On September 29 at SROA in Boston, Heather Turner and I gave an hour on buying AI in a radiation oncology program. I built it around a deal I sold from the vendor side of the table and the five places it broke.

Some of you saw the start of this story on LinkedIn. It was a mid-size, multi-vendor cancer center, and a physician there wanted a workflow layer on top of the OIS to cut clicks. The product was real, and it worked. Then it went up on a screen as AI-enabled, and nobody could tell me what that meant, or how much time it saved, or whose.

The physician had never asked those questions either, and not for lack of sense. She had seen a demo, and the demo felt like it answered all of them.

What does it do? The best way to answer it: have the vendor read the product description with the word AI taken out. If nothing is left, there is no product under it. If something is left, that is what you are buying.

In your own program, it starts with the problem you are trying to solve. In Boston we used one: a head and neck case is late, plan approval slips, and dosimetry works after hours to catch up.

Name the specific step you want fixed; "treatment planning" is too big to measure. Pick the number that will show it moved, CT to approved plan in our case, and measure it before anything goes in. One person, a clipboard and about a week: who touches the case, in what order, how long each touch takes and how long it sits between touches.

That record matters because no vendor has it. The tool was designed somewhere else, at an academic center or in Europe or at a community center that looks nothing like yours.

The person who wants the tool most cannot be the only one grading it, because they have every reason to say yes. That person stays involved. Before the vendor comes in, physics, dosimetry, the physician and the manager each write down the answer that would work from their chair: physics on what it validates and how often, dosimetry on how long review takes, the physician on what they are signing, the manager on whether it can be staffed.

Then you grade the demo against answers you already wrote. Without them, the demo looks amazing, because the demo was designed to look amazing.

Programs know the OIS rhythm: one major release a year, a project manager, an interface team. Very few have anyone managing the AI side, and the models underneath change far more often. FDA's predetermined change control pathway lets a manufacturer pre-authorize future model changes and ship them without a new submission. The pathway works as designed. It does not tell your department when the model changed.

Your physicist checks the LINAC against a tolerance table. There is no tolerance table for the AI model. Treat the software like the machine: commissioning, a QA schedule, an owner.

Who owns the model? Before signature, ask what it is cleared for: the label, like a drug label. Ask what it was trained on, and whether those institutions and that patient population look like yours. Ask them to run it on your own cases, ones your dosimetrist already knows the answer to. Get it in writing that they will tell you before the model changes. The vendor does not have to offer that; you have to ask.

After go-live, rerun a fixed set of cases every quarter and after every update. Physics should own that; it already lives with tolerance tables, commissioning and QA.

At 7:15 on a Tuesday, a physician gets an output that doesn't quite match what she needs, and calls the help desk.

"Who do I talk to?"

"I'll open a ticket."

"I've got a patient at eight o'clock."

"Got it. I understand."

Who owns the service? Your LINAC contract prepared you for one failure: the system is down. AI adds a second: it is up and the answers do not make sense. Ask who at the vendor can explain an output, and how fast you reach them. Ask what happens to the patient while they figure it out; that answer is a workflow, a way back to doing it by hand. Physics should be able to pull the tool from clinical use the way it would pull a LINAC.

Four hours is a rescheduled patient. Longer, and you have rescheduled your week.

The engineers who build these tools are good people doing good work, and most have never taken that call with a patient on the table. The contract is where you put the patient back in that discussion.

When a tool goes sideways, one of two things follows. Nobody catches it, because nobody designed a review step. Or everybody catches it, because dosimetry is on overtime and physics is redoing work.

Nothing in radiation oncology works automagically. Not contouring, not planning, not adaptive, not the note. Every output is reviewed by someone on your payroll. A vendor who says the review is going away is selling a product that does not exist.

What happens when it is wrong? Who catches it, who signs it, who is named. Design that step before go-live: who looks, when, how many minutes per case, and what they stop doing to make room. Write down a name; piling it onto the dosimetrist's day pays for the work twice. When an ROI shows review in year one, spot checks in year two and nothing in year three, budget review time for the whole term. And measure the whole path from CT to approved plan, because a vendor's figure for one step describes that step. Heather's column below is the floor's side of that path: seeing where a plan stands between simulation and first treatment.

What does it cost in year three? This newsletter took the AI line apart in July, so briefly: if it was not carved out up front, nobody knows what it should cost at renewal. Ask for the standalone price before you sign. It is free to ask.

Measure the AI by the workflow it changes: the same number before go-live, then at 30, 60 and 90 days. If the tool does not pay for itself, that can be fine; it may keep staff, or send a physician home earlier. Call it what it is. Your CFO can deal with whatever the answer is. What a CFO cannot deal with is not knowing it.

