Strategy18 min read

Ambient Scribe vs Clinical Document Extraction: Two Inputs

Dilr Mira is a class of private clinical small language models from DILR.AI that read clinical documents already on file. This guide separates that job from ambient scribing, which turns live consultation speech into draft notes, and shows where the same NHS safety rules apply to both and where they do not.

On this page 11 sections
Ambient Scribe vs Clinical Document Extraction: Two Inputs DILR MIRA · HEALTHCARE Ambient Scribe vs Clinical Document Extraction: Two Inputs 01 Speech in room 02 Draft note 03 Document on file 04 Fields extracted dilr.ai/blog

A clinical team weighing up AI for its paperwork usually meets two kinds of product in the same month. One is an ambient scribe: a tool in the consulting room that listens to the conversation and drafts the note and the letter. The other is a document extraction model: a tool that reads documents already sitting in the record, a lab report, a discharge summary, an intake form, and turns them into structured fields. Both are pitched at administrative load, both are built on language models, and both end in output a person has to check. They are easy to confuse, and a buyer who expects one to do the other's job ends up disappointed with both.

The cleanest way to tell them apart is to look at what goes in, not what comes out. An ambient scribe starts from live speech in a room. Document extraction starts from a document that someone, somewhere, has already written. The outputs overlap more than the labels suggest, and NHS England's own guidance lists coding and record population among a scribe's possible outputs, so the input is the line that holds. This guide is written for a CIO or Head of Digital at an NHS trust, a GP federation or a private provider who has to decide which problem they are actually solving.

The scope is deliberately narrow. The NHS ambient scribing rollout in South West London is covered in our analysis of the London AI scribing programme, and the MHRA's testing of ambient voice is covered in our note on the MHRA AI Airlock. Hazard logs and the DCB0129 safety case for an extraction model belong to our clinical safety guide for AI extraction, and the on-premise argument for extraction belongs to our clinical document extraction pillar. This post holds one thing: the split between speech and documents on file.

This guide is shipped by the team behind Dilr Mira, a family of small, private language models that read clinical documents on hardware you control and return schema-valid JSON for a person to review. Or see DATS, the five-stage method we use to place AI inside an organisation and run it.

What is the difference between an ambient scribe and clinical document extraction?

An ambient scribe and clinical document extraction solve different input problems. An ambient scribe captures live speech during a consultation and drafts a new note or letter from it. Document extraction reads a clinical document that already exists, such as a lab report or discharge summary, and returns structured fields from it. Both outputs need human review, but the scribe creates a record while extraction reads one.

That one sentence explains most of the practical differences. A scribe has to be present when the conversation happens, so it lives in the consulting room, on a dedicated microphone or a smartphone, usually integrated with the electronic patient record. Extraction has no appointment to attend. It works on whatever lands in the record, from any sender, at any time: a referral letter from another trust, a pathology report from a lab network, a form a patient filled in at reception, historical letters already in the record.

The two also differ in who is present. A scribe records a patient, a clinician and sometimes a relative or carer, and NHS England's guidance says the patient should be told and can ask for it not to be used. Extraction reads a document the organisation already holds, so there is no conversation in which to raise it. The data protection questions are real for both, but they arise at different moments.

Finally, they differ in the shape of the output. A scribe produces prose that a clinician edits: a consultation note, a referral letter, a patient letter. Extraction produces fields: a list of medications, a set of lab values with ranges, a diagnosis list, an encounter date. One is written for a reader; the other is written for a system. The extraction pillar linked above explains why structured, schema-valid output is the point of the second job.

What does NHS England's guidance count as an ambient scribing product?

NHS England's guidance on AI-enabled ambient scribing products defines them by speech. It describes tools that turn spoken words into text and other outputs with minimal user intervention, used for clinical or patient documentation and workflow support. Its input functions are capturing speech and adding secondary information, and its outputs run from transcripts and summaries to letters, clinical codes and record entries.

The guidance was first published on 27 April 2025, and when we checked it on 18 November 2026 the page showed a last-updated date of 18 November 2026 and a Version 3 label. Its definition, from section 2.1 of NHS England's guidance on AI-enabled ambient scribing products, describes ambient scribing products as:

"tools that use advanced speech technologies to automatically convert spoken words into text and other outputs, requiring minimal user intervention"

Every word of that definition points at speech. The guidance then splits a scribe's functions into inputs and outputs. The inputs are capturing or recording speech interactions, and incorporating secondary information such as context from health records. The outputs are converting recordings into transcripts, generating summaries from transcripts, formatting outputs to a style or template, extracting and linking terms to clinical codes, generating referral and patient letters, populating information in health records, and detecting or suggesting actions such as scheduling or referrals.

