Migrating from PowerScribe: What Radiology Groups Should Evaluate Before Switching

Most radiology groups asking about migrating from PowerScribe are not actually unhappy with speech recognition. They are unhappy with how long the report takes after the dictation is done: structuring it, remembering which follow-up guideline applies, keeping the impression consistent with the findings, and documenting technique and indication in the words the coder needs. Those are drafting problems, and an enterprise dictation system was never built to solve them.

This checklist is for the practice manager, medical director or IT lead evaluating an AI report-drafting tool for a radiology group. It assumes you will keep reading on your current PACS and that the decision is whether to add a drafting layer, replace the dictation system for part of the group, or do nothing. It is written by a radiologist who reads a full clinical schedule with such a tool, so it is opinionated about what matters and honest about what the tool does not do.

First, be clear about what you are comparing

PowerScribe is enterprise speech recognition with deep PACS and RIS integration: you dictate the complete report and it transcribes it accurately. An AI drafting tool inverts the work: you give it a scratch read, a worksheet or the images, and it writes the structured draft for the radiologist to review, edit and sign. One automates transcription; the other automates drafting. The two are not substitutes for each other in every workflow, and a vendor who tells you otherwise is selling.

Three situations come up repeatedly:

• The hospital pays for PowerScribe and the integration works. Keep it. Add a drafting layer that writes into it, so the radiologist still signs in the system the hospital expects. The question becomes whether the drafting layer earns its cost in minutes saved and errors caught.
• Independents, teleradiologists and locums reading for three groups on three systems. Here the drafting tool often is the reporting system, because there is no enterprise platform to integrate with. The evaluation is about quality and compliance, not integration.
• A group considering a full switch to save licence fees. Slow down. Model the dictation path first: who transcribes the parts of the report the AI does not draft, and how does the signed report reach the RIS. If the answer is "copy and paste", that is fine for a solo reader and a real cost for a 40-radiologist group.

The side-by-side comparison covers where each approach is genuinely stronger.

1. Workflow fit

Ask the vendor to show the path from images to signed report on your actual setup, not a demo environment.

• Where does the draft land? In the dictation field, in a browser window beside PACS, or in a separate application the radiologist has to switch to?
• Does it write back into the dictation system you keep, and with one keystroke or several?
• Can it read the study off the workstation (DICOM import or viewer capture) or does the radiologist screenshot it?
• What happens on a hospital network with SSL inspection or a locked-down workstation? Many tools work in the vendor's office and fail behind a health-system firewall.
• How does a locum who reads for several groups keep style and rules separate per group?

Red flag: a demo that only ever starts from a clean pasted report. Real input is a messy scratch dictation at 2 a.m.

2. Measured report quality, not demo quality

The number that matters is how much of the draft the signing radiologist keeps, measured on reports that were actually signed, with the method published. Ask for three things:

• Edit rate on signed reports — the share signed without any change, and the share of draft text retained when there were changes — with exclusions stated (imports, re-runs, challenges).
• The study mix behind the number. A 95% untouched rate on normal chest radiographs says nothing about a CT abdomen with six findings.
• Your own fifty cases. Any vendor worth a contract will let two of your radiologists run fifty de-identified cases during the trial. Count edits yourself.

MyRadAgent publishes its method and live figures on the validation page; hold every vendor to the same standard, including us.

3. Guidelines, coding and the parts the coder sees

A drafting tool that does not know Fleischner, Lung-RADS, TI-RADS, LI-RADS and Bosniak is a chatbot with radiology vocabulary. Check that the guideline fires only when the modality and finding warrant it (TI-RADS on ultrasound, not CT), and that the decision table is cited in the draft so the radiologist can verify it.

Then check the unglamorous part: does the draft carry codable indication wording, the required elements for the exam type, and the technique statement a CPT code needs (contrast phase, views, Doppler, post-processing)? Avoidable downcodes and denials come from exactly these lines. The MIPS quality-measure guide describes which measures a drafting tool can document for you.

