The Crown Prosecution Service: Beam Notes
Beam Notes is a digital solution designed to help prosecutors quickly and securely review video recorded evidence through transcription and summarisation of digitally recorded media.
1. Summary
1 - Name
Beam Notes
2 - Description
Beam Notes is an AI‑powered transcription and summarisation tool developed by Beam to support prosecutors in reviewing video interviews. Through using this tool, the CPS aims to reduce administrative burden, improve consistency, and speed up case progression by automating the generation of transcripts and producing structured summaries.
3 - Website URL
4 - Contact email
CPS Cyber Security Team <CPSCyberSecurityTeam@cps.gov.uk>
Tier 2 - Owner and Responsibility
1.1 - Organisation or department
Crown Prosecution Service
1.2 - Team
Digital & Information Directorate
1.3 - Senior responsible owner
Deputy Director of Digital Delivery
1.4 - Third party involvement
Yes
1.4.1 - Third party
Beam Up Ltd
1.4.2 - Companies House Number
10637337
1.4.3 - Third party role
Beam provide technical development and deployment support for Beam Notes.
1.4.4 - Procurement procedure type
Framework agreement call offs
1.4.5 - Third party data access terms
Beam’s designated team is provided with controlled access to CPS-uploaded data using the Beam Notes application. This access is strictly limited to the processing, transcription, and summarisation of video evidence for the purpose of generating transcripts and summaries. Beam does not integrate Beam Notes with other CPS systems, and all data processing is carried out in compliance with data protection legislation. Access is restricted to authenticated CPS staff via Single Sign-On, and Beam Notes retains data only for a limited period (up to 30 days) before deletion. All Beam staff and sub-processors are subject to contractual and technical controls, including ISO27001 and CyberEssentials Plus certification, to ensure data security and appropriate use.
Tier 2 - Description and Rationale
2.1 - Detailed description
Beam Notes enables CPS to process video evidence more efficiently by providing an end‑to‑end workflow for uploading video files and generating time‑stamped transcripts with structured summaries. This supports faster handling of video testimonies and reduces administrative burden. All summaries produced by Beam Notes are subject to a full human review and verification process prior to use. Beam Notes uses commercial large language models for speech‑to‑text transcription and summarisation. CPS and Beam co‑designed templated structures that provide clear guardrails, ensuring that summaries align with CPS requirements. These models are used exclusively for transcription and summarisation; no automated decision‑making is performed, and data is not retained or used for model training by Beam or its sub-processors.
2.2 - Benefits
The use of Beam Notes provides prosecutors with searchable, AI‑enhanced transcripts and summaries, that enable significantly faster and greater consistency in evidence reviews. The tool enhances efficiency, reduces administrative burden, and improves accessibility while ensuring users remain fully in control of the information they review and use.
2.3 - Previous process
The previous process for reviewing video recorded evidence was entirely manual and time‑consuming: prosecutors watched long recordings, repeatedly pausing to take handwritten or digital notes and often transcribing key sections verbatim, with transcripts usually only available at a later stage of the case review process.
2.4 - Alternatives considered
Beam Notes was chosen for its tailored functionalities to support CPS’ video evidence review as a one-off solution. Other functional capabilities and alternative solutions continue were explored and considered as part of the wider programme.
Tier 2 - Deployment Context
3.1 - Integration into broader operational process
Beam Notes is not a decision‑making tool and does not automate, infer, or advise on case outcomes or charging decisions. It supports the evidence‑review stage of the prosecution process by helping prosecutors more efficiently access and navigate video evidence.
Beam Notes is not a decision-making tool and does not automate, infer, or advise on case outcomes or charging decisions. It supports prosecutors during the evidence-review stage by helping them review video-recorded evidence more efficiently. Users manually upload evidential video recordings into the Beam Notes application, which generates a time-stamped transcript and structured summary (key details, such as names and dates or birth and a chronology of events). These outputs help prosecutors navigate recordings, identify relevant evidence and prepare for case review activities. Beam Notes is used as a review aid only and supports, but does not replace, review of the original video evidence.
3.2 - Human review
All Beam Notes summaries and transcripts must be reviewed by users before use in their case review or moving the content into the case management systems, which is reinforced during training. This “human-in-the-loop” design means outputs are reviewed before entering case systems and are never auto-populated. Users are trained and expected to correct errors, e.g. misspelled names, using the chat function within Beam Notes. Prosecutors must watch the video evidence and Beam Notes does not replace this requirement. A further human-in-the-loop mechanism is applied with users (CPS prosecutors) reviewing and scoring summaries via an in-app rating model, which provides immediate feedback on issues if they arise.