Two things follow for a buyer. First, the list of outputs overlaps with extraction: coding terms and populating the record are things an extraction model also does. So "scribes write prose and extraction makes fields" is not a safe rule. Second, the list of inputs does not overlap. A scribe's primary input is a recording of people talking. Secondary information from the record can sharpen its output, but nothing in the guidance describes a scribe whose job is to read the inbound document pile.

NHS England's Digital pages describe the same product class in plainer words. Its information governance guidance for ambient scribes says these tools listen in the background to capture speech or conversations and produce a task-specific output, such as a summary or a letter. That is a description of a conversation tool. Our approach to AI placement starts from the same question: which workload is the problem, before which product.

Where do the two jobs overlap, and where do they not?

Ambient scribes and document extraction overlap at the output end and separate at the input end. Both can produce coded terms, structured entries and drafts that a clinician reviews before anything reaches the record. They separate on input, timing, who is present, which national route applies and how each one fails. A buyer should compare them on those five points rather than on the output alone.

The table below sets the two jobs side by side. The ambient column follows NHS England's guidance; the extraction column follows how a model such as Mira-Q2 is described on the Dilr Mira product page.

DimensionAmbient scribeDocument extraction
InputLive speech in a consultation, plus optional context from the recordA document already written, from inside or outside the organisation
TimingDuring or straight after the encounterWhenever the document arrives, or in a batch over a backlog
Who is presentPatient, clinician, sometimes a carer or relativeNobody: the document is read where it is stored
Main outputDraft note, letter or summary for a clinician to editStructured fields for a person to check
National route in EnglandNHS England AVT Supplier Registry and ambient scribing guidanceNo national registry we could find; general NHS digital safety and data rules
Typical failureMishearing accents, dialects or impaired speech; mistaking a hypothetical for a factDocument types it never saw in training, such as operative notes

The output row matters as much as the differences. Because both tools can put structured information into a record, both need the same discipline at the point of entry: a person checks the output before it becomes part of the patient's story. The guidance is explicit that users review and approve scribe outputs before further action, and Dilr Mira's page is explicit that every Mira-Q2 output is a draft for a person to check.

The difference row that surprises most buyers is the timing. A scribe's value is bounded by the number of conversations it hears. Extraction's value is bounded by the number of documents that arrive or already sit unread, and much of that pile was never spoken aloud in the receiving organisation. The same reasoning applies to governance: an owner and a review step belong to each workload rather than to each tool, which is the kind of RACI an AI operating model sets out.

Which clinical documents arrive that no scribe ever heard?

Many clinical documents reach an organisation without any conversation it could record. Lab and pathology reports come from laboratory systems. Discharge summaries and clinic letters come from other providers. Intake forms are filled in by patients. Historical records predate any scribe. Document extraction exists for this inbound and archived pile, which an ambient scribe in the consulting room never hears.

Think about where a GP practice's or a hospital department's paperwork actually comes from. A fraction is generated inside its own consultations, and that is the fraction an ambient scribe can help with. The rest is written by someone else: the hospital that discharged the patient, the laboratory that ran the blood test, the specialist who saw them last year, the patient who completed a form in the waiting room. None of those documents passes through a microphone in the receiving organisation.

Two input problems in one clinical record
01ConsultationSpeech in the room02Scribe draftNote or letter, clinic…03Inbound documentsLabs, summaries, forms04Extracted fieldsDraft for review
An ambient scribe works on the consultation it hears; document extraction works on documents that arrive or already sit in the record. A person reviews the output of both.

This is also where the two tools meet, and the meeting point needs care. A scribe's letter, once signed off and sent, becomes a document on file for whoever receives it. In principle an extraction model at the receiving end could read it. In practice, whether a given model reads a given document type well is an empirical question. Mira-Q2, for example, is strongest on the types it was trained on: lab reports, medication lists, discharge summaries, pathology reports, intake forms and progress notes. Its page does not list narrative clinic letters among them, and its published score on real physician prose across 39 specialties is lower than on its training distribution, as the failure section below sets out.

