AI for Durable Medical Equipment Suppliers: Referral Intake and Calls

A durable medical equipment supplier can use AI to pull key fields from incoming referrals into a checklist, flag missing paperwork, draft chasers to referrers and patients, and run delivery-window confirmation calls. Humans still approve every order, decide whether documents satisfy the payer, and handle any call involving health concerns.

Most of the delay in this trade is not the equipment. It is the loop of incomplete referrals, unsigned orders, missing notes and phone tag that sits between "referral received" and "delivered". That loop is made of repeatable steps, which is where AI helps, and every step touches patient information and payer rules, which is why a person must stay in charge. The three jobs below run in order: intake, chasing, and delivery calls.

Follow me on Instagram@sagnikteaches

Where referrals get stuck: map it before you automate

Pull the last 30 referrals that took longest to deliver and note, for each one, where it waited. Do this on a spreadsheet with a column for each stage and a date for each hand-off. You will almost always find the same few holdups, and these decide where AI is worth adding.

Connect on LinkedInSagnik Bhattacharya

If your slowest referrals wait mostly on payer authorisation, AI will not shorten that part, and you should say so honestly. If they wait on missing paperwork or unreachable patients, you have found the best starting point.

Subscribe on YouTube@sagnikteaches

Job one: a referral intake checklist AI can fill and a person confirms

The foundation is a standard intake record, the same fields every time, whatever format the referral arrives in. Build it from the payer rules you actually bill under, as each payer publishes its own documentation list and updates it. Your compliance lead or billing manager should own the master list. What follows is a typical shape, not a rule for any particular payer.

Fields in the standard intake record

• Patient: name, date of birth, contact numbers, delivery address, preferred contact method, best times to call, anyone authorised to speak for the patient.
• Coverage: payer name, member or policy number, secondary coverage if any, whether authorisation is required for the item.
• Referrer: name, practice, contact details, the person who signs orders.
• The order: item or category, quantity, date written, prescriber name and identifier, whether it is signed, and length of need where the payer asks for it.
• Clinical documentation: which notes accompany the order, their date, and whether they support the item according to the payer's documentation list. (This is a check for presence, not a clinical judgement.)
• Delivery details: access notes, stairs, who will be present, any set-up requirements.
• Status: complete, incomplete (with the list of what is missing), or on hold.

Extraction: a prompt and what to expect

Most AI tools used for document extraction work by reading the page and filling the fields you ask for. With a general model, the prompt below shows the idea; in practice, your intake system or a document-capture tool may do the same job inside the software. Use only tools that have passed your privacy review.

Read the referral text below and fill ONLY these fields. If a field is not
clearly present, write NOT FOUND. Never guess or infer.

Patient name:
Date of birth:
Phone:
Delivery address:
Payer and member number:
Referrer name and practice:
Item or category ordered:
Quantity:
Date order written:
Order signed (yes / no / unclear):
Clinical notes attached (list titles and dates):
Anything unusual or contradictory:

Then list any fields where two parts of the document disagree.

REFERRAL TEXT
[paste]
END

The line "never guess or infer" matters, because a model asked to fill a date will sometimes supply a plausible one. "NOT FOUND" turns a hidden error into an obvious gap. Here is an illustrative output for a referral whose scan is clear except for a smudged line:

Patient name: [as written]. Date of birth: NOT FOUND (smudged). Item ordered: walking frame. Quantity: 1. Date order written: 14 October. Order signed: unclear (signature line appears blank). Anything unusual: delivery address and referral cover sheet show different house numbers.

That output is useful because it tells the coordinator exactly where to look: the missing date of birth, the unclear signature and the address clash. The coordinator opens the original scan beside it and confirms or corrects each field, and only then marks the record complete.

Check the extraction before you rely on it

Run 50 past referrals, with names redacted if you are testing in a sandbox, through the process and have a coordinator mark each field right or wrong. Do this separately for your five most frequent referrers, because their forms and fax quality differ. If a field is wrong more than a handful of times in 50, do not automate that field for that referrer; keep it manual. Fields with the highest cost of error, such as date of birth, member number and quantity, always get a human check regardless.

Job two: chasing missing paperwork without nagging

A missing signature or note is the commonest reason for delay, and chasing it is an ideal job for drafted, scheduled reminders. The aim is a short, specific request, sent at a sensible cadence, with every attempt logged so a person can step in. The tutorial on chasing missing records covers the general method; here is how it fits a supplier.

The chaser table

Cadence is illustrative. Set your own, and respect any limits on how often you contact an office. Secure channels matter here: a message that names a patient and an item is health information, so use the channel your privacy rules allow, and keep identifiers to the minimum needed.

