Verpleegkundige-rapportage ROI: fourteen ranked mistakes

It is a Tuesday in March 2026. A bestuurder at a 138-bewoner zorginstelling in Gelderland is reading a three-year offer from a Nuance reseller. The number on page four is €172,000. The sales note says "verpleegkundigen save 47 minutes per shift on dictation". The SOFI her-accreditatie is booked for September. The kwaliteitsmanager wants the rollout signed off before the audit team walks in.

They sign on Friday. Six months later, they are spending roughly five times what an equivalent Whisper-large-v3 plus Sonnet rapport-structuur stack would have cost. The per-bewoner ROI has landed at half the deck's projection. And one clause buried in the framework agreement is going to make the next SOFI cycle longer, more expensive, and harder to pass.

We have sat at four of these tables in the last twelve months. The same fourteen misreadings turn up almost every time. Here they are, ranked by how hard the mistake is to walk back.

What the rapportage-agent is actually doing

Before the field-guide: the thing being bought. A verpleegkundige-rapportage agent takes a nurse's spoken or typed note ("mevrouw De Vries had vannacht onrust, om 03:15 omgedraaid, huid intact, geen pijnscore boven 2") and turns it into a structured entry in the ECD: Nedap ONS, PinkRoccade Cura, Carenzorgt, whichever your instelling runs. Two steps. Transcribe the speech, then structure the prose into the rapport-vorm your kwaliteitsmanager has signed off.

Dragon Medical One is sold on step one. The pitch is medical-vocabulary accuracy and on-prem deployability. Step two, the structuring, ships as rigid, template-bound, hard to change after rollout. The Whisper + Sonnet stack flips the ratio. Whisper-large-v3 handles Dutch medical speech at sub-5% WER once you feed it a 200-term glossary of zorg-Nederlands via the initial_prompt parameter. The structuring runs as a Sonnet call with a cached system prompt: the rapport-vorm is the cache, written once, billed as input tokens at the discount tier on every subsequent rapport.

The transcription is the cheap part. The structuring is the work. Almost every mistake below is a variant of pricing the cheap part as if it were the work.

Group A: reversible in a single workflow-config

Seven of the fourteen. The kwaliteitsmanager can fix all of them before the next shift-wissel without touching procurement.

1. Quoting license cost instead of cost-per-rapport

The reseller deck shows €99 per seat per month. That number is a distraction until you divide by rapporten written. Forty seats writing nine rapporten each across two shifts a week comes to roughly 32,000 rapporten per cluster per year. Cost-per-rapport on the Dragon licence: about €1.49. Cost-per-rapport on the Whisper + Sonnet stack, glossary and structuring cache included: roughly €0.01. Same kwaliteit at the kwaliteitsmanager's signed-off rapport-vorm. Fix: rewrite the business case in cost-per-rapport. One spreadsheet column.

2. Skipping the rapport-structuur cache

The SOAP layout, the ABCDE layout, your instelling's own intake-rapport. None of these change between rapporten. They are the system prompt. Anthropic's prompt cache discounts cached input tokens by 90%. If your stack is rebuilding the rapport-vorm prompt on every call, you are paying full price for tokens that have not changed since the kwaliteitsmanager signed them off. Fix: add the cache_control breakpoint. Fifteen minutes of engineering, recurring monthly saving in the four-figure range.

3. Confusing transcription accuracy with structuring quality

The Nuance pitch fixates on WER. WER is the wrong number. The verpleegkundige does not care that decubitus got transcribed correctly if the rapport still buries the wondzorg observation in the wrong veld. The metric that matters: percentage of rapporten that land in the right ECD-veld on the first save, with no rework. Measure that. Fix: replace WER on the kwaliteitsdashboard with first-save-correct rate.

4. Treating the agent as autocorrect instead of a structuring layer

Half the rollouts we have audited bolted the agent in as a smart Notepad. The verpleegkundige dictates, the rapport appears in a textarea, the verpleegkundige edits and saves. That ships none of the structuring win. The agent should be writing directly into ECD-velden over the integration API, with the textarea as a confirmation step. Fix: rewire the front-end to render the structured payload, not the freeform transcript.

5. Counting minutes saved without booking the after-hours overspill

The 47-minute number in the deck is a per-shift average. The actual saving the bestuurder cares about is the verpleegkundige who used to stay until 17:45 finishing rapporten and now leaves at 17:00. Book the saving in after-hours minutes, not total minutes. Fix: add an end-of-shift logout timestamp to the kwaliteitsdashboard. Three lines of SQL against the ECD audit log.

6. Claiming Whisper does not know zorg-Nederlands

It does not out of the box. It does the moment you give it a 200-term glossary via the initial_prompt parameter. Decubitus, MDL, ADL, BIG-nummer, the names of the twenty most-used medicijnen on your bewoner-lijst. Fix: write the glossary once with your verpleegkundige-team in an afternoon, version it in git, regenerate quarterly when the medicijn-overzicht changes.

