One staff-facing chat surface (Carcin) on top. Underneath: three Phase 1 agents running daily across enforcement, inactivity detection, and settled-case banking. A human strategist setting the play, supervisors keeping authority. The Repo is yours from the moment of creation, source code, prompts, skills, memory, configurations. Work-for-hire.
Before we go deeper, here's how we use a few words. The proposal makes more sense once we share the vocabulary. The Repo contains everything else; the rest are the components inside it.
Actionstep is the system of record. It is not a system of detection. Tasks go overdue without anyone noticing. Cases go inactive for weeks before a supervisor catches it. Settlements sit in CLS for 30+ days because nobody's tracking the milestone clock. The original SteinLaw Workflow Build Spec Pack identified these as ten distinct workflows. We start with the three that hurt most today, the Phase 1 enforcement layer, and we operationalize the fix continuously rather than discovering it in a monthly review.
Every overdue task, every stalled case, every settled-but-unbanked matter is sitting in Actionstep right now, waiting to be detected. What's missing is a layer that scans for them three times a day, routes the alert to the right supervisor, escalates if uncorrected, and produces a digest at end of day. The data exists. The accountability ladder doesn't. Phase 1 builds that ladder for the three workflows where the cost of a missed signal is highest, missed deadlines, stalled files, and delayed banking.
This isn't a generic legal-AI playbook. The decision logic is built against SteinLaw's actual operating model, the supervisor hierarchy, the no-noise alert discipline, the Florida Bar UPL boundary, the IOLTA trust-account rules, the existing role of the Unit Operations Head. AI never crosses into legal advice or client-facing communication. AI never moves money. Detection and recommendation only. Humans keep enforcement authority. The Phase 1 build hardens that boundary into the architecture from day one.
For SteinLaw, the hub is Actionstep: the system of record where every matter, task, note, and case event lives today. Phase 1 doesn't replace it. Phase 1 reads from it on a schedule, mirrors what we need into a local Postgres store on a Mac mini inside the firm (the privileged-data residency layer), and runs detection from the cloud orchestrator, then writes alerts and digests back through the proper accountability ladder. Hybrid by design: sensitive data stays in the building, elastic compute runs in the cloud.
These are the six domains the firm's AI ecosystem will eventually span, mapped from the original Workflow Build Spec Pack. Phase 1 ships agents inside three of them: Case Operations, Supervisor Enforcement, and Banking. The other three (Compliance scaffolding, Reporting digests, Audit Trail) get the substrate built in foundation but the workflow agents inside them are deferred to later phases. The framework absorbs the rest later without rework.
The original Workflow Build Spec Pack ranked all ten workflows by complexity, time-to-value, and dependency. Phase 1 selects the three with the highest weighted score given today's pain at SteinLaw, missed deadlines, stalled files, and slow banking. The same engine produces the ranking when Phase 2 (Workflows 4-6) and Phase 3 (Workflows 7-9) come up for re-scoping. Output is a deterministic ranked list. We work top-down.
Before any workflow agent is deployed, SteinLaw receives The Repo, the substrate (hybrid cloud + local), the Actionstep mirror, the compliance scaffolding, the audit trail, the eval harness, and the Carcin daily digest. The 6-month sprint covers building all of that plus shipping the three Phase 1 workflows, then a short operate-and-tune window. Workflows 4-10 from the original Build Spec Pack, plus the Case Vision migration and LawScore revival sprints, are tabled for a future engagement.
The original Spec Pack defined ten workflows across four phases. Phase 1 ships the three enforcement-layer workflows: Missed Task Enforcement, Case Inactivity Detection, Settled Case Banking. Foundation gets built first (M1-M2); workflows roll out M2-M5; M5-M6 is operate, tune, and Phase 2 scoping decision.
The Repo is the substrate, the owned codebase, the Actionstep API integration, the Postgres mirror, the compliance scaffolding, the audit trail, the eval harness. On top of it sits Carcin, the staff-facing chat surface for the daily digest and ad-hoc queries. Underneath: three Phase 1 agents running autonomously across the firm's task, matter, and banking pipelines, with a human strategist setting the play and supervisors keeping enforcement authority. Cancel or scale back any agent at any time.
