AI for Patient Intake in Pharma

A referral form arrives in the study coordinator's mailbox late on a Friday. A consultant at a tertiary centre has written three lines about a patient who might suit the trial, attached a clinic letter, and asked whether the site can see them next week. The protocol it is read against runs to well over a hundred pages and carries a long list of inclusion and exclusion criteria, several of which depend on laboratory values taken inside a defined window, on prior lines of therapy in a particular order, and on a washout period counted from a date nobody has written down.

Nobody at the referring clinic calls that intake. On the sponsor side it is the front end of enrolment, and the same shape repeats in patient support, where the form comes from a prescriber's office and eligibility turns on diagnosis coding, prior therapy and reimbursement status rather than a protocol.

Unrelated on the org chart, near identical in practice

Clinical trial recruitment and access programme enrolment look unrelated on an organisation chart and behave almost identically. Both begin with an unstructured arrival: a referral, a self referral from a campaign, a fax from a prescriber. Both assess it against criteria that are long, conditional and partly ambiguous. Both produce a record whose quality is tested only when something goes wrong, usually years later and in front of someone else.

What differs is the governing artefact. Recruitment answers to the protocol, the approved consent form and the screening log. Access answers to the prescriber enrolment form, the benefits verification outcome and the case intake record. An agent built for one transfers to the other with the rule set swapped rather than rebuilt.

Reading the criteria as written rather than as remembered

Criteria live in the protocol. Day to day they live in a summary sheet somebody produced at study start up, and that sheet drifts. An amendment changes a laboratory threshold, the protocol is updated, the sheet on the coordinator's desk is not.

An agent can hold the criteria as discrete checkable questions derived from the protocol text, each carrying its section and amendment version. For a candidate it answers what the documents support, states what it cannot answer, and names the source needed. The output is a gap list rather than a verdict, which is the more useful half, because the assessment was never the bottleneck.

Pre screening against records you do not control

The data that answers most criteria sits in a hospital record system the sponsor has no access to, across clinic letters, discharge summaries, pathology reports and laboratory portals, many of them scanned. Pre screening is therefore reading, not querying.

Given the referral packet, an agent can extract candidate values with a pointer to where each was found, note the date each was taken, and compute the derived items people get wrong by hand: whether a value falls inside the required window, how many prior lines of therapy are documented, whether a concomitant medication appears on the prohibited list, when a washout actually started.

The extracted value is a reading aid, never a data entry. Nothing reaches the data capture system without a person confirming it against the source document.

The screen failure that was decided before the visit

Screen failures get discussed as if they were biology. A meaningful share of them are reading errors. A criterion was read loosely at site, a value was accepted from outside its window, prior therapy was counted in a way the protocol does not support, or a date came from the wrong letter. The patient came in, consented, had bloods drawn, and failed on something visible in the referral.

Two agent tasks follow. First, check the hard criteria before the screening visit is booked, so avoidable failures are caught while they are cheap and before anybody has been consented. Second, read across the screening log and report which criteria drive failures at which sites. That second output is an input to a protocol amendment conversation, and it stays invisible while failure reasons sit in free text.

Consent version control belongs to the data

Informed consent is a conversation, and it is also a document with a version, an approval date per site, translations, and local amendments. A participant consented on a superseded version is a deviation. Found late and in a batch, it becomes a finding with a remediation plan.

An agent can maintain the register of which version was approved for which site and language and from what date, then check each signed form against the version effective on the date it was signed. It can also run the completeness checks that catch most queries: every required page present, dates written by the participant rather than back filled, the signature of the person who conducted the discussion. Where an amendment requires re consent, it can list the active participants affected and who is due at their next visit.

Feasibility answers and the forecast stacked on them

Site feasibility questionnaires are completed from memory, often optimistically, often by someone who will not run the study day to day. Enrolment forecasts are built on those answers, and the gap shows up at the point where it costs the most.

An agent can pre fill the questionnaire from what the site actually did on comparable protocols, flag answers inconsistent with that history, and extract from the protocol the criteria likely to limit that site's population. The forecast stays a judgement, because the variables that break it are relationships, competing studies and staffing, none of which appear in the questionnaire.

