Every AI idea we're considering, building, or have ruled out

A signed download that walks a new clinic through installing blueHealthcare on its own machine. Setup itself needs no model at all; an optional AI layer can answer setup questions, from the bundled manuals only.

Note: off until two people at the clinic turn it on; suggests only; a named person confirms every step; every answer shows its manual source; never patient data, because none exists yet at setup.

A clinician types what they mean; Blue ranks real codes from the clinic's own loaded terminology; the clinician picks.

Note: the first working slice of Blue — every candidate is checked against the real terminology, so nothing invented can reach the clinician's screen. Suggests only; a named person decides; sources shown; off by default.

Turns a clinic's own procedure, medication, lab and test-menu codes into candidate matches against SNOMED, ICD and LOINC, for the clinic's own reviewer to confirm before go-live.

Note: named first, independently, by every session that looked at it — most clinics outside the US cannot code at all without this.

Points Blue at another system's spreadsheet export and it proposes which column is which.

Note: Blue proposes a rule, never a row value — the distinction that keeps this permitted.

For the clinic that has its old data and no documentation for what any of it means.

Note: the migration case for a clinic that inherited data nobody left an explanation for.

Reads a scanned paper register and proposes rows of data; nothing becomes live until a person clears it.

Note: every proposed row stays unconfirmed and not live until a person checks it against the page.

Reads a rejected HL7 message and says, in plain language, what the sending system got wrong.

Note: being wrong here costs an afternoon of re-reading a spec, never a patient.

Suggests how a foreign system's message fields line up with ours, for an analyst to confirm.

Note: only admitted with a test attached — a mapping that can't be checked can't be accepted.

Answers a staff question from our own manuals, citing the page it came from.

Note: no IT staff, no vendor within a day's travel — this is what keeps a clinic on the software. Suggests only; a named person decides; sources shown; off by default.

Same idea, over the clinic's own runbook, for a site with no connected help desk.

Note: runs over a bundle that carries no patient data at all.

Walks a new hire through the software in a practice environment, never the real record.

Note: the cheapest way to onboard a new clinician where there is nobody experienced to ask.

Turns "adult clinic Tuesday and Thursday mornings, twenty-minute slots" into a template a person edits and saves.

Note: no patient data, and it shortens onboarding for a clinic with nobody to build schedules by hand.

Turns a manager's plain-language staffing rule into something the scheduler's own solver can use.

Note: the solver stays the product; the resulting roster stays fully explainable.

"Her potassium results since March" becomes a structured search of a patient's record.

Note: Blue writes only the query — the clinic's existing access rules, unchanged, decide what comes back.

Machine translation shown beside the original, for a clinician who doesn't share a language with a colleague's note — with numbers, doses and drug names kept out of the model's reach entirely.

Note: the most valuable idea on this whole list outside the United States, and effectively unusable inside it under current interpreter requirements.

When someone cannot open a patient's chart, explains why — "you're not on their care team", "ask the attending to add you", "record a coverage period" — with the next step.

Note: reduces support load; today clinics share logins to get around unclear denials.

Would have read an incoming fax or scan and routed it to the right patient automatically.

Reason: deciding which patient a document belongs to is patient-identity matching, which is never allowed at any setting. This was briefly on a build list and should not have been — sorting a document by kind (a lab report, a referral) for a person to then attach themselves is a separate, still-open idea.

Helps a reviewer draft a coverage rule for blueChart and checks it against past cases before it is adopted.

Note: adopted by edition, never by a single run, so a stranger can re-run the same edition later and get the same answer.

Drafts a companion explanation of a policy for blueClaims and checks that every statement in it traces back to the source document.

Note: no personal data anywhere in this one.

Would highlight the exact sentence in a record that a coverage or claims verdict rests on, instead of only naming the document it came from.

Note: blocked until the record's own evidence-anchoring reaches individual sentences rather than whole documents.

