AI for Collections in Healthcare
A patient opens a statement for eight hundred pounds and rings the number printed on it. The balance is real in the sense that the ledger holds it. Whether they owe it is a different question, and the person answering the phone usually cannot tell inside the first minute. The claim may still be open with a secondary payer. The primary may have adjudicated against a plan the patient left in March. A denial may have been posted to patient responsibility because nobody worked it in time.
Healthcare collections goes wrong there, not at the point of chasing. In most industries a receivable is presumed correct and the work is recovery. Here a large share of patient balances is a billing artefact, and chasing it produces complaints and refunds rather than cash.
Two books sharing one ageing report
Payer accounts receivable and patient responsibility are separate trades wearing the same report. Payer receivable is institutional: reason codes, filing windows, contract terms. Patient responsibility is a household problem: a person, a statement, an income, and often an illness that is still ongoing.
Run them in one queue and attention follows the largest balances, nearly always payer balances, while the patient book ages quietly towards agency placement.
The first useful thing an agent does is keep the two separated at line level, track every movement of value from payer responsibility to patient responsibility, and record the remittance that caused each move. Few organisations hold that mapping in a form anyone can query.
The first question is whether the money is the patient's at all
Several checks can be run mechanically before a statement is produced, and most teams run them only after a patient complains. Was the claim adjudicated at all, or is this an unbilled charge that drifted to self pay. Does the patient responsibility figure reconcile to the explanation of benefits, line by line. Is there a secondary plan that was never billed. Is the balance in fact a denial posted to the patient, which is a billing failure and not patient debt. Was the episode covered by a different scheme entirely: a motor accident, an employer liability claim, a research protocol, or a bundled arrangement the charge should never have left.
An agent can hold that gate on every account. The output is a hold list with a stated reason per account, routed to whoever fixes that category, not a tick that releases the batch. Expect the self pay book to shrink the first time you run it, and warn finance before it does.
Coordination of benefits is the quiet manufacturer of false patient debt
More patient balances are created by coverage errors than by patients refusing to pay. A spouse plan and an employer plan billed in the wrong order. Two parents' plans sequenced by a rule nobody at registration applies. Retroactive eligibility granted months after the episode. A payer suspending every claim for a patient because a coordination questionnaire went out in the payer's own vocabulary and never came back.
These leave a signature: a zero payment with a coordination reason code, a denial pointing at other coverage, claims for the same patient paid by a different payer in the same period, an eligibility response naming a plan the account does not hold.
An agent can assemble a coverage timeline per patient across the episode, compare it against what was billed and in what sequence, and propose the correction with the evidence attached. Where the missing piece is that questionnaire, it drafts a short request in ordinary language explaining what is asked and why it matters to them, which is the version that gets returned. A person sends it.
The statement is the collection instrument
A great deal of collections effort is spent recovering from a statement nobody could read. Three arrive for one operation because the hospital, the anaesthetist and the pathology service bill separately. Each lists charges, adjustments, payments and a balance in the order the ledger keeps them, in adjudication vocabulary, and the explanation of benefits the patient already received states a different number again.
An agent is well suited to producing the version a person can act on: a short narrative per episode covering what was done, what the insurer paid, what the plan left with them and why, and what is still open with the payer and therefore not yet theirs to settle. Where the systems allow, consolidate across billing entities so one episode produces one account of itself.
Make the agent compare the statement against the explanation of benefits and hold anything that disagrees. Clarity is not a courtesy here. It is the highest yielding intervention in the patient book, because most non payment is confusion rather than refusal.
Financial assistance screening belongs before collections, not after it
Most patients who qualify for assistance never apply, because the application asks for documents at the point in someone's life when they are least able to assemble them. The result is a household paying instalments it cannot afford on a balance it should never have carried.
Obligations differ by country and organisation type, and many providers must offer an assistance policy and let an interval pass before certain recovery actions. Treat this as the shape of the problem and involve compliance early.
