ Throughout: weekly written status notes, monthly executive reports, full audit trail, end-of-Phase-1 retrospective document with named outcomes
Decision point
End-of-Phase-1: jointly decide whether to expand (Phase 2). If outcomes hit Phase 1 success criteria → Phase 2 SOW. If they don't → clean break with the substrate, audit trail, and operator-graded retrospective document yours to keep.
Expansion Pipeline
Add workflows to the operating substrate one at a time. Each new workflow follows the same Shadow → Supervised → Trusted graduation pattern. Pipeline grows as workflows compound; the substrate stays the same.
What happens
- ▸ Per-workflow scoping: 2-4 week SOW per new workflow (Discovery → Build → Shadow → Trusted)
- ▸ Steady-state operation of all in-flight workflows on the shared substrate
- ▸ Quarterly business reviews: pipeline health, audit findings, governance posture, expansion decisions
- ▸ Continued integration with your evolving stack (new tools added, schemas changed, team rotations) — substrate adapts; workflows extend
Decision point
Per-workflow expansion decisions happen on a rolling basis. You're never locked into a multi-workflow commit. Workflows can be paused, decommissioned, or migrated back to your team at any time.
Steady-State Operation
Operating substrate is a steady-state utility. Multiple workflows in production; new workflows added on cadence that matches your business; existing workflows monitored, audited, and improved on schedule.
What happens
- ▸ Continuous operation across all production workflows
- ▸ Audit-evidence delivery on board / regulator / customer cadence (quarterly default; on-demand available)
- ▸ Annual architecture review: 10-Layer Assessment re-run with your team; substrate gaps identified; remediation roadmap
- ▸ Workflow lifecycle: graduate workflows to your in-house team when they're stable enough to operate without our involvement (hand-back is the success outcome, not lock-in)
Decision point
Annual renewal decision. Either side can exit on 90-day notice. Hand-back documentation provided in any exit scenario — your operating capability outlasts our engagement.
§2 — Three safety nets, active from day 1
AI will be wrong sometimes. The architecture assumes that and provides three independent catch-and-correct mechanisms.
Approval gates
Any workflow declares approval gates at any step. The agent pauses, full context surfaces to the named approver (input, draft output, prior-step trail, model version), and the approver explicitly approves, edits, or rejects before any external-facing emission. Routable to specific people, roles (CIO / CTO / Director-Eng / Architect), or escalation chains. Single-approver + multi-approver policies both supported.
Audit trail with hash-chain integrity
Every system + human action logged with input, output, model version, timestamp, and HMAC-signed for tamper detection. Optional hash-chain integrity for SOX / FDA / EU AI Act evidentiary contexts. Append-only by design; export to your SIEM (Splunk / Sentinel / syslog). If something looks wrong six months in, you can reconstruct exactly what happened, when, by whom, with what data.
Kill switch + structured-feedback loop
One click flips any workflow back to Shadow Mode or disables it entirely. Plus: when an operator rejects or edits an AI action, the structured-feedback record routes to the agent's policy memory so the same class of mistake gets caught earlier next time. The system learns from corrections, not just from successes. Every kill-switch event triggers an incident ticket in your tracking system.
Active in Shadow Mode, in Supervised mode, and in Trusted mode. Different layers do different work at each graduation level — none of them turn off.
§3 — What's NOT in scope
Operator-honest exclusion list. The work we explicitly do not take on.
- ✕ We do NOT operate your entire IT estate. We operate bounded workflows; you continue to own the broader stack.
- ✕ We do NOT replace your in-house engineering team. We work alongside them; they integrate; we operate; capability transfers back to them over time if you want it to.
- ✕ We do NOT bring our own LLM. We use Anthropic, OpenAI, or Google — your choice; provider-portable workflows by design. BYOK supported.
- ✕ We do NOT lock workflows into our platform. Hand-back documentation is part of every engagement. Your data, your prompts, your audit trail — yours.
- ✕ We do NOT operate workflows we can't audit. If a workflow can't be instrumented for audit-evidence emission, we won't take it on. Audit-evidence is the operating-cost floor.
- ✕ We do NOT do "AI strategy consulting." Our delivery is operating substrate + workflows in production. If you want strategy decks, hire a strategy consultancy.
These boundaries are why our customer roster stays small and why the engagements that work, work. If your situation needs something on this list, we'll say so honestly and point you at the right kind of partner.
§4 — Compliance + regulated-data posture
10-Layer Governance framework + BAA-eligible HIPAA workflows + EU AI Act risk-tier mapping.
The 10-Layer Governance framework is the architecture we use to operate AI for paying customers in regulated industries. The same framework is published publicly at /10-layer-assessment so you can baseline your own posture before, during, or after any conversation with JieGou.
Cross-framework mapping: SOC 2 Common Criteria (especially CC6 Logical & Physical Access + CC7 System Operations), EU AI Act risk tiers + conformity assessments, NIST AI Risk Management Framework, ISO/IEC 42001 AI Management System. HIPAA-aligned workflows are BAA-eligible. GDPR data-residency controls configurable per workflow. SOX + FDA evidentiary contexts supported via hash-chain audit-trail integrity.
Audit-evidence emission is the operating-cost floor: if a workflow can't be instrumented for audit evidence, we won't take it on. This is non-negotiable across phases.
FAQ
CIO + CTO questions on Phase 1, hand-back, and the catch-and-correct architecture.
See the architecture. Or schedule the discovery call.
Reference architecture is operator-grade detail you can evaluate without scheduling anything. Discovery call is 30 min, no deck, no demo — we look at your existing AI stack with you. Honest either way.