Drafting the chaser

Give the model the intake record's missing-items list and nothing else, and ask it for a message in a fixed template. A good chaser is short, polite and exact. Here is an illustrative draft for a missing signature:

Hello [first name], we received your referral for [item] for [patient initials and date of birth]. To go ahead we still need: (1) the order signed and dated by the prescriber; (2) the clinic note from the visit on or around [date]. You can reply to this message with both attached. If anything here looks wrong, call us on [number]. Thank you, [your name], [supplier name].

What would you fix in this? Check that it does not contain more patient detail than necessary, that "(2) the clinic note from the visit" matches what your payer list actually requires for that item, and that the phone number goes to a person. The model can tidy wording, but the list of what is required comes from your checklist, never from the model's memory of what suppliers "usually" need.

Before and after: a voicemail that gives nothing away

A first draft of a delivery voicemail read: "Hello, this is [supplier] calling about your hospital bed and oxygen concentrator delivery on Thursday." Anyone who overhears the message learns about the patient's health. The corrected version says: "Hello [first name], this is [supplier name] with a message about your delivery on Thursday. Please call us on [number] or reply to our text to confirm the time." It names no equipment and still gives the patient everything needed to act.

The chaser that asked for the wrong document

In this illustrative case, a supplier let a model draft chasers from a generic prompt ("ask the doctor's office for the missing paperwork") and the drafts asked for a document the payer did not require for that item. Referrers sent it, the coordinator filed it, and the real gap, an undated order, stayed open for another week. The fix was to generate every chaser from the record's own missing-items field, which a coordinator had confirmed, so the AI never decides what is missing.

Job three: delivery-window calls that get a yes

Delivery planning has the highest phone volume per order: confirm the address, offer a window, handle a reschedule, remind the day before. These are well-defined and repeat daily, which suits an automated caller or text assistant. They are also calls where a patient might mention a health concern, so the safety rules matter more than the script.

What the delivery assistant may do

• Introduce itself as an automated assistant from your business and offer a person at any point.
• Confirm the delivery address and who will be at home.
• Offer the windows your dispatch has released and book the one the patient picks.
• Reschedule within the same rules, and flag anything outside them.
• Send a reminder the day before with the window, what to prepare (clear space, access) and a number to call.

What it must never do

• Give advice about how to use equipment, adjust it or whether it is suitable.
• Answer questions about symptoms or recovery.
• Discuss coverage, cost or what the patient owes beyond a scripted line that a person will call.
• Continue a call where the patient sounds distressed, confused or mentions a fall, pain or an emergency. It transfers to a person, or, if none is available, tells the patient to contact emergency services where appropriate, using wording your clinical lead approved.

A sample delivery call script

Hello, this is the automated assistant for [supplier name], calling for
[first name] about the delivery of your [item category]. This call may be
recorded. I can confirm a time, or I can put you through to a team member
at any point. Is now a good time to talk?

[if yes] Your delivery is ready to schedule. We have Thursday between
9 and 12, or Friday between 1 and 4. Which would you prefer?

[after choice] Thank you. We'll deliver to [address]. [Your receiving
rule, e.g. an adult must be home to sign.] Is there anything about access,
such as stairs or a gate code, we should tell the driver?

[at any mention of pain, a fall, equipment trouble, or a request for
advice] I'm going to connect you to a team member now.

The patient's replies are logged in the order record as free text, and anything flagged is read by a person the same day. If a patient does not answer, the assistant leaves a short voicemail that avoids naming the equipment, and tries a text by the patient's preferred method.

Sixty referrals a week: one supplier's rollout

Illustrate with a small supplier handling about 60 new referrals a week, with three intake coordinators and one dispatcher. Everything below is an illustrative scenario with round numbers.

1. Baseline. A review of the last 30 slow referrals shows 18 waited mainly on missing signatures or notes, 7 on authorisation and 5 on unreachable patients. The average referral that needed a chase took about nine working days to reach a delivery date.
2. Intake record. The billing manager writes the master checklist, one page per item category, from the payers' own documentation lists. Coordinators start filling the standard record by hand and spend two weeks refining it.
3. Extraction pilot. The team tests extraction on 50 old referrals. Date of birth and quantity prove reliable for the four biggest referrers; handwriting-heavy forms from a fifth are wrong too often, so that referrer stays manual. Coordinators keep checking every field against the scan.
4. Chasers. Chasers are generated from the confirmed missing-items field and sent on the cadence above. A coordinator approves each batch before it goes out in the first month; later, routine day-2 reminders go automatically while day-4 calls stay human.
5. Delivery calls. Confirmation calls run for stock items first, such as basic mobility aids. Anything requiring set-up by a technician stays with staff.
6. Result to track. The team compares days from referral to delivery date for chased referrals, the number of chasers that had to be corrected, and the number of calls transferred to staff. They treat any rise in corrections as a reason to pause, not push on.

Even a modest improvement matters here, because each day a patient waits is a day without equipment they need. But the supplier makes no promise about how much faster things become; they measure it.

Human sign-off: where a person must press the button

Define sign-off points in writing. Put them in a table on the wall.