7. Booking the saving on overhead hours instead of direct-care hours

If 47 minutes per shift goes back into the verpleegkundige's day and that time is spent in the koffiekamer, the per-bewoner ROI is zero. The saving has to be visibly reallocated to direct-care minutes: an extra ronde, an ADL-handeling that used to get skipped at 16:30. Fix: include "direct-care minutes per bewoner" on the same dashboard as the rapportage-tijd. The two numbers move together or you do not have ROI.

Group B: procurement-level renegotiation

Five of the fourteen. Reversible, but it takes a renegotiated framework, not a workflow-config commit. Three months of legal and finance work.

8. Buying per-seat when the actual unit is per-shift

Forty verpleegkundigen on the rooster, fourteen on at once across dag, avond and nacht. Per-seat licensing pays for the other twenty-six. Per-shift licensing, or token-based pricing on Whisper + Sonnet where you pay per actual call, matches the real demand curve. Fix: in the next negotiation cycle, push for concurrent-user pricing or move off the seat model entirely.

9. Forgetting night-shift dictation volume

The night-shift verpleegkundige handles every bewoner-event during the eight-hour stretch. Average dictation length per nacht-rapport is shorter, but the rapport count is higher. We have measured night shifts running 1.6 to 2.2 times the rapport volume of day shifts on the same ward. Sales decks model day-shift averages. Fix: re-model the volume forecast off night-shift logs from your existing ECD before signing anything.

10. Pricing Dragon Medical at list

The list price on the offer is not the price your peer instellingen are paying. VECOZO-aligned zorg-tier pricing, multi-year discounts, and the explicit "we are piloting Whisper next quarter" lever bring the real number down 30 to 40%. Fix: ask three peer bestuurders in your regional koepelorganisatie what they are paying. Bring those numbers to the next call.

11. Ignoring the ECD-integration line item

The agent is worthless if the structured payload does not land in Nedap ONS or PinkRoccade Cura without manual copy. Both vendors charge a one-off integration set-up and a monthly API fee. We have seen this line item left off the original deck and tabled at €18-24k after signature, when leverage is gone. Fix: ask for the full integration TCO in writing before signature, not after.

12. Picking voice as the primary modality when the team types

In four of the instellingen we have worked with, between 55% and 70% of the verpleegkundige-team types faster than they dictate. They are under 35, fast on a phone keyboard, and have never used Dragon. Voice as the primary interface is a generational assumption that the sales rep is too polite to challenge. Fix: ship both modalities to the same structuring backend. Transcription cost is the same either way; UX preference is per-verpleegkundige.

Group C: forces a fresh DPIA and a SOFI her-accreditatie

Two of the fourteen. These are the ones the bestuurder wakes up at 04:00 worrying about. Both involve documents on file with external partijen: the kind that need the kwaliteitsmanager, a jurist, and at least one external auditor to walk back.

13. Locking a 5-year term against a 4-year SOFI cycle

The Nuance framework is sold on 60-month terms. The SOFI cycle is four years. The lock means the next her-accreditatie lands inside a contract you cannot yet exit, with no leverage to fix anything the audit flags. The DPO and the kwaliteitsmanager then have to document why the instelling is paying for a tool the auditor questioned. Fix: never sign a clinical-tooling term longer than the her-accreditatie cycle. If you have already signed, open the exit-clause review now, not in month 48.

14. Skipping the AVG/DPIA review until after the pilot

This is the one. The verpleegkundige-rapportage agent processes bijzondere persoonsgegevens (gezondheid) by definition. The Autoriteit Persoonsgegevens classifies this as high-risk processing. A DPIA is not optional. If the pilot ran without one and the SOFI auditor asks for it, you are reconstructing the DPIA retroactively against a system already in productie. That is months of work, an external DPO review, and depending on what surfaces, a Verwerkingsregister update that triggers its own her-accreditatie review.

Fix: there is no workflow-config fix. The fix is "do the DPIA before the pilot." If you are already past that point, schedule it for next week and budget for a six-to-nine-month overhang on the SOFI prep.

The five-minute audit you can run tomorrow

Pull the last 200 rapporten from your ECD. Count: how many landed in the right veld on first save? How many needed verpleegkundige rework? What was the average tijd from dictation-start to ECD-save? Three numbers. They tell you whether the tool you are paying for is doing the structuring work or just the transcription work. If it is the second, you are overpaying, usually by a factor between three and seven.

When we built the rapportage-agent stack for a 96-bewoner instelling in Brabant earlier this year, the thing we ran into was exactly mistake 4: the previous vendor's tool was running as a smart Notepad, and the structuring win was leaking out in the textarea step. We rewired it to write directly into the ECD-velden via the integration API and the first-save-correct rate jumped from 38% to 81% in six weeks. The work itself is what we do at ABN's AI agents practice; the three-number audit above you can run without us.