3.3 - Frequency and scale of usage
Beam Notes will regularly be used by CPS Prosecutors across all 14 areas across England and Wales as part of reviewing video evidence in approximately 11,000 cases per year, enabling consistent and efficient evidential review at scale.
3.4 - Required training
End‑users were actively involved in shaping Beam Notes, providing feedback on usability, accuracy, workflows, and template design. All users receive structured, mandated training before access is granted and written guidance, video demonstrations and support materials ensure consistent and safe use. Roles and responsibilities are clearly defined and reinforced during training.
3.5 - Appeals and review
The tool will not be used for decision making. Existing CPS policies and procedures for feedback, appeals and complaints continue to apply.
Tier 2 - Tool Specification
4.1.1 - System architecture
Beam Notes is a secure, cloud-based evidence-processing tool. CPS-uploaded files are stored in encrypted cloud storage and processed through an AI pipeline that produces time-stamped transcripts and structured summaries. A bespoke orchestration layer manages ingestion, transcription, summarisation and the application of CPS-designed templates. All data is encrypted in transit (TLS 1.2/1.3) and at rest (AES-256 standard), and is processed and stored within the UK or EEA. Data is retained according to the retention periods agreed with CPS as data controller; once a retention period is reached, deletion runs automatically. Backups are retained for 7 days, after which data cannot be retrieved. Sub-processors operate zero data retention, with the exception of Azure, which retains data for 30 days solely for misuse detection. Role-based access controls restrict access to authenticated CPS users and to a named, SC-cleared Beam project team. The architecture includes monitoring, audit logging and a defined support model to ensure secure, resilient national operation.
4.1.2 - System-level input
The inputs to Beam Notes consists of CPS-uploaded evidence video files via the secure Beam Notes application. These files may contain victim, witness or suspect statements and will therefore contain information disclosed by those individuals during the statement taking process. Inputs are received in standard digital video formats (e.g., MP4), accompanied by system‑generated metadata such as upload time, file name, and authenticated user credentials. The tool processes only the content uploaded by CPS users and does not capture or record any information itself.
4.1.3 - System-level output
Beam Notes generates a speech-to-text transcript and a structured summary of each evidential video uploaded by CPS staff. These outputs are accessed via the Beam Notes application. Users review transcripts and edit the summaries before manually copying the verified content into our case management system. The outputs are designed to support the early evidence‑review stage of the prosecution process by helping prosecutors more efficiently access and navigate video evidence. Beam Notes does not automate any decisions, and outputs are not used for automated processing by sub‑processors.
4.1.4 - Maintenance
Beam Notes is subject to continuous operational review and a formal evaluation at the end of the project period with monthly and indepth quarterly oversight. There is no local model re-training schedule, as AI models are managed externally and not retrained on CPS data. Security and compliance are maintained through regular oversight.
Beam holds enterprise agreements with all third-party AI model providers, which guarantee that no customer inputs or outputs are used for model training or product development, and that no data is retained by the provider after processing
4.1.5 - Models
Beam Notes uses Deepgram’s Nova 3 model, self-hosted by Beam within its own GCP infrastructure in the UK, for speech-to-text transcription. Summarisation uses EEA-based OpenAI GPT models accessed via the OpenAI API. Fallback models are OpenAI Whisper for transcription and an OpenAI GPT model for summarisation, both hosted within Beam’s Azure infrastructure in the EEA.
Tier 2 - Model Specification: Deepgram Nova 3
4.2.1. - Model name
Pre-trained. Deepgram Nova 3 (self-hosted) - This model is developed by Deepgram but is self-hosted by Beam on their Google Cloud Platform (GCP) infrastructure within a dedicated CPS-only GCP project.
4.2.2 - Model version
Nova 3 (latest version as provided by Deepgram)
4.2.3 - Model task
Speech-to-text transcription - The model converts audio recordings of conversations into text transcripts
4.2.4 - Model input
Video files uploaded via the Beam Notes web application
4.2.5 - Model output
The output is a text transcript of the video, including speaker separation (diarisation) and timestamps.
4.2.6 - Model architecture
Deepgram Nova 3 is a proprietary automatic speech recognition (ASR) model developed by Deepgram. It uses deep learning techniques optimised for speech-to-text conversion. For the CPS deployment, the model runs on servers/GPUs inside Beam-managed GCP infrastructure. It is not a public API call to Deepgram; the model is self-hosted, meaning all audio/video processing occurs within Beam’s controlled environment. No customer data leaves Beam’s infrastructure for this processing step. Further information: https://deepgram.com/learn/nova-3-speech-to-text-api
4.2.7 - Model performance
Deepgram Nova 3 has demonstrated industry-leading accuracy on standard speech recognition benchmarks. Each template developed for CPS is thoroughly tested by Beam’s prompt engineering team alongside CPS team members to ensure the transcription model produces the desired output consistently.