The five are for the vendor meeting. One question sits above them, for you, before the meeting starts: what are you looking for it to change in your program, and what number will show you it did?

If you cannot answer it, you are not evaluating software yet, and you should not be buying it. If you can, and the tool moves your number without a new bottleneck, with review and QA built in, at a cost you can carry, say yes. Some of these tools are really good. The point of the hour was to let you say yes, and own the yes.

On "saves time, clicks and money."

The Pitch: "It saves time. It saves clicks. It saves money."

The Reality: That was the promise on the slide in the deal behind this issue, and nobody could put a number on any of the three. The product underneath was real and did real work. The number was missing. A time-saved figure is a measurement somebody made, so ask for the study behind it: how many cases, whose department, which modality mix, and who ran it. A figure from one step in someone else's department describes their path, and the program buying the tool lives with its own. Hold the claim against your own CT-to-approved-plan number, taken the same way by the same person before go-live and again at ninety days. If the tool moves that number, the slide was right.

With Heather Turner, RT(T), PMP

The Dosimetry Planning Black Box: Why your treatment planning pipeline needs visible milestones.

In many radiation oncology departments, the dosimetry department operates like a black box. A CT simulation is completed, the case is sent to the planning queue, and the rest of the department simply waits for a plan to emerge. When treatment therapists don't know where a plan stands, they can't accurately forecast their first day treatment schedule.

This lack of visibility creates artificial panic. Dosimetrists get constantly interrupted by phone calls and messages asking if a plan is ready, which breaks their concentration and slows down planning throughput. Meanwhile, physicians end up reviewing plans late at night, pushing physics QA into the early morning hours right before the patient arrives. Worse yet, this delay can force the clinical team to pull the patient off the schedule and postpone their first fraction to another day.

To build a predictable, low stress planning pipeline, consider establishing transparent milestones across the dosimetry workflow using three clear operational standards:

Define Standard Turnaround Tiers by Case Complexity: A straightforward palliative spine treatment shouldn't follow the same planning timeline as a multi target stereotactic brain plan. Work with your dosimetry and medical physics leads to establish realistic, tiered planning windows. Set targets like 24 hours for urgent palliative cases or 3 to 4 days for complex IMRT and VMAT, then build those milestones directly into your OIS workflow tracks.

Enforce Hard Contouring Deadlines for Physicians: Dosimetrists cannot build a plan without target volumes and organs at risk. When physician contouring drags past the designated window, the entire planning timeline collapses. Establish a clear operational expectation that physicians must finalize contours within a certain timeframe, such as 24 to 48 hours post-simulation, to protect the dosimetry team's planning time.

Eliminate Status Inquiries with Visual Queues: Use your OIS task list or electronic status boards to display real time plan progress across clear stages like Contours Complete, In Dosimetry for Planning, Physician Review, and Physics QA. When the entire team can see case statuses at a glance, phone calls to dosimetry drop, interruptions disappear, and plans move through the pipeline faster.

Opening up the planning black box isn't about micromanaging your dosimetrists. It is about protecting their time so they can focus on high quality plan design. When you bring transparency to your planning milestones, you eliminate last minute physics scrambles, reduce staff stress, and ensure your patients start their treatment on schedule.

Before anyone books the demo, write one page with three lines on it.

The step. The name of the step you want fixed, as narrow as it goes: normal-structure contouring for head and neck, if that is the one.

The number. One measure that shows it moved. CT to approved plan, timed for a week by one person with a clipboard before go-live, then again at ninety days by the same person, the same way.

The reviewer. The name of the person who reviews the output, on which shift, and how often.

Then take five questions into the vendor meeting and write down what they said, in their words. Where they have no answer, write "did not provide." That is a finding, and it goes on the page next to everything they did give you.

Q1. What does it do? What problem are we trying to solve? Take the word AI out of the sentence and ask them to say it again.

Q2. Who owns the model? What changes after we buy it, and who owns that change?

Q3. Who owns the service? Who do we call when it fails? Walk us through the first time it fails, not whether.

Q4. What happens when it is wrong? Who catches a bad output before it reaches a patient, who signs it, and who is named?

Q5. What does it cost in year three? What is the standalone price, what does year three renew at, and what caps the increase? Which invoice line does it live on?

And the one before them, for you: What are you looking for it to change in your program, and what number will show you it did?

None of these are technical. Watch how they answer. The one-page version, to carry into the meeting or forward to your CFO, is at medsolvedynamics.com/sroa.

A Medsolve read starts with one document the program already has, here the AI proposal or the quote with the AI line on it, and comes back as one page finance can act on, at no charge.