AI in the platform
The paperwork reads itself. A person still signs off.
Requisitions reach you as forwarded emails, faxed PDFs, phone photos and spreadsheets somebody maintains by hand. Our platform reads all of it into a structured draft order — locations matched to your saved sites, service, temperature profile, window, STAT flag — and then stops and waits for a human to confirm. That pattern holds everywhere a model touches the work.
Three sentences that govern every model in the system
The model drafts
It reads, extracts, ranks and explains. Every output is a proposal with its source attached.
A person confirms
Nothing is ordered, dispatched, sent or closed until someone with the authority to do it presses confirm.
The engine decides
Qualifications, handling rules, capacity and cutoffs are settled by deterministic code. A model never decides eligibility.
Order intake
Send it the way it already arrives
Forward the email. Drop the PDF requisition. Photograph the fax on the bench. Upload the pickup list your coordinator keeps in a spreadsheet. Or type one sentence. What comes back is a structured draft order with every field traced to the words it came from, and it stays a draft until you say otherwise.
What you can send
-
A forwarded email
Ordering staff keep writing the email they always wrote.
-
A PDF requisition
Attached, or dropped straight into the portal.
-
A photo of a fax
Taken on a phone at the bench, skewed and half-toned.
-
A spreadsheet
Tomorrow's pickups, in the file your coordinator already maintains.
-
One typed sentence
"Pickup at Quest Edison tomorrow 10am, 3 refrigerated specimens to LabCorp Raritan."
What comes back
Draft order
FCL-4418-A · unconfirmed
- Pickup
-
Quest Diagnostics — Edison, NJ
matched to your saved site QD-EDI-02
- Delivery
-
LabCorp — Raritan, NJ
matched to your saved site LC-RAR-01
- Service
-
Medical specimen transport
from "3 specimens"
- Temperature
-
Refrigerated, 2–8 °C
from "refrigerated"
- Window
-
Tomorrow, 10:00 – 10:30
from "tomorrow 10am"
- Priority
-
Routine
no STAT language found in the request
- Packages
-
3
from "3 refrigerated specimens"
In the portal
Paste, type or upload on the New Request screen and watch the fields fill in front of you.
By email
Every account gets its own intake address. Forward a requisition and the draft is waiting in your portal.
Through the API
Post raw text or a document and receive the same structured draft your coordinators would review.
The source document and the draft it produced are both kept on the order, so an auditor can see what arrived, what the model made of it, who confirmed it and which fields they changed.
Where else it earns its place
Each of these is a model doing the reading, the ranking or the writing, and a person doing the deciding. The guardrail is written next to the feature because it is part of the feature.
Dispatch copilot
The board ranks who should take a job and says why in a sentence a dispatcher can argue with: "Maria is eight minutes out, already routed past this stop and phlebotomy-certified for this lab. Jamal is closer, but his shift ends before the window." Step away from the board and a short digest tells you what changed while you were gone.
The engine clears who is eligible. The model only orders and explains the candidates it is handed.
Why is this late
Ask any running job why it is behind and get the combination in plain language: traffic on the lane, dwell at the current stop, and what this site has historically done at this hour.
Every explanation quotes the timestamps and readings it was built from.
Exception triage
Likely late, failed pickup, temperature excursion, tracking lost — each alert arrives with a suggested next action and a drafted client message that carries no patient information: "Your 2 pm pickup is running about 15 minutes behind because of traffic on I-95. New ETA 2:20." One click to send it, one click to rewrite it.
Drafted under a no-PHI constraint, and no message leaves the system until a dispatcher sends it.
Your account manager, at 2 a.m.
The portal answers the questions people normally phone in: where is my order, what did we ship last week, what does a STAT to Newark cost.
Answers are assembled from tool calls against your own scoped data. If the data cannot answer it, it says so instead of composing something plausible.
Quality and compliance drafting
A quality event arrives with a drafted root cause and corrective action built from its own timeline. SOP updates are drafted from what changed. An audit-log query comes back summarised in the language an auditor asked the question in, with evidence packs assembled per control for ISO 27001 and 9001.
A draft is unsigned work. Your quality owner edits and signs it, and every line links back to the timeline entry behind it.
Credential reading at onboarding
A courier or phlebotomist uploads a licence, an insurance certificate or a certification and the platform pulls out the holder, the issuing body and the expiry, flags a name that does not match the application, and puts the reviewer in front of a decision instead of a PDF.
A human reviewer approves every credential. Extraction shortens the reading, not the decision.
Demand and staffing forecast
Volume per service area and per hour, built from your own history, so route planning and phlebotomist staffing are argued from a curve rather than from memory.
The forecast is classical statistics on your data. The model writes the note that explains the shape; it does not invent the shape.
Voice in the field
A courier holding a cooler taps the microphone and says "arrived, patient not home, left card." It lands as a structured outcome with a note attached, in the seconds it would have taken to unlock a phone.
The transcript is kept beside the structured outcome, and the worker sees what was recorded before it is filed.
What the model is never allowed to do
- Decide whether a worker is qualified, a vehicle is suitable or a window is achievable.
- Create, dispatch, reassign or close an order on its own.
- Send anything to your clients or patients without a person pressing send.
- Invent an ETA, a temperature reading or a scan that no device produced.
- See patient or clinical detail that the task in front of it does not require.
Your data is not used to train public models, and every model-assisted action is written to the same audit log as a human one, naming the model, the input it read and the person who confirmed it.
What is live, and what is in build
These capabilities reach accounts in stages. Before anything is signed we put in writing which of them is available to you on day one and which is still in build, because a demonstration on your own paperwork should not turn into a feature that quietly arrives later.
Does any of this touch patient data?
Only where the task requires it, and only inside the platform. Client-facing drafts are generated under a constraint that excludes patient information, transport records stay transport records, and the access rules that apply to a person apply to a model acting on that person's behalf.
What happens when the model reads a requisition wrong?
The same thing that happens when a person reads it wrong, except earlier: the draft is on screen with each field traced to the text it came from, so the mistake is visible before the order exists. Corrections are recorded against the draft, and they are what we use to improve the extraction.
Can we turn parts of it off?
Yes. Intake, the portal assistant and the drafted client messages are switched per account. Some organizations run intake and nothing else; that is a normal configuration, not a downgrade.
Is the AI deciding who picks up my specimens?
No. Eligibility and assignment rules are deterministic code, and a dispatcher makes the call. The model ranks the candidates that code has already cleared and explains why it put them in that order.