The receiving side of a discharge summary has its own detailed treatment in our discharge summary extraction guide, which sets the hospital that writes the summary against the practice that reads it. Coding, which turns a finished record into national classifications, is a third job again, covered in our post on how clinical coding differs from extraction.

Do the same NHS safety and data rules apply to both?

The core NHS rules apply to both tools. NHS England's ambient scribing guidance says safety requirements apply to all digital products used in the NHS, whether or not they are medical devices, and it expects a DPIA and DCB0160 work from the adopter and DCB0129 from the supplier. What differs is a set of scribe-specific steps: the AVT Supplier Registry, telling the patient, and a time limit on pilots.

Start with what is shared. The guidance tells adopters of ambient scribing products to complete DCB0160 documentation, including a safety case, hazard log and monitoring framework, and a Data Protection Impact Assessment, and to ensure the supplier has completed DCB0129. It also says plainly that safety requirements apply to all digital products used in the NHS regardless of medical device status. An extraction model deployed into an NHS workflow therefore does not escape clinical safety work by not being a scribe. Our DCB0129 guide for extraction models covers that work in full, so it is not repeated here.

The same is true of data protection. Both tools process special category health data under UK GDPR, both need a lawful basis and a DPIA, and both raise the question of where the processing happens. The guidance names the risk of externally hosted AI models, where an organisation lacks control over processing and data handling, and then adds a fair balancing point: on-premises models also carry risks, to be managed with appropriate security and risk controls. Running a model inside the perimeter changes the risk; it does not remove it.

Now the parts that are specific to ambient scribing. NHS England's AVT Supplier Registry, available since 16 January 2026, is a national self-certified list of ambient voice technology suppliers with a repository of their evidence. NHS England states that it is not a commercial framework and that it does not endorse the suppliers on it. We found no equivalent national registry for document extraction models, so an extraction buyer should expect to assemble that evidence locally.

Patients are the second difference. NHS England's information governance guidance says a professional using an ambient scribe should tell the patient at the start of the session, and that the patient can ask for it not to be used. That step fits a tool that records a conversation the patient is part of. Extraction has no consultation moment to raise it in, so how an organisation tells patients about document processing is a question for its DPIA and its privacy information.

The third difference is the pilot clock. The guidance says pilots or proofs of concept for ambient scribing products can continue only if they are time-limited, up to four months, and are not used to bypass compliance. That limit is written for ambient scribing; nothing in it sets a pilot length for extraction. A sensible extraction pilot still has a fixed end date and a measured exit test, and our AI execution office is the embedded, long-term delivery team that carries work from pilot into production.

Device status is decided the same way for both: by the intended purpose the manufacturer defines. The guidance does not determine which ambient products are medical devices, and on 29 July 2026 NHS England and the MHRA wrote to the system about the MHRA's publication of guidance on the regulatory status of AVT products. For any extraction tool, the same intended-purpose test applies, and that call belongs to whoever writes the intended purpose.

How does each one fail, and how would you notice?

Ambient scribes and document extraction fail in different places, so they need different checks. A scribe can mishear speech, struggle with accents, dialects or speech impairments, and mistake a hypothetical remark for a confirmed fact. An extraction model can falter on document types or layouts it never saw in training. Each needs monitoring designed around its own weak point, run independently of the supplier.

On the scribe side, NHS England's guidance is specific. It warns that ambient voice technologies may have varying success with some accents or regional dialects, where participants speak English as a second language, or where speech is affected by disorders or impairments. It also flags contextual misunderstanding, giving the example of a patient's hypothetical statement mistaken for a confirmed diagnosis. And it asks a question every buyer should put to a scribe supplier: does the product summarise directly from the transcript, which the guidance calls extractive, or combine information to create new outputs, which it calls abstractive? The answer changes how much a reviewer has to check.

On the extraction side, the weak point is the document, not the voice. Dilr Mira's page states that Mira-Q2 was trained on synthetic data, learned that data well, and struggles with document types it never saw, such as operative notes and physical examinations. Its published evaluation puts JSON validity at 100% on 200 held-out documents from the same distribution as training, 94.7% on 150 real physician documents that match its schema, and 85.8% on 282 real physician documents across 39 specialties. The drop is published rather than hidden, and the page says retraining on real documents from a design partner is what closes the gap.

