Why we put a pipeline in the middle
Most systems that touch service requests do it one of three ways, and all three lose context at the seam. AI made a fourth option viable: draw a line, put the pipeline there, and every downstream system gets the request already analyzed instead of raw.
How most operations wire this today
There are three shapes an intake-to-fulfillment operation usually takes. Each one works. Each one costs something the others don't. Verdikt is a fourth shape, made possible by AI carrying context that used to be a human's job.
The cost in one line each: inside the CRM, every change is a build; one-off per channel, five codepaths instead of one; siloed teams, humans re-read the same documents at every hand-off.
See each default in full
Default 1
Everything inside the CRMCase management platform becomes the system-of-record for intake, decisioning, correspondence, and fulfillment orchestration. Great for reporting; every change is a CRM build. Rules are hard-coded, letters live in template tools, decisions are humans clicking buttons.
Costs: every new program or channel is a project. Rules and letters drift from the CRM code that reads them.Default 2
One-off per channelEmail intake gets its own scripts. Fax has its own OCR pipeline. Portal writes directly to the case DB. EHR integrations go through a middleware. Each channel is normalized differently before landing in the same case store.
Costs: five extract-and-map codepaths instead of one. Bugs discovered in one channel don't automatically get fixed in the others.Default 3
Siloed teams stitching systemsIntake team hands off to an eligibility team, who hands off to a fulfillment team, who hands off to a compliance team. Every hand-off is an email, a spreadsheet, or a shared inbox. The case is the same; the context isn't carried across.
Costs: humans re-read the same documents at every step. Turnaround measured in days, not seconds.The bit that used to be the human's job
All three defaults exist because context (what does this case mean, is it decidable, which rules apply, what happens next) used to be something only a human could produce reliably, so every architecture pushed that work to people and left the systems as dumb pipes. Modern extraction, retrieval, and reasoning models change what a machine can carry across a seam: documents read into structured fields, rules retrieved and cited, a verdict produced with its explanation. Not perfectly, but reliably enough that downstream systems can trust context has been established and act on it. The pipeline in the middle stops being aspirational.
A logical boundary, not another silo
The line runs between "reading the case in context" and "writing the reply." AI carries the case across the first half; deterministic code carries the reply across the second half. Everything upstream of the line normalizes into one case; everything downstream receives it already analyzed.
The pipeline is the seam. AI stops before the reply is written; the deterministic render carries bytes-forever guarantees to every downstream integration.
Five things that were structural blockers before
These aren't features; they're properties of the shape. Once the line is drawn, the rest follows.
Every request enters analyzed
How
The pipeline reads the case, cites the rules, and produces the verdict before anything downstream sees it. Downstream systems act on meaning, not raw attachments.
Context flows across the seam
How
The verdict, the rationale, the rule citations, and the byte-identical reply travel together as one payload. When a CRM ticket opens, it's already annotated. When Pharmacy OMS receives a dispense order, it knows why.
New channels plug in without a rebuild
How
Fax, mail, portal, EHR all normalize into the same case shape before the pipeline runs. Adding a channel means writing one adapter, not rewriting the whole decision path.
New programs onboard as data
How
Program rules, extraction schemas, and reply templates are configuration, not code. A second program with a stricter cap and a different template is a folder, not a project.
Triage is a pipeline stage, not a human job
How
The pipeline classifies each case (new intake · refill · appeal · missing-info reply · not-decidable) and routes on what's actually inside. Refills short-circuit to Pharmacy OMS, appeals go to the CRM appeals queue, not-decidable cases are rejected with an audit row, and fresh intakes and missing-info replies continue down the full Decide → Deliver path.
The design, in three principles
Zoom out and the whole shape reduces to three choices: where AI belongs, how it's deployed safely, and where humans belong. Together, they let the operation pull the real power of AI while keeping the deterministic, auditable outcomes this industry demands.
-
01
AI upstream, deterministic downstream
The CRM and every downstream system stay AI-free. They receive verdicts, replies, and cited rules as data. All the AI work lives upstream of the line, moving cases toward a decision instead of summarizing decisions after the fact. The audited, regulated part of the stack never has to reason.
-
02
AI in the right place, with fallbacks
Not every step is an AI step. The pipeline picks the stages where AI genuinely adds signal (read, triage, verdict rationale) and leaves the rest deterministic. Every AI call has a fallback path so the pipeline degrades safely, not stops.
-
03
Humans in / on / over the loop
In-the-loop for edge cases the model isn't confident about. On-the-loop for silent oversight of routine cases. Over-the-loop for the reviews that actually need judgment: exceptions, appeals, trend calls. Humans stop doing the reading. They start doing the reviewing.