4. Style learning and the group rulebook

Two different things, and you want both:

• Per-radiologist learning — every accepted edit teaches the tool that radiologist's phrasing, ordering and sign-off style, without a template-maintenance chore.
• A group rulebook — the conventions the practice enforces across readers (how to report incidental nodules, what the critical-result line says, section order), injected into every draft and editable by a practice administrator, not the vendor.

Ask whether one radiologist's corrections can ever leak into another's drafts. The right answer is no, and the practice rules are the only shared layer.

5. Security and compliance

This deserves its own review; the security checklist for AI radiology tools walks through twelve questions. The short version for a group contract:

• PHI de-identified before any model call, with the identifier categories named.
• Business Associate Agreements executed with every AI vendor and the host, with dates, and a BAA with your practice.
• No training on your data, in writing.
• Retention and deletion schedules, export on demand, per-account isolation.
• Audit log and an export of it for your compliance officer.
• A plain statement of what the vendor does not yet hold (attestations, penetration tests, SAML). Vendors who claim everything have usually verified nothing.

MyRadAgent's answers are on one page at /security, including the "not yet" list.

6. The cost model

Enterprise dictation is priced per seat through the institution. Drafting tools are priced per radiologist, per report, or on pooled annual volume. For a group, pooled volume is usually the right shape: busy nights and part-time days balance out and nobody hits a cap mid-shift. Questions to settle before the pilot:

• What is a "report" for billing, and are failed generations counted?
• What happens at the cap: a hard stop mid-case, metered overage, or a warning?
• Is there a contract term, or can the group leave at the end of a month?
• Who pays: the group, the hospital, or the individual radiologist for their independent work? Published self-serve tiers make the individual answer easy; a group needs an invoice.

Published pricing, including the group tier, is at /pricing.

7. Run a pilot that answers the question

A good pilot is small, short and measured:

• Two to four radiologists, including one sceptic, reading their normal mix.
• Two weeks. Long enough for style learning to show, short enough to hold attention.
• Baseline first. Time per report and a subjective quality score on the week before, so the comparison is against your own numbers.
• Count rejections, not just edits. Drafts the radiologist discards without signing are the error signal that signed-report statistics never see. A tool that logs them and tells you why is doing your evaluation for you.
• Exit criteria written down in advance: a draft-retention floor, a time saving per report, zero unresolved compliance questions. Decide what "no" looks like before you start, or the pilot will drift into a permanent maybe.

8. Exit and data portability

Before signing, confirm you can leave: export of every signed report and style rule in an open format, deletion on request within a stated period, and what the vendor keeps afterwards (audit logs, usually, for a statutory period). Read the privacy policy's subprocessor table, not the sales deck.

9. Change management is smaller than you think

Thirty minutes per radiologist is a realistic onboarding if the tool runs beside the existing PACS and writes into the existing dictation field. No macro migration is needed, because the tool drafts from scratch; the group rulebook captures the conventions the macros were enforcing. One clinical champion who reads with it daily matters more than any training programme.

What a sensible migration looks like

Pilot with a few readers in the browser, no IT project. If the numbers hold, roll the drafting layer out to the group with the practice rulebook in place, and add the desktop client for the readers who sign in PowerScribe so the draft lands in the dictation field. Keep the dictation system for as long as the hospital pays for it and the integration earns its keep. Revisit the licence question a year later with your own data instead of a vendor's.

If you read for a group and want to run that pilot, the fastest route is a short call: request a workflow audit, or start on the free trial yourself first and bring your own fifty cases.

PowerScribe is a registered trademark of Nuance Communications, Inc., a Microsoft company. MyRadAgent is not affiliated with, sponsored by or endorsed by Nuance or Microsoft; the mark is used only to identify the product being discussed. Every MyRadAgent output is a draft for review and signature by a licensed radiologist.