The Hard Part Isn't the Technology Anymore

An animal, hundreds of miles away, took out a network line.

That's the story from the KLAS Enterprise Imaging Summit I keep coming back to. Not because it's dramatic. Because it's the point. The hard part of cloud and AI isn't the technology anymore. It's the circuits, the people, who breaks a tie, and the day after go-live.

If you lead imaging, here's what the people who've made the move would tell you. I'm not naming anyone. People spoke candidly, and that's worth protecting.

Go-live isn't the finish line

Cloud and AI are tools. Getting them isn't the achievement. What you do with them is.

Many projects still end at go-live. The contract's signed, the switch is flipped, everyone celebrates, and nobody tracks what happens next. The best examples measured ER turnaround, radiologist productivity, even whether radiologists liked the job more. One group took post-processing and routine steps off the radiologist's plate, so reading was all they did.

Try this: before your next project starts, write down the one or two numbers that will tell you it worked a year from now. Then name who watches them.

Watch the workarounds you've stopped noticing

Engineers design systems to fail safe, or to keep running when something breaks. The real danger is slower. A small workaround becomes routine. Then another. Nobody ever decides to accept the risk. It's called the normalization of deviance.

Imaging is full of it. The downtime procedure nobody has rehearsed. The alert everyone ignores. The manual step that quietly became policy.

Try this: ask your team one question: what have we stopped noticing? You'll learn more from the answers than from most dashboards.

The keynote added one more way to fail, the one clinical care needs: failing right. Things will go wrong. Meet it with honesty and humility, and keep the patient at the center. We don't treat patients, but we decide how much clinicians can trust their tools.

Nobody's asking whether AI works anymore

They're asking how to run it. A few ideas came up again and again:

• Judge AI as a portfolio. Few algorithms pay for themselves alone. A group of them can.
• Don't let vendors grade their own homework. Use independent monitoring to catch drift and uneven performance.
• Have a backup for critical algorithms. If a vendor disappears, stroke detection shouldn't go with it.
• Be willing to turn things off. A tool that misses its targets should go.

Reporting AI got the most attention. It saves real time once it sits inside the dictation workflow, not next to it.

The bigger shift is from findings to context. Most AI stops at "there's a nodule." The next step is what that means for this patient, given their history, labs, and prior studies. That used to take hours. It's starting to take seconds.

One speaker offered a five-rung ladder for judging any imaging tool, from a 1991 paper by Fryback and Thornbury. Does the image look good? Is the finding right? Did it change the doctor's thinking? Did it change the treatment? Did the patient do better? Most AI stops at rung two.

Fragmentation is a governance problem

One health system ran more than ten cancer centers as separate islands. Every region had a "number one" priority, so small projects won and big ones stalled. What fixed it wasn't new software:

1. One intake form with scope, IT impact, and funding spelled out up front.
2. A cross-regional committee of clinical, operational, and product leaders.
3. One person who breaks ties.
4. Two IT tracks. One keeps today stable. One evaluates what's new.

The ask to vendors was clear and fair: real support for DICOM, FHIR APIs, and IHE profiles, so the data can move.

One longtime informatics leader shared a wish list for the next archive. It's a good one to hold up against your own roadmap:

• One cloud-native, zero-footprint viewer, with AI results inside it.
• Real support for every specialty, nuclear medicine included.
• Local caching for slow connections, plus streaming when bandwidth allows.
• A home for AI results, not just images.
• Access rules that fit outpatient reality.

Order your circuits before you sign

The cloud advice was the most practical of the summit:

• Order circuits early. Lead times can run six months or more.
• Build real redundancy. Diverse paths, not two lines in the same trench. (Remember the animal.)
• Double your estimate for data discovery. Moving data is easy. Finding it isn't.

One academic health system went live on hybrid cloud in under five months. Their lessons:

1. Decide what's moving. It doesn't have to be all or nothing.
2. Give clinical operations a seat in governance.
3. Spend real time on the design. Don't accept the default.
4. Bring your security and network teams in early.
5. Migrate in phases, most important data first.
6. Reset what your team does. When someone else runs the servers, your people can move closer to the clinical work.

A pilot is a decision, not a rollout

A pilot asks one question: are we keeping this? An initial site starts a rollout everyone has already committed to. Mix them up and people hold back.

To get out of pilot purgatory, both sides need something at stake:

• Put money on the table, even a little. A paid proof of concept, credited back if you buy, brings in procurement and starts the clock.
• Agree on success up front, with the contract ready once you hit it.
• Decide what you'll measure before you start. Most ROI projections are never checked.

And remember who decides. A physician may ask for a tool, but techs, nurses, and admin staff decide whether it survives.

AI can only go as far as your data can

The closing fireside chat mapped AI on two axes. One is what it can understand, from a single narrow algorithm up to models that read images, text, and the patient record together. The other is what it can do, from a standalone tool up to running workflows on its own. Most of radiology sits near the start of both.

Moving further takes three things: access to data across the enterprise, a place inside the workflow, and permission to act. Access comes first. We've written about why in The data has to move first.

The hopeful part: when a profession gets more productive, demand tends to grow. The case made was that radiology's role gets bigger, not smaller.

Eight things to do before your next project

• Write down the one or two outcomes that will tell you it worked, and who watches them.
• Ask your team what workarounds you've all stopped noticing.
• Order network circuits before you think you need them.
• Double your time estimate for finding legacy data.
• Bring security and network teams in before the cloud design is set.
• Name one person who can break ties between sites.
• Have someone other than the vendor monitor your AI.
• Agree on what success looks like before any pilot starts.

Thanks to everyone who shared so openly, on stage and around the tables.

One question before you go: what's the workaround your team has stopped noticing? Tell me on LinkedIn. I'd like to know if it's the same one I keep hearing about.