Case study · Patient intake automation · Phase one live
The systems were fine. Everything between them was being done by hand.
A specialist physician practice with eighteen years of accumulated case records, a capable EHR, and a team that was still retyping the same information into three places. Patient intake ran on memory, email and a spreadsheet. The practice is not named here.
Phase one, measured
Running unattended in the practice's own Google Workspace since September 22, 2026. Counts read from its logs, October 4, 2026.
referral emails logged to a row, none missed
emails refiled automatically after a service pause, recovering 30 referrals
privacy-notice emails caught and set aside for a person
Phase two, the follow-ups and escalations, has run in shadow mode since September 25, 2026, writing only to copies of the practice's log until they approve it. These are coverage numbers, not time saved: nobody recorded a before, so I do not print one.
What the audit found
The practice was not disorganized, which is the usual assumption and the usual mistake. The EHR handled scheduling, records and reporting well. The billing software was genuinely good. The case log carried eighteen years of hard-won structure. The cost was in the gaps between those systems, where a person was doing the joining.
Referrals logged and chased from memory
Every referral entered by hand, one at a time, and followed up when somebody remembered. No single list of what was outstanding, so the chase depended on whoever was paying attention that week.
Email as the coordination layer
Messages scattered across departments, so the workaround had become copying everyone on everything. Things still slipped. The clearest example was a week lost because the one person who needed to see a message was the one person not copied.
No single at-a-glance view
Status lived in people’s heads and in a spreadsheet that had to be re-typed from what the EHR already knew. The same fact was being entered two and three times.
Reporting quietly wrong
The same physician’s name spelled four different ways across records, so any report grouping by provider was subtly incorrect. Nobody had noticed because the numbers looked plausible.
The last one is worth dwelling on. A data hygiene problem that nobody can see is more dangerous than a workflow problem everybody complains about, because decisions get made on the bad numbers with full confidence.
What was safe to automate, and what was not
Safe, and worth doing first
- +Routing incoming email to the right person automatically, so nothing depends on remembering to copy someone
- +Logging a referral in one action instead of by hand
- +Generating the weekly list of what is outstanding and who to chase
- +Producing a summary for a scheduled review, from records that already exist
- +Normalizing provider names so reporting stops being wrong
Off the table entirely
- −Anything patient-facing. No AI talking to patients, in any channel
- −Anything clinical. No triage, no prioritization by medical urgency, no interpretation of records
- −Any decision a clinician is professionally accountable for
- −Moving patient information anywhere it was not already held
The dividing line was simple enough to state in one sentence to the practice: the AI works in the back office only, sorting and summarizing and reminding. Never patients, never clinical decisions. A boundary a physician can repeat from memory is a boundary that survives contact with a busy week.
The finding that changed the plan
They could build the first phase without a new software vendor.
The default assumption going into a project like this is that new capability means new vendors, new vendor agreements, and a legal review that adds months. Checking rather than assuming turned up something better: the practice had held a suitable agreement with their existing platform provider since 2018, and the entire first phase could be built inside it.
That single finding reshaped the sequencing. The work that needed no new vendor, no new system for staff to learn, and no patient data moving anywhere new became phase one. The only new paperwork was between the practice and me: a services agreement and a business associate agreement, signed before phase one touched real mail. The work that would require a separate agreement, covering the more advanced querying, was pushed to a later phase to be scoped only once the earlier work had proved itself.
This is the sort of thing an audit is for. It is not a technical insight and it is not clever. It is just the result of reading the existing agreements before designing the system rather than after, and it removed the slowest step from the project.
The plan that came out of it
Three phases, each one gated on the last having earned its place. The practice can stop after any of them and still be better off than before. Here is where each one stands today.
Running since September 22, 2026. Copies of mail from the intake and records inboxes land in one account inside the practice's own Workspace. Every 15 minutes, referrals, records requests and surgery mail are logged to the practice's sheet, one row per email thread, so a firm's "any update?" reply updates the row instead of adding a new one. Mail the rules cannot place is left for a person. A Monday digest lists what is still open. Routing each message to a named person switches on once the practice sets who handles what.