4.2.8 - Datasets and their purposes
Beam does not train, fine-tune or develop the Deepgram Nova 3 model. It is a pre-trained, off-the-shelf model that Beam self-hosts. No CPS data is used for training or improving the model. The model’s original training data is managed by Deepgram.
Tier 2 - Model Specification: OpenAI GPT models via the OpenAI API (EEA-based)
4.2.1. - Model name
OpenAI GPT models via the OpenAI API (EEA-based)
4.2.2 - Model version
Currently OpenAI GPT‑4/GPT‑4o is used for CPS Beam Notes. Models are selected per template based on testing as certain models perform better at certain tasks.
4.2.3 - Model task
Text summarisation - The model takes the transcript produced by the transcription model and generates a structured summary based on a template defined by Beam and the CPS team.
4.2.4 - Model input
The model input is the transcript of the video, and a system prompt containing template instructions that define the structure and format of the summary; this template is co-developed by Beam’s prompt engineering team and the CPS team.
4.2.5 - Model output
The output is a structured text summary of the transcript, formatted according to the CPS-defined template.
4.2.6 - Model architecture
OpenAI GPT models are large language models (LLMs) based on the Transformer architecture. They are pre-trained on large, diverse text datasets and are accessed via the OpenAI API. Further information: https://platform.openai.com/docs and https://trust.openai.com/
4.2.7 - Model performance
Beam evaluates summaries generated from transcriptions to verify that they include all factual details from the source. Beam employs LLM evaluation tools that alert engineers when accuracy scores fall below threshold. Beam engineers continually refine the configuration to ensure outputs are based solely on the content of the original source.
4.2.8 - Datasets and their purposes
Beam does not train, fine-tune or develop OpenAI’s GPT models. These are pre-trained, off-the-shelf models accessed via the API. No CPS data is used for training or improving the models. The models’ original training data is managed by OpenAI; details can be found at https://platform.openai.com/docs.
Tier 2 - Model Specification: OpenAI Whisper
4.2.1. - Model name
OpenAI Whisper - A model developed by OpenAI but is self-hosted within Beam’s Azure infrastructure.
4.2.2 - Model version
Whisper (latest stable version) is only used as a fallback if Deepgram Nova 3 is unavailable.
4.2.3 - Model task
Speech-to-text transcription - This is activated only if the default Deepgram transcription service is unavailable.
4.2.4 - Model input
Digital video files, same as the default transcription model.
4.2.5 - Model output
The output is a text transcript of the videos and the format is consistent with the default transcription model (Deepgram).
4.2.6 - Model architecture
Whisper is an open-source automatic speech recognition model developed by OpenAI. It uses a Transformer-based encoder-decoder architecture trained on a large dataset of diverse audio. Further information: https://openai.com/index/whisper/
4.2.7 - Model performance
Whisper is a widely benchmarked Automatic Speech Recognition (ASR) model with strong performance across multiple languages. It is used only as a fallback model, with the same accuracy monitoring and evaluation processes applied as for the default transcription model.
4.2.8 - Datasets and their purposes
Beam does not train, fine-tune or develop the Whisper model. It is a pre-trained, off-the-shelf model that Beam self-hosts. No CPS data is used for training or improving the model.
Tier 2 - Model Specification: OpenAI GPT model hosted in Beam’s Azure infrastructure (EEA)
4.2.1. - Model name
OpenAI GPT model hosted in Beam’s Azure infrastructure (EEA)
4.2.2 - Model version
Azure-hosted GPT model (latest stable version deployed by Beam), used only as a contingency.
4.2.3 - Model task
Text summarisation (contingency) is activated only if the default OpenAI API summarisation service is unavailable.
4.2.4 - Model input
Text transcript from the videos uploaded by CPS prosecutors and system prompt, same as the default summarisation model.
4.2.5 - Model output
The output is a structured text summary, consistent with the default summarisation model output.
4.2.6 - Model architecture
This is an OpenAI GPT model hosted within Beam’s Azure infrastructure, with the same protections in place for data retention and model training as the default summarisation model.
4.2.7 - Model performance
The same evaluation and monitoring processes are applied as for the default summarisation model.
4.2.8 - Datasets and their purposes
The same evaluation and monitoring processes apply as for the default summarisation model.