Mira-Q2 JSON validity by test set
100%Held-out100%Dialect94.7%Real, on-schema85.8%Real, 39 specialties
Published evaluation of Mira-Q2, 782 documents across four test sets: held-out documents from the training distribution, a different formatting dialect, and two sets of real physician documents. Source: DILR.ai, Dilr Mira page and Mira-Q2 model card (checked 18 November 2026)

The practical consequence is a different monitoring plan for each. For a scribe, sample consultations across the accents, languages and conditions your patients actually present with, and compare the draft with what was said. For extraction, sample by document type and sender, because a model can falter on document types it never saw in training. In both cases the guidance's principle holds: the deploying organisation's monitoring data should be collected and analysed independently of the manufacturer's.

Where does Dilr Mira fit, and where does it not?

Dilr Mira fits the documents-on-file side of the split. Mira-Q2, the current release, reads a clinical document such as a lab report or discharge summary on hardware the customer controls and returns ten schema-valid JSON fields for human review. It does not listen to consultations, transcribe speech or draft letters, so it is not an ambient scribe and does not compete with one.

Dilr Mira is DILR.AI's family of small, private language models for clinical document extraction. The ten fields Mira-Q2 returns are the document type, the patient's age and sex, the encounter, vitals, labs, medications, diagnoses, procedures, allergies and an extraction notes field. The patient object is de-identified by design, returning age and sex but never a name or record number, and the identifier leak rate was 0.0% across all 782 evaluated documents.

It runs where the documents already are: on-premise or air-gapped, and once the weights are downloaded, inference needs no network connection. Mira-Q2 is a 120 MB adapter on Qwen2.5-3B-Instruct, published on Hugging Face. Its licence has two parts: the adapter is Apache-2.0, while the Qwen base model is licensed for non-commercial use, so a commercial deployment has to check the base model's terms first. It reads English only.

What Dilr Mira does not do matters as much. It has no microphone, no transcription step and no letter-writing function. It is not an ambient voice technology, so the AVT Supplier Registry is not its route. It does not make clinical decisions; every output is a draft for a person to check. Mira-3, which aims at more languages, any schema without setup and identifiers removed before reading, is in development and is not available today. Our overview of Dilr Mira and the Mira-Q2 evaluation write-up give the detail behind each figure.

A private deployment of Mira is scoped per engagement and fitted to an organisation's own schema and documents. That work runs through DATS, our consulting system, and starts with a scoping conversation through our contact page. For organisations still mapping where AI belongs at all, our AI solutions page sets out the six recurring problems we see and what we do about each.

What is the best way to choose between an ambient scribe and document extraction in 2026?

The best choice in 2026 depends on where the documentation burden starts. If clinicians lose time writing up consultations, an ambient scribe from the NHS England AVT Supplier Registry is the route with national guidance, a registry and a large NHS study behind it. If staff lose time reading and re-keying documents that arrive from elsewhere, document extraction fits. Many organisations need both, scoped as two projects.

Four criteria settle most decisions. First, the input: count how many hours go into writing up conversations versus reading inbound and archived documents. Second, the national route: ambient scribing has NHS England guidance, a supplier registry and national funding through regional teams, while extraction is assembled locally. Third, data location: decide whether the documents may leave your environment at all. Fourth, review capacity: both tools produce drafts, so somebody has to own the check.

On the evidence, ambient scribing wins its own ground clearly. A study led by Great Ormond Street Hospital across nine NHS sites in London, using the TORTUS AI scribe and evaluating more than 17,000 patient encounters, reported a 23.5% increase in direct patient interaction time and an 8.2% reduction in overall appointment length, according to the GOSH announcement of 4 September 2025. For consultation documentation, a specialist scribe is the right tool, and an extraction model is not a substitute. Our healthcare guide to where AI pays places that evidence in the wider picture.

Extraction wins a different ground: the documents that arrive with no conversation behind them, especially where they must not leave the organisation's network. A cloud extraction service can be the pragmatic answer for a team that has a data processing agreement in place and is content for documents to be processed off its premises, and our Google Healthcare NL API comparison works through those cloud options. An on-premise model fits the team that cannot send documents out. DILR has not run a head-to-head benchmark against other clinical models, so we do not call Mira the best; the right test is a sample of your own documents.

Our approach to placing AI starts from the workload, not the tool, and that is the honest answer here. Map the paperwork first, then choose. Our strategy articles cover the operating decisions that follow.

Frequently asked questions

Can an ambient scribe replace document extraction?