# The Repo dispatches the work, humans & agents execute against Actionstep $ ad-apt run --client "[CLIENT NAME]" --scope "phase-1" # Carcin, staff-facing chat surface, sits on top of The Repo Carcin → daily digest · role-scoped queries · production # Phase 1 enforcement layer · 3 workflow agents W1 · Missed Tasks → Actionstep · 3x daily sweep · supervisor routing · escalation ladder W2 · Inactivity → daily inactivity scoring · Yellow/Orange/Red tiers · Red same-day W3 · Settled Banking → CLS pipeline tracking · day-30 / day-90 breach flags · IOLTA-safe (track only) # Substrate agents that protect every Phase 1 workflow Compliance Layer → BAA-tier model selection · redaction · UPL guardrail · conflict screen Audit Trail → per-decision lineage · daily Merkle root mirrored to GitHub Eval Runner → labeled set per workflow · regression gate · drift detection # Tabled at client request, not in this engagement Case Vision Bridge → DEFERRED LawScore Healer → DEFERRED # Human gate, supervisors and strategists approve before anything ships Strategist → APPROVED [ DEPLOY ] → LIVE
Carcin and the agent stack do the production. Three humans set the play, approve before anything ships, and own the outcome.
These three workflows come directly from SteinLaw's original Workflow Build Specification Pack. We are not inventing them. We are building exactly what was scoped, with every required field, decision rule, alert discipline, and escalation path preserved. Below is each workflow expanded against the spec, so the firm sees precisely what gets shipped.
One ongoing retainer covers everything for Phase 1, building The Repo, shipping the three Phase 1 workflow agents, running Carcin, iterating on prompts, monitoring evals, supervisor onboarding, monthly review. Six-month sprint cadence. The retainer ramps up M1-M3, peaks at M3-M4 during build + UAT, then declines through M5-M6 as the system stabilizes. Cancel or scale back any workflow at any time, hours saved, retainer reduced, AI/API spend lowered.
SteinLaw owns The Repo from the moment of creation. Source code, prompts, skills, memory, configurations, work-made-for-hire. Hardware (if local path is chosen) is SteinLaw's. Cloud accounts (Hetzner, AWS, Anthropic, OpenAI) registered to and paid by SteinLaw directly. The retainer covers all of the work below, end to end, building, tuning, running, and iterating.
| Month | Focus & deployments | Retainer | Infra | AI/Tokens | Total |
|---|---|---|---|---|---|
| M1 |
Foundation kickoff · hybrid infra stand-up
Discovery + Actionstep field auditCloud VPS provisioningCompliance scaffolding start
|
$3,500 | $200 | $200 | $3,900 |
| M2 |
Foundation finish + Carcin live + W1 build
Carcin daily digestAudit trail + eval harnessW1 Missed Task Enforcement build
|
$4,500 | $250 | $250 | $5,000 |
| M3 |
W1 ships + W2 build · optional local hardware
W1 Missed Task Enforcement3x daily sweep + escalation ladderW2 Inactivity scoring build+ Mac mini (optional, as needed)
|
$4,500 | $250 +$2,000 opt |
$300 | $5,050 $7,050 w/ hw |
| M4 |
W2 ships + W3 build
W2 Case Inactivity DetectionYellow/Orange/Red severity tiersW3 Settled Banking build
|
$3,000 | $250 | $300 | $3,550 |
| M5 |
W3 ships + UAT + supervisor onboarding
W3 Settled Case BankingPer-workflow acceptance gatesSupervisor onboarding + training
|
$2,500 | $200 | $250 | $2,950 |
| M6 |
Operate + tune + Phase 2 scoping
Eval drift monitoringPrompt tuning + friction reviewPhase 2 / Case Vision / LawScore re-scope
|
$2,000 | $200 | $250 | $2,450 |
| Total | Foundation + 3 Phase 1 workflow agents, deployed against Actionstep with hybrid (cloud + local) infrastructure. Case Vision and LawScore sprints tabled at client request. | $20,000 | $1,350 +$2,000 opt |
$1,550 | ~$[TOTAL_12MO] ~$24.9K w/ hw |
The whole point of The Repo is that it's SteinLaw's. The whole point of the retainer is that it adjusts. Three guarantees keep both true, at any point in the engagement, the firm can pause work, scale it back, or walk away with everything we've built. Re-scoping Case Vision or LawScore in a future engagement uses the same flexibility.
This is the data payload powering the rendered proposal. Phase 1 only, 6-month sprint, hybrid infrastructure. Case Vision migration and LawScore revival sprints are tabled at client's request. Optional one-time hardware spend (~$2K in M3) covers the local Mac mini for privileged data residency.