2.4.3. Development Data
4.3.1 - Development data description
Beam does not train or develop its own AI models. All models used in the CPS deployment are pre-trained by their respective providers (Deepgram, OpenAI). No CPS data is used for model development. Beam pays for enterprise access to these models, which contractually guarantees that no inputs or outputs are used for training or product development by the model providers.
4.3.2 - Data modality
N/A - The underlying models are trained by their respective providers on diverse datasets including text, video and audio. Details are available from the respective model providers.
4.3.3 - Data quantities
N/A - As no CPS data is used to train or develop models, the training data volumes are managed by the respective model providers.
4.3.4 - Sensitive attributes
N/A - Model providers have their own processes for handling sensitive data in training pipelines.
4.3.5 - Data completeness and representativeness
N/A - Beam relies on established, large-scale commercial AI models that have been pre-trained on extensive and diverse datasets by their respective providers. Beam supplements this by conducting its own evaluations for accuracy and bias. CPS also conducted their own evaluations throughout the pilot to assess accuracy, bias and efficiency, and continue to do assurance checks regularly post-deployment.
4.3.6 - Data cleaning
N/A as Beam does not train or develop models.
4.3.7 - Data collection
N/A - The third-party models are trained by their providers on large, diverse datasets. Details are available from the respective model providers at request.
4.3.8 - Data access and storage
N/A - Beam does not store or have access to model training data. All model training data is managed by the respective model providers.
4.3.9 - Data sharing agreements
N/A - Beam does not share data for model development purposes. Beam’s enterprise agreements with providers explicitly prohibit the use of customer data for training or product development.
Tier 2 - Operational Data Specification
4.4.1 - Data sources
Data sources for the CPS deployment of Beam Notes include: 1. Video evidence: CPS prosecutors upload recorded video evidence via the Beam Notes web application, which is uploaded securely to a CPS‑dedicated GCP environment. 1. User account data: CPS user identities (names and email addresses) are authenticated via Kinde using CPS’ Entra ID single sign‑on. CPS credentials are entered on a CPS‑controlled page and are not handled by Beam. 1. User interactions: CPS Prosecutors review transcripts, refine summaries using AI editing features, and query transcripts via the chat interface to support evidence review. 1. Template instructions: Summary templates and system prompts are co‑designed by CPS and Beam to define and develop the required structure and format of outputs.
4.4.2 - Sensitive attributes
Due to the sensitive nature of CPS casework, all data processed by Beam Notes is treated as sensitive by default. This may include personal data (such as names, contact details and CPS staff information), special category data (including health, racial or ethnic origin, sexual orientation, religious or philosophical beliefs, and other sensitive disclosures), and criminal offence data relating to investigations and proceedings. All data is protected through strong technical and organisational measures. Data is encrypted at rest (AES‑256) and in transit (TLS 1.2/1.3) and processed within a fully segregated CPS environment, including a dedicated GCP project, VPC, database, web application and object store. Access is restricted to a named, SC‑cleared Beam project team, with no access granted to other Beam staff. Role‑based access controls ensure prosecutors can only access their own recordings and summaries, with alerts triggered for any role or access changes.
4.4.3 - Data processing methods
Video processing: Video is uploaded by CPS Prosecutors in a supported format (e.g. MP3, WAV, MP4 etc.) to Beam Notes which is hosted in Beam’s CPS‑dedicated GCP infrastructure. Transcription: Video is transcribed using the self‑hosted Deepgram Nova 3 model within Beam’s GCP environment, producing a timestamped, speaker‑diarised transcript. If unavailable, transcription falls back to a Whisper model hosted in Beam’s Azure environment. Summarisation: The transcript is sent to the OpenAI API with the CPS‑approved template instructions to generate a structured summary. If the OpenAI service is unavailable, an Azure‑hosted GPT model is used as a fallback. No additional pre‑processing or data transformation is applied beyond this transcription and summarisation pipeline.
4.4.4 - Data access and storage
Beam Notes processes sensitive operational data, including video recordings of evidential interviews, transcripts, and structured summaries. Access is restricted to authorised CPS Prosecutors through secure authentication, while administrative access is limited to a named, security cleared Beam project team. Data is retained within the CPS dedicated GCP project in the UK, in line with CPS defined retention periods (30 days) and is deleted automatically once the agreed period expires. Data is retained for 7 days in the backup and then entirely irretrievable. Reviewed outputs are manually transferred into the CPS’ Case Management System (CMS) by Prosecutors in accordance with CPS retention policies. Data is protected through encryption, strict access controls, and contractual restrictions on sub processors including ISO27001 and CyberEssentials Plus certification, to ensure data