[ SERVICES 03 / 08 ] M E S Six playbooks
000%Scroll
00:00:00 UTCDistributed
[ 03 — Services ] Six playbooks

Agents we have shipped.

Each of these is a playbook we designed to run again and again. The agent looks different at every client — but the architecture, the eval harness, and the operating model are the same.

Continue
Index

Jump to a playbook.

01 · Sales SDR 02 · Support triage 03 · Invoice ops 04 · Content scheduling 05 · Deal underwriting 06 · Prior auth
Anatomy

How an agent actually works.

Every agent we ship looks roughly like this — a signal, a brain, the right tools, an action, and an audit log. The pieces change. The shape doesn't.

Signal
EMAIL · TICKET · WEBHOOK
Reasoning
CLAUDE · GPT
Tools
CRM · ERP · EHR
Action
SEND · POST · SUBMIT
↓ BEFORE

A person, doing it slowly.

→ Read ticket / lead / invoice
→ Switch to CRM / ERP / EHR
→ Look up context (5 tabs)
→ Write reply / fill form
→ Double-check, send, log
≈ 12 — 45 minutes per item
↑ AFTER

An agent, doing it constantly.

→ Signal received
→ Context retrieved (parallel)
→ Reasoning + tool call
→ Action executed
→ Logged + observability
≈ 3 — 30 seconds per item
SVC · 01 / SALES

A sales rep that never sleeps.

The agent ingests your ICP, pulls account signals from a dozen sources, drafts personalised first-touch and follow-up emails, books meetings on your calendar, and updates your CRM. It works the whole pipeline, 24/7, for less than a single rep's salary.

  • Account research from Apollo, Clay, LinkedIn, news, funding data
  • Personalised outreach written in your brand voice
  • Reply classification + automatic follow-up cadences
  • Calendar booking via Cal.com or Chili Piper
  • Two-way sync with HubSpot, Salesforce, Attio, Pipedrive
[ LIVE FEED ]
→ research(acme.co)
→ draft(tone="founder")
✓ sent · 2m ago
✓ booked · 30 min
SVC · 02 / SUPPORT
[ TICKET # 4451 ]
"My export to CSV is stuck again..."
bugpriority · P2
"Hey Tara — checked the logs, the queue is backed up. Restarted it. Try again now?"
✓ AUTO-SENT · 23s

Triage every ticket. Resolve the easy ones.

The agent reads every ticket, classifies intent and priority, retrieves the relevant context from your help-center and past tickets, then drafts a reply in your brand voice. Auto-sends or routes to a human with everything they need.

  • Intent, sentiment, and priority classification
  • RAG over help-center, past tickets, and product docs
  • Draft-first or full-auto reply modes
  • Escalation rules and human-in-the-loop checkpoints
  • Native in Intercom, Zendesk, Front, HelpScout, Plain
SVC · 03 / OPS

Make the back office disappear.

An OCR + LLM pipeline that ingests any document — invoices, contracts, packing slips, expense receipts — extracts every line item, validates against your master records, flags anomalies, and posts to your ERP. We've processed 120K documents to date.

  • Vendor-agnostic invoice extraction (PDF, image, email, EDI)
  • PO and GL coding with confidence scoring
  • Three-way match and anomaly detection
  • NetSuite, QuickBooks, Xero, SAP B1 integrations
  • Human review queue for low-confidence items
[ INVOICE EXTRACT ]
VendorDALMOR INC
PO match✓ PO-4451
Total$ 18,420.00
Confidence98.4%
Posted→ NetSuite
SVC · 04 / CONTENT
[ THIS WEEK · 7 POSTS ]
MON 09:00 · LinkedIn
"Three lessons from shipping our first agent..."
TUE 11:30 · X
"Our agent booked 14 meetings last night..."
THU 08:00 · Newsletter
"The hidden cost of dashboards..."

A content engine that sounds like you.

Trains on your founder's voice and your brand library, then produces a weekly content calendar across LinkedIn, X, blog, and email. You approve from a single queue. The agent learns from edits.

  • Brand voice fine-tuning from existing posts
  • Topic mining from sales calls, support, product roadmap
  • Multi-channel formatting (carousel, thread, blog, newsletter)
  • One-click approval flow plus scheduling
  • Performance feedback loop on engagement signals
SVC · 05 / REAL ESTATE

Underwrite a deal in nine seconds.

Paste an OM or address. The agent pulls comps from CoStar, MLS, and public records, models rent rolls, runs cap-rate sensitivity, and outputs a one-pager with the go/no-go signal. Built for syndicators, family offices, and broker shops.

  • OM parsing and T-12 rebuilds from PDF
  • Comp pulls from CoStar, Reonomy, MLS, public records
  • Rent roll and expense reconstruction
  • Cap rate, IRR, cash-on-cash sensitivity
  • Risk flagging on title, zoning, environmental
[ DEAL · WILLOW HEIGHTS ]
IRR
17.4%
CAP
6.8%
CoC
9.1%
DSCR
1.42
SVC · 06 / HEALTHCARE
[ MRN 884301 ]
MRI · Lumbar w/o contrast
BCBS · TX
Criteria met5/5
Evidence attached3 items
Status→ Submitted

Prior auths that submit themselves.

HIPAA-eligible workflow that reads patient charts in your EHR, identifies the payer's medical necessity criteria, fills the form, attaches the right clinical evidence, and submits. Currently in production at three regional health systems.

  • Epic, Cerner, athenahealth EHR integrations
  • Payer policy ingestion + criteria mapping
  • Evidence retrieval from chart, labs, and notes
  • Status tracking and appeal automation
  • BAA-signed, audit log, full traceability
07 / Stack

Boring tech. Boring on purpose.

Anthropic
OpenAI
LangGraph
Temporal
Postgres
Vercel
AWS
GCP
Pinecone
Braintrust
Datadog
Sentry

Don't see your
workflow?

If you can describe a repeatable, rule-driven workflow in a 30-minute call, we can probably build an agent for it.

Describe your workflow