Case studies
Four workflows, four numbers
Two running in production today, one in build, one in design. Names withheld; we walk through any of them live on a call.
Support calls, answered day and nightIn production
Consumer wellness brand. Every support call answered, day or night. Routine work resolved end to end on the phone: orders, subscriptions, account changes, payments.
What runs: twin voice agents on the brand's existing support number. Callers verify identity first; a silent classifier routes them; agents do the work through 39 backend tools into the brand's commerce, ticketing, email, and payment systems. Edge cases warm-transfer to eight human queues by day and become tickets by night. Every call is logged with a transcript and outcome, transcripts get read, and fixes ship as versioned releases with regression checklists.
By the numbers (one production week):
- ≈600 calls routed through the agent selector
- 2 errors across 2,967 backend runs and 17,711 operations
- 216 designed conversation nodes, 39 backend tools
- 74 versioned production releases and counting
When the agent got a call wrong, the review loop caught it, shipped two guards as a versioned release, and flagged the customer for a human callback. The loop is the product as much as the agent is.
News that finds the analystIn production
Financial research team. Analysts scanned six news sources by hand for every company on a large coverage list. Most reading time went to duplicates, stale stories, and junk domains, with no audit trail behind judgment calls.
What runs: a pipeline that sweeps all six sources every 15 minutes and delivers a ranked, deduplicated, scored feed plus scheduled digests. Five narrow LLM jobs, each answering one question. Two model lanes: a private lane on a self-hosted open model so data never leaves the client's network, and a public frontier lane with failover. Analysts recalculate scores on demand and edit prompts without a deploy; a live dashboard shows AI spend.
By the numbers:
- Full sweep across six sources every 15 minutes
- 25 parallel workers under hard rate limits
- 5 narrow LLM jobs per article, not one big one
- 100% of AI calls logged with prompt, response, model, and cost
The system knows when it doesn't know. A page it can't read is flagged, not guessed.
The order queueIn build
Mid-market distributor. Orders arrive five ways: shared inbox, rep email, phone, a legacy form, an EDI trickle. Everything was keyed into the ERP by hand.
What the assessment found (2,504-ticket export): 94.6% of open tickets stuck at one stage, the oldest 138 days; 35 to 90 minutes of human time per multi-line order; 25 to 33% rework from missing details and wrong units; the real pipeline state living in three printed desk stacks.
What's being built: an intake layer of four agents behind one front door. Triage reads every channel and submits clean orders; a validator catches unit, pricing, and freight errors before entry; document linking kills the print-scan-reupload cycle; status notifications end the “where's my order” calls. Everything runs behind confidence gates: low-confidence orders route to the client's team with the agent's reasoning attached, and every action is logged with a reversal path.
The number: touchless order rate, with the intake backlog target set at under 500 open tickets from 2,368.
The client's team stops retyping and starts handling exceptions. Nobody at SectorFlow touches an order decision blind, and neither does the AI.
The playbook that answers backIn design
Channel sales enablement, B2B services vendor. A 60-page partner playbook is out of date the quarter it ships, and a rep in a live deal is not flipping to page 47.
What's designed: the service, industry, and persona matrix becomes a governed knowledge base. A rep types “mid-size manufacturer, no CISO, worried about ransomware” and gets the services to lead with, the discovery questions, and the objection responses. Leadership gets a weekly readout of what reps ask, which objections recur, and where deals stall. The same knowledge base generates the 20 to 25 page core playbook document, refreshed instead of reprinted.
The number: weekly active reps.
A binder reports nothing. The agent feeds go-to-market strategy every week.