Reading the body and attachments of an email, not just its subject line. Referrals with no appointment get follow-up nudges, and a firm that keeps asking is escalated. A reductions flag stays up until someone has checked the balance and signed off. Records requests of both kinds are tracked until they close, and signed HIPAA forms are filed. Built and installed on September 25, 2026 in shadow mode: it writes what it would log into copies of the practice's log tabs, and records what it would file without filing anything, so the practice can compare it with their own log before it touches anything real. It is not switched on to run by itself yet.
A link to the EHR so a new patient entered there appears on the log without being typed twice, plus a provider roster that keeps the rest current. Scoped last deliberately: a test connection to the EHR vendor's sandbox works, but the practice's live system is not connected, and nothing is promised until the vendor opens it up.
What this shows about the diagnosis phase
The answer is usually smaller than the pitch
Nothing in phase one is technically impressive. A routing inbox and a tracker are not the future of medicine. They are what would actually give this practice its week back, and an audit that recommended something grander would have been serving me rather than them.
Read the agreements before designing the system
The most valuable finding was that they could build inside what they already had. That came from reading paperwork, not from technical work, and it removed months from the timeline.
Name what you will not automate, early and plainly
Stating the boundary in one memorable sentence did more for confidence than any capability demonstration. In regulated work, the limits are the product.
Find the errors nobody is complaining about
The provider-name inconsistency was not on anyone’s list of problems. It was quietly corrupting every report. Those are the findings worth paying for, because by definition nobody was going to ask for them.
Further reading
medical practice automation: where the audit trail starts
accuracy is the wrong first question for medical practice automation. here's why the audit trail has to come first, and what it actually looks like.
medical billing automation without the handover of decisions
medical billing automation can speed up claims without letting a model decide them. the fix is splitting extraction from decision, with a trail either way.
why an audit trail matters more than accuracy
a 99% accurate model that can't show its work is a bigger liability than a slower system you can audit line by line. here's the actual split.
This is what the audit produces.
A map of where the manual work is, a clear line around what should stay human, a compliance answer, and a phased plan you can stop after at any point.
See the service: healthcare workflow automation for medical practices.
Questions people ask
Is this a finished project?
No. Phase one has run unattended in the practice's own Google Workspace since September 22, 2026, logging referrals, records requests and surgery mail as it arrives. Phase two, the follow-ups, was built and installed in shadow mode on September 25, 2026: it writes what it would log into copies of the practice's log tabs, and nothing touches the real log until the practice has compared the two and approved it. The diagnosis is still the center of this page, because it is the part most people have never seen and it is exactly what the Workflow Audit produces.
Does phase one use AI?
No. Phase one sorts mail with plain rules the practice can read, and it makes no calls to any outside service. That is deliberate: the first job was to stop things slipping, and rules do that without a model in the loop. If a later phase adds a model, it will only pull facts out of an email. The rules will still decide, and a person will still make every call.
Is this patient intake automation?
It is the audit that comes before it. What the practice had was a patient intake path running on manual effort: referrals entered one at a time and chased when somebody remembered, email as the coordination layer between departments, and status living in people's heads plus a spreadsheet retyped from what the EHR already knew. That is the process patient intake automation is meant to replace, and mapping it properly is what tells you which parts of it should be automated at all. The first phase has since been built and is running inside the practice's own Workspace.
What do the numbers measure?
Coverage and accuracy, read from the practice's own logs on October 4, 2026: whether every referral email got a row (95 of 95) and what happened after a service pause (1,826 emails refiled automatically, recovering 30 referrals). They do not measure hours saved. Nobody recorded a before, so a time-saved figure would be a guess, and this page does not print one.
What was ruled out, and why?
Anything patient-facing, anything clinical, any decision a clinician is professionally accountable for, and moving patient information anywhere it was not already held. Those were off the table from the start rather than negotiated away later, because automating a clinical judgment is not a feature, it is a liability.
Did the practice have to buy new software?
No, and that was the most useful finding in the audit. The assumption going in was that new capability means new vendors, new agreements, and a legal review that adds months. Checking rather than assuming turned up a suitable agreement the practice had already held with its existing platform provider since 2018, and the entire first phase is built inside it. The only new paperwork was my own: a services agreement and a business associate agreement between the practice and me.
What does an audit like this cost?
The Workflow Audit is $2,500 and takes about two weeks. If you go on to a build within ninety days, the full fee comes off the build price.