Case Study: Replacing a Clinic's Paper Workflow With a Decision System
Article Hossein Narimani November 18, 2026 Operational Intelligence

Case Study: Replacing a Clinic's Paper Workflow With a Decision System

The clinic didn't need software. It needed to know who owned a patient between two visits. Note: this case study is a composite of several similar engagements in healthcare operations, with client details removed....

The clinic didn't need software. It needed to know who owned a patient between two visits.

Note: this case study is a composite of several similar engagements in healthcare operations, with client details removed. The pattern and the figures represent what recurs in this class of project.

Context and the constraint we could not change

A single-site clinic with six physicians and four administrative staff. Around 90 patients a day. Paper records, telephone scheduling, and post-visit follow-up that depended on a receptionist's handwritten notes.

One constraint was clear from the start and non-negotiable: the physicians would not interact with any system during a consultation. Any solution that added data entry to the consultation minute was rejected.

That constraint looked like an obstacle at first. In practice it improved the design, because it forced us to answer which data was genuinely required.

What we measured before touching anything

Two weeks of observation with no changes. Three numbers recorded:

  • Return rate for follow-up visits: of patients where the physician recommended a second visit, how many came back.
  • Staff time spent locating information: every occasion someone hunted for a file or a note.
  • No-show rate: the easiest number to obtain and, as it turned out, the least connected to the real problem.

The first number was the shock: roughly 41 percent of patients who should have returned had not. The clinic did not know this, because nowhere was it recorded who should return.

That is the central point of the whole project: the problem was invisible because no data existed to see it with.

The architecture decision: system of record before automation

The obvious temptation was building automated reminder messaging. Built first, it would have failed — because no source existed to say who should be reminded.

The correct order was:

  1. One record of state per patient. Not a medical record — operational state only: awaiting follow-up, followed up, closed.
  2. One owner per state. Every patient in "awaiting follow-up" carried a staff member's name.
  3. One deadline. Follow-up within four working days of the visit.
  4. Only then, automation.

What we deliberately did not automate

Three things stayed manual, each a deliberate decision:

The content of the follow-up call. No templated message. Staff had a framework and spoke themselves. The reason was simple: response rates to automated messaging in this domain are poor and it costs trust.

Deciding who needs follow-up. That is a clinical decision and stayed out of the system. The physician ticked a box, and that was all.

Exceptions. Any patient outside the normal pattern left the automated flow and landed in a human queue.

In total, two things were automated: building the daily follow-up list, and raising an alert when the four-day deadline passed. The entire automated logic was under fifty lines.

The resistance we did not expect

The initial assumption was that resistance would come from the physicians. It did not — because the system asked nothing of them beyond a tick.

It came from the administrative staff, and their reason was sound. Until then, knowing which patient needed follow-up was a personal skill. The experienced receptionist knew, and that knowledge was what made her valuable. The system made it public.

This is not a technical problem and training does not solve it. What worked was a change of role: that same person became the owner of follow-up quality rather than merely its executor. Instead of replacing her knowledge, the system converted it into a formal role.

In operational projects this pattern recurs almost universally. Wherever a process depends on one person's tacit knowledge, making that knowledge explicit reads as a threat unless something replaces it.

Results

MetricBeforeAfter 90 days
Follow-up return rate59%81%
Follow-ups missed within deadlineNot measured7%
Daily staff time searching for information~70 minutes~20 minutes
No-show rate14%13%

The last row matters and usually gets deleted from reports. No-show rate barely moved. We expected it to improve and it did not — no-shows had a different cause that this project never touched.

A report showing only the improved numbers is useless, because it prevents you from learning which mechanism actually worked.

What we would do differently

Two things.

First, two weeks of initial observation was not enough. We missed the clinic's seasonal pattern and hit volumes in month two that we had not designed for. Four weeks would have been correct.

Second, we underestimated staff training. The technical system was ready in three weeks; real adoption took about seven. The second number was predictable and should have been in the plan from the start.

Maintenance cost after the project ends

A system that gets delivered and has no owner returns to the previous state within a year. I have watched that happen more than once.

Here, one staff member became the owner of follow-up state, and a monthly review of the return rate was added to the clinic's existing monthly meeting. Total maintenance cost is about twenty minutes a month.

What would have happened without it was predictable: in busy months follow-ups slip; because nobody watches the number, the backlog normalises; and within a few months the system becomes an unread list.

What transfers out of this project

The core pattern is not specific to healthcare. Any business where work hangs between two customer contacts — a repair shop, a law practice, a technical services firm — has the same problem.

And the core lesson is the same: before automating anything, something has to exist that states the current state and who owns it. Automation over an unowned process only raises the speed at which things get lost.

Why not use off-the-shelf clinic management software?

It was evaluated first. The blocker was the physicians' constraint: every option required data entry during the consultation. The layer we built sat on top of the existing paper process instead of replacing it.

What was the total cost and timeline?

About ten weeks from the start of observation to stability, seven of which were organisational adoption rather than technical work. That ratio is normal in operational projects.

Was this built with AI?

No. No part of it required a language model. A state table, an owner and a deadline were the entire architecture.

Closing

The most expensive thing in this clinic was not paper. It was that nobody could say how many patients were awaiting follow-up today. That question needed neither a tool nor a model. It needed a definition.

Comparable projects are listed on the case studies page.

Ready to apply this in your own product? Book a Strategy Call and get a clear roadmap for your next sprint.

Comments (0)

Be the first to leave a comment.
Login / Sign up