An ambient scribe cannot replace document extraction, because it only works on conversations it captures. NHS England's guidance defines ambient scribing products by speech, and documents that arrive from laboratories, other providers or patients never pass through a scribe's microphone. A scribe may use record context to improve a note, but reading and structuring the inbound document pile is a separate job.

Some scribe products can populate the record and suggest codes from what they heard, which narrows the gap for the organisation's own consultations. It does nothing for the referral letter that arrived this morning. Our about page explains who we are and why we build for that second problem.

Is a document extraction model covered by the AVT Supplier Registry?

A document extraction model is not what the NHS England AVT Supplier Registry is for, since NHS England describes it as a registry of ambient voice technology suppliers. An extraction buyer in England therefore assembles supplier evidence locally: clinical safety under DCB0129 and DCB0160, a DPIA, data security assurance and a monitoring plan, with no national self-certified list to start from.

That local assembly is ordinary procurement work, and it is where most extraction projects slow down. An AI placement diagnostic works out where AI belongs before that procurement starts, and an AI operating model sets out the governance, RACI and lifecycle once the model is live.

Want to see this on your own documents? Read about Dilr Mira private clinical models, start with the clinical document extraction guide, see our DATS methodology, or talk to the operators about a private deployment.

Talk to the operators

Read the documents you cannot send out.

30-min scoping call · No deck · Confidential. We will tell you which paperwork is a scribe problem, which is an extraction problem, and where a private model fits.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.

  • ambient scribe vs document extraction
  • healthcare ai
  • ai scribe vs document ai
  • nhs ambient scribing guidance
  • clinical document extraction
  • ai scribe reddit
  • best clinical document extraction 2026
  • dilr mira

What does this article answer?

The questions this article answers, in short

  • An ambient scribe and clinical document extraction solve different input problems. An ambient scribe captures live speech during a consultation and drafts a new note or letter from it. Document extraction reads a clinical document that already exists, such as a lab report or discharge summary, and returns structured fields from it. Both outputs need human review, but the scribe creates a record while extraction reads one.
  • NHS England's guidance on AI-enabled ambient scribing products defines them by speech. It describes tools that turn spoken words into text and other outputs with minimal user intervention, used for clinical or patient documentation and workflow support. Its input functions are capturing speech and adding secondary information, and its outputs run from transcripts and summaries to letters, clinical codes and record entries.
  • Ambient scribes and document extraction overlap at the output end and separate at the input end. Both can produce coded terms, structured entries and drafts that a clinician reviews before anything reaches the record. They separate on input, timing, who is present, which national route applies and how each one fails. A buyer should compare them on those five points rather than on the output alone.
  • Many clinical documents reach an organisation without any conversation it could record. Lab and pathology reports come from laboratory systems. Discharge summaries and clinic letters come from other providers. Intake forms are filled in by patients. Historical records predate any scribe. Document extraction exists for this inbound and archived pile, which an ambient scribe in the consulting room never hears.
  • The core NHS rules apply to both tools. NHS England's ambient scribing guidance says safety requirements apply to all digital products used in the NHS, whether or not they are medical devices, and it expects a DPIA and DCB0160 work from the adopter and DCB0129 from the supplier. What differs is a set of scribe-specific steps: the AVT Supplier Registry, telling the patient, and a time limit on pilots.
  • Ambient scribes and document extraction fail in different places, so they need different checks. A scribe can mishear speech, struggle with accents, dialects or speech impairments, and mistake a hypothetical remark for a confirmed fact. An extraction model can falter on document types or layouts it never saw in training. Each needs monitoring designed around its own weak point, run independently of the supplier.
  • Dilr Mira fits the documents-on-file side of the split. Mira-Q2, the current release, reads a clinical document such as a lab report or discharge summary on hardware the customer controls and returns ten schema-valid JSON fields for human review. It does not listen to consultations, transcribe speech or draft letters, so it is not an ambient scribe and does not compete with one.
  • The best choice in 2026 depends on where the documentation burden starts. If clinicians lose time writing up consultations, an ambient scribe from the NHS England AVT Supplier Registry is the route with national guidance, a registry and a large NHS study behind it. If staff lose time reading and re-keying documents that arrive from elsewhere, document extraction fits. Many organisations need both, scoped as two projects.
Written by Dilr.ai EngineeringEngineering team

Who can help us place AI where it pays off?

DILR.ai consulting takes AI from placement to production

The DATS system runs from a 4 to 6 week placement diagnostic through to embedded delivery, so AI reaches production instead of staying a pilot.