Prescriber enrolment forms, arriving half finished

Patient support and access intake has a volume problem rather than an ambiguity problem. Forms arrive by fax and portal with a missing prescriber signature, an absent diagnosis code, an unticked consent box, a blank date, or handwriting that will not resolve. Each gap costs a round trip to a busy office.

An agent checking completeness at the door changes the economics. It validates the form against what the programme requires, drafts the specific request for the missing item rather than a generic notice, and catches the duplicate where the same patient was submitted by two prescribers under different spellings.

What eligibility turns on once a payer is involved

Access eligibility is a different calculation. It rests on the diagnosis code as submitted, documented prior therapy and the intolerance or failure that justifies the request, the payer and plan, the coverage position, and for free product the income attestation.

An agent can assemble that position with each element tied to the document that supports it, and flag the mismatch that causes the most downstream trouble: a diagnosis code on the enrolment form that the attached clinical note does not support. Raising that before submission is worth far more than correcting it after a denial.

Benefits verification and the waiting that follows

Benefits verification is a sequence of enquiries whose answers arrive at unpredictable intervals, often followed by a prior authorisation whose requirements sit in a payer policy document.

An agent can read that policy and list what this payer wants for this indication, assemble the pack from the case record, flag what is absent, and keep the pending queue honest by knowing which enquiries have gone unanswered beyond their usual turnaround. Submitting and appealing stay with the team.

The adverse event that arrives mid conversation

A caller asking about their shipment mentions a rash. A coordinator confirming a screening appointment hears about a hospital admission. The reporting clock starts when the company becomes aware, and awareness includes an agent operating on the company's behalf. This is the part of intake nobody can afford to automate carelessly.

The design that works is asymmetric. The agent is tuned for recall, not precision: it flags every candidate mention of an event, a special situation, a product complaint or a pregnancy, preserves the verbatim text, captures when awareness occurred, and routes it to pharmacovigilance for a trained person to assess. It is never permitted to conclude that something was not an event, to classify seriousness or to assess causality. A false positive costs a triager minutes. A suppressed case is a regulatory failure, and that asymmetry should be built into the system rather than left to a threshold.

The log that has to survive an inspection

The screening log must account for everyone screened, the outcome, the reason for each screen failure and the dates, each entry traceable to source. In practice it is maintained in arrears, and the enrolment tracker, the consent log and the data capture system quietly diverge from it.

Keeping those four in agreement suits an agent and defeats people. Run the comparison on a schedule and surface the specific disagreements: a participant consented with no screening log entry, a screen failure with no reason recorded, a date in the tracker that does not match the signed form. Reconciling continuously costs a fraction of reconstructing it the week before a visit.

Where AI fails here

Eligibility determination is the investigator's medical judgement and cannot be delegated. An agent can show that every criterion has been addressed and where the evidence sits. The decision to enrol a specific person into a specific protocol belongs to a named clinician accountable for it.

Consent is not a form check. The conversation, the chance to ask questions and the assessment of capacity are the substance. Automating around the document while treating the discussion as administrative gets the priority backwards.

Safety assessment stays human. Seriousness, expectedness, causality and the narrative are trained judgements, and pharmacovigilance intake carries validation expectations that detection assistance does not satisfy.

The agent cannot reconcile contradictory source documents. Where the clinic letter and the laboratory report disagree, it will present both and should not choose. Nor can it reach a health system record it has no lawful basis to see, fix a protocol whose criteria cannot be answered from routinely collected data, or take responsibility for telling a patient they do not qualify.

Where to start

Pick one protocol at a small number of sites and scope the agent narrowly: read the referral packet, answer what the documents support, produce the gap list with source pointers. Leave the eligibility call and the data entry alone. You will soon know what share of referrals fail on something visible before the visit, and that figure is a better argument for change than any projection.

Consent version checking usually follows, because the rules are objective and the findings it prevents are expensive. On the access side, start with form completeness and duplicate detection. Safety detection comes last, after pharmacovigilance has set the recall bar.

Adopt AI Partners builds these agents inside the systems clinical operations and patient access already use, so a pre screening summary lands on the screening record rather than in another tool. If it would help to talk it through, describe how referrals reach your team to hello@adoptaipartners.com.