Reads a trial's eligibility criteria against a clinic's patients and proposes candidates for a research coordinator to screen.

Note: named as a permitted shape from the start; waiting on the record being able to say a program was involved before it ships.

Points to the calibration dates on a certificate PDF so a person can confirm what was read rather than retyping three dates per instrument, with every date quoted from the certificate.

Note: the single most-retyped document in the lab software, and one of the most common hold-causing typos. Suggests only; a named person decides; sources shown; off by default.

Runs known, planted errors through a clinic's own human-plus-model review process and reports the real catch rate — recomputed whenever the model changes.

Note: would let a clinic decide whether to trust an AI feature by its own measured results, not by taking our word for it.

Pulls together what a clinic needs to show an outside inspector.

Note: no patient data and no clinician approval needed — a one-doctor clinic's actual, ordinary problem.

Turns a donor's or ministry's reporting form into a specification that computes every number itself.

Note: the specification produces every number, never the model, and it is proven against a real example before it is trusted.

Reads a contract and proposes its renewal date and notice deadline for a person to confirm.

Note: the auto-renewed contract nobody read is the most common avoidable loss at a small organization. Suggests only; a named person decides; sources shown; off by default.

Turns an invoice's line items into structured data a computer can check.

Note: Blue extracts the numbers; ordinary arithmetic still decides whether anything is actually wrong. Suggests only; a named person decides; sources shown; off by default.

Identifies re-submitted invoices or re-scanned pages that a clinic might otherwise pay twice.

Note: a common source of waste at a small clinic, especially during migration from paper.

Highlights an invoice line that does not match this vendor's usual pricing or volume patterns, for a staff member to screen before payment.

Note: a warning only — arithmetic and a manager's approval still decide whether anything is actually wrong.

Reviews a clinic's own signed record against its own charges and flags a charge that looks supportable but was never billed.

Note: chosen over similar ideas because its only possible mistake is billing less, never more.

When a patient writes in their language, a clinician sees a translation beside the original, so they can act on the message at all.

Note: runs only where clinic law allows; the original is always shown; routing stays identical to every other message.

Would have summarized an incoming patient portal message for the clinician reading it.

Reason: in tension with the rule against a model doing any triage — deciding what part of a message is worth summarizing is itself a triage judgment.

A clinician types "chest film" or "sugar test" or the local-language name and sees candidate order items to pick from, instead of scrolling through the full catalogue.

Note: candidates come only from the real catalogue; the human still reads and clicks the real item.

The model would point at a region of a photographed paper form or a device's screen; a person reads that region and types the number themselves.

Note: the single idea that would most change what a paper-based clinic can do — waiting on a ruling about whether a pointer counts as a value.

Would have let a model compile, de-duplicate or reformat a patient's medication list, even from data the clinic already held in structured form.

Reason: ruled out outright, with no exception — compiling a medication list is itself a clinical judgment, even when nothing has to be read from free text.

The stored bytes of a scanned document get a derived text layer for search and display beside the image, keeping the original scan as the record.

Note: stored bytes are already retained; this layer helps a clinic clerk who otherwise re-keys a 6-page discharge summary.

A clinician speaks; speech-to-text fills a note's section; the original audio is kept alongside it.

Note: the named exception that would decide whether a clinician can dictate at all instead of typing.

Would quote a note's own words as a candidate entry for a patient's problem list, for a clinician to accept or reject.

Note: held for review, not being built right now. A medication-list version of the same idea is decided against outright and does not survive in any form.

Would have had a model write a clinical note's narrative sections from a clinician's description or a conversation.

Reason: rejected on clinical advice. The structured intake questionnaire built instead needs no model at all and delivers most of the same benefit.

Would have had a model compare a signed note against a patient's medication and allergy lists and propose updates to them.

Reason: its medication and allergy halves are ruled out outright by the same clinical decision that ended model-assembled medication lists; the version described here does not proceed.