The Install.
Build the layer. Deploy it. Instrument it. Hand off to the retainer.
Where the work actually happens. The Sweep filed an install plan; the Install executes it. Five layers deployed in sequence, instrumented against KPIs the executive sponsor signed off on, run by the same tiger team that scoped the work.
Deliverables.
- →Models picked per workflow — frontier where judgment matters, self-hosted where volume and margin do
- →Knowledge layer: firm knowledge embedded with retrieval tuned to your domain
- →Agents built for every workflow in scope, with eval harnesses and human checkpoints
- →The workflows themselves rebuilt around the agents — not bolted on top
- →Dashboards on every workflow, audit-grade lineage on every model call, governance signed by IT and legal
- →Team training and change management for the roles whose work is shifting
How the the install actually runs.
Models + knowledge online
Models picked per workflow — frontier for high-judgment work, self-hosted where volume and margin demand it. Firm knowledge ingested from engagement files, KBs, and documentation. Retrieval evaluated against a held-out set of internal queries. Governance signed off by IT.
Agents built and live
Agents built workflow by workflow. Each one tested against historical data before it talks to a customer or a partner. Voice and tone reviewed with the executive sponsor. Everything versioned in a Git repo — you have read access throughout.
Workflows rebuilt
The business process itself gets reshaped around the agents. Some workflows fully encoded; some left as hybrid human-plus-agent loops where judgment still wins. Affected teams trained in 30–60 minute sessions — not a six-week curriculum.
Dashboards and governance live
KPI dashboard goes live, instrumented to every workflow in scope. Audit logs route into your existing security tooling. Model-interaction lineage captured for every call. Governance memo updated to reflect the live system. QBR scheduled.
Phased rollout and handoff
We run the new system alongside the old one for 1–2 weeks per workflow before retiring the old path. Once everything is stable, the engagement formally transitions to the retainer. Same operators, different contract structure.
What you get on paper.
ANONYMIZED, REAL-LOOKING. ENGAGEMENT-SPECIFIC ARTIFACTS LIVE IN CLIENT NDA.
Live KPI dashboard
Hosted at a private subdomain. Real-time metrics on every workflow in scope. Reviewed monthly during the Retainer.
Knowledge layer (Pinecone / pgvector)
Your firm's knowledge, ingested, embedded, retrieval-tuned. Yours; we just run it.
Agent configs (Git repo)
Every agent's prompt, tools, guardrails, and evals — versioned. You have read access; we have write.
Governance memo v2
The Sweep memo, updated to reflect the live system. Defendable in regulatory exam, version-controlled with the agent repo.
Team training materials
Short videos + reference cards for each role whose workflow changed. Hosted in Notion.
Runbook for Retainer
What's automated, what's not, where to escalate, who owns what during ongoing operation.
Inside this phase.
Toolchain selected per-engagement. This is the typical stack.
- Anthropic Claude Sonnet/Opus
- OpenAI GPT-5
- Self-hosted Llama variants
- Pinecone / pgvector
- Inngest / Trigger.dev (agent orchestration)
- Vercel / Cloudflare (edge)
- Sentry / Datadog (observability)
- Stripe MCP / vertical MCPs
Your team's role during this phase.
- →Executive sponsor: 4 hours/week — workflow review, agent voice signoff, QBR prep.
- →Operational leads: 6–10 hours/week during their domain's build window; tapers off afterward.
- →IT lead: governance approval at Week 3 and Week 14 milestones.
- →Affected staff: 1 hour for training per role; 1 hour for shadow-run feedback per workflow.
- →Nobody quits their day job. The Install runs in parallel with the existing operation.
The exit.
At the end of the Install, every workflow in scope is running on the AI-native operating system. KPIs are measured. Audit logs are flowing. Your team is trained. The Tiger Team that built it transitions to running it on retainer — same operators, different contract structure.
Book an Ops Call.
30 minutes. Operator-to-operator. No deck. No follow-up nurture sequence designed to wear you down.
Book an Ops Call →