Keep a record of each sign-off in the order system. If your software cannot hold an approval field, add one before adding automation, because an automated step without an audit trail is hard to defend later.

Privacy and vendor checks before any patient data moves

Referrals and delivery calls are full of health information. Before connecting an AI tool, put the vendor through a short check, and get your compliance adviser's view on the rules that apply to you.

1. Where are documents, audio and transcripts stored, and for how long?
2. Is any of it used to train models? The business plans of ChatGPT, Claude and Gemini in Google Workspace say business data is not used for training by default; get the same answer in writing from any intake, fax or phone vendor.
3. Who at the vendor can access it?
4. Will the vendor sign the data-handling agreement your health-privacy rules require?
5. Can you delete records on request and export them if you leave?
6. What happens to a call recording if a patient asks you to delete it?

The privacy-terms tutorial gives the longer list. If a vendor cannot answer these clearly, that tells you something. Start with tools you already use for patient data, and only add new ones that pass.

Sizing the cost against coordinator hours

Costs vary widely by tool and call volume, so I will not quote prices; each vendor's current plans are on its own pricing page. The main expense is usually staff time to write the checklist, test extraction and review outputs, rather than the subscription. A reasonable way to size it: count the hours your coordinators spend each week on chasing and delivery confirmation, estimate what fraction is routine, and compare the tool's monthly cost against that, adding review time. If the sums only work with optimistic assumptions, wait.

Start small, with one channel and one item category, run it quietly first, and widen only when the numbers hold. The shadow-mode pilot approach is made for this: the AI drafts, staff compare, and nothing reaches a patient until the drafts match what a good coordinator would send.

Ten things to check each week

• How many referrals were marked complete by a coordinator after AI extraction, and how many fields were corrected.
• Which fields are most often wrong, and for which referrer.
• How many chasers went out, and how many produced a reply within two working days.
• How many chasers asked for the wrong item, and why.
• Delivery calls: completed, no answer, transferred to staff.
• Every transfer that involved a health comment, and whether the right person called back the same day.
• Any complaint from a patient or a referrer's office.
• Referral-to-delivery days for chased referrals against the baseline.
• Whether the payer documentation list was updated, and the checklist with it.
• Any message sent with more patient detail than the minimum.

If you are tidying records before adding any of this, clean up customer records first; extraction is only as good as the data it fills. And if you would like to map your own intake loop with someone, that is what my AI implementation consultation covers.

Questions equipment suppliers ask about AI in intake and delivery

Can AI read referrals that arrive as faxes or scanned PDFs?

Often, but accuracy varies with scan quality and layout, and handwriting is the weak spot. Treat the extraction as a draft that an intake coordinator confirms against the source document. Measure how often fields come out wrong for your most common senders before you rely on it, and keep the original image attached to every record.

Should AI decide whether a referral meets coverage rules?

No. Coverage rules differ by payer and change often. AI can check that listed documents are present and flag gaps, but a trained person decides whether an order is complete and compliant. Use the payer's own documentation list as the checklist, and update it whenever the payer changes its requirements.

Is an automated delivery-window call acceptable for patients?

For confirming and rescheduling windows, many patients accept it if it is clear, short and offers a person. Disclose that the caller is automated, check any rules on automated calls and texts that apply to you, and always transfer anyone who mentions equipment trouble, pain or an emergency.

What privacy checks come before connecting an AI tool to patient data?

Confirm where data is stored, whether it is used to train models, who can access it and whether the vendor will sign the agreement your health-privacy rules require. Take advice from your compliance officer. Do not put patient details into any tool that has not passed those checks.

How do we prove a human signed off?

Record who approved each referral and each outbound message, with a date and time, in your order system. The AI's draft and the person's final decision should both be kept. If your system cannot store that trail, add a simple approval field before you add any automation.

Further reads

• Medical Equipment Lease: Compare Quotes and Check Terms With AI — Compare equipment lease quotes and check terms with AI.
• Can AI Fill In Forms and Supplier Portals for You? — What AI can and cannot do with forms and portals.
• Cutting Parcel-Tracking Calls With AI Delivery Updates — Cut delivery-tracking calls with automatic updates.
• How Independent Pharmacies Use AI to Shorten Phone Queues — A health-supply business shortening its phone queues.
• AI Answering Service for Home Care Agencies: After-Hours Call-Outs — After-hours call rules for a care business.
• AI Governance for a Small Business: Who Decides, Approves, Checks — Who decides, approves and checks AI output in a small team.
• AI for an Equipment Rental Business: Setup, Cost and What to Automate — Setup and cost for an equipment business adding AI.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: a large public payer's order-requirements page (checked 4 November 2026, used only to confirm the typical order fields) and the business-data pages of OpenAI, Anthropic and Google Workspace. Payer documentation requirements vary and must be taken from each payer's own current guidance. No vendor prices are quoted.