An agent can screen the self pay book against your own published criteria and produce a shortlist where the supporting evidence is already gathered, so a counsellor offers assistance rather than waiting for the patient to discover it exists. It can pre fill the application and name the two documents missing instead of listing nine.
Presumptive determinations, and which way to err
Where your policy allows a determination without a full application, an agent can propose one from the evidence available; the counsellor decides. Set the tolerance deliberately, because the two errors are not equal: a discount granted to someone who did not strictly qualify costs money, while one refused to someone who did sends a sick household into a recovery process. Build the screening so it can only recommend help, never withhold it, and keep it structurally separate from anything that scores accounts for pursuit.
Payment plans for balances that keep arriving
Healthcare debt is episodic, so it does not behave like consumer instalment debt. A patient on a plan for a March admission has a June procedure, and the standard system response is a second plan with a second payment instruction and a second way to fail. Plans should absorb new balances and re amortise, not multiply. Dividing a balance by twelve is a target rather than an affordability assessment, and where treatment is ongoing the plan should be built knowing more balances are coming.
An agent is good at the surrounding work: watching for quiet failure such as an expired card, noticing the first missed instalment rather than the third, gathering what changed before anyone contacts the patient, and preparing a revision for approval. Keep the schedule arithmetic in ordinary code rather than in the model, since a wrong subtotal produces a plan that breaks by month two.
Contact, consent and the ethics of chasing someone who is ill
This is where automation needs its hardest limits.
A balance message carries clinical information by implication. A voicemail or a letter about a hospital account discloses to a household that someone received care, and for some service lines that disclosure causes real harm. Channel and wording follow the minimum necessary principle, and the agent needs no sight of the clinical detail behind a charge in order to write about the charge.
Certain conditions must act as stops rather than weights in a score: a patient recorded as deceased, a patient in active treatment for a serious condition, balances attached to sensitive services, and accounts with an open dispute or an assistance application under review. A stop takes the account out of every automated path and routes it to a trained person, and no balance is large enough to outvote it.
No automated outbound contact should reach a patient without a person approving it. That is not a caution for the first phase of a deployment but the permanent design.
One boundary is specific to this sector: collections must have no influence over care. The agent needs no write access to scheduling or the clinical record, and should be incapable of flagging a patient in any system a clinician sees.
Deceased patients and estates
A death recorded in one system has to stop dunning in all of them immediately, and that reconciliation is work an agent should own. The failure mode is a chaser arriving at a bereaved household. Estate recovery, where it applies, is a separate path with a human owner.
Where AI fails here
Identity and coverage data defeats it first. One patient under three record numbers, a guarantor record nobody updated, a plan captured correctly and then overwritten. Everything downstream inherits the error and states it confidently.
Hardship is not detectable from a ledger. A model can see that someone stopped paying, not why, and it should not be asked to guess. That judgement belongs to a person who can ask.
Readable and wrong is the dangerous output. A clear explanation containing an incorrect figure is worse than a confusing accurate one, because the patient believes it and acts on it. Bind every number in a patient facing draft to a value read at the moment of drafting.
An agent will not repair a weak front end either: if coverage capture at registration is unreliable, automation produces wrong statements faster and in better prose.
Measuring it, and where to start
Most reporting here counts activity: statements issued, accounts touched, calls made. Measure what shows the book is honest instead. Patient balances resolved with no contact needed. The share of statements that generate an inbound query. Complaints and refunds per thousand statements. Assistance applications started against those completed. Plans alive at six months. Bad debt placement volume, which should fall.
Do not set recovery per account as the objective. Optimise for that and the system learns to press hardest where pressure works, which here means households already paying for care they are still receiving.
Start with pre statement verification, including the coordination of benefits work. It touches no patient, it has a checkable right answer, and it removes debt that should never have been pursued. Statement explanation into a review queue is a strong second, and assistance screening follows. Outbound contact comes last, if at all.
If your self pay book is growing and nobody can say how much of it is genuinely patient money, describe the process to Adopt AI Partners and we will tell you which parts an agent should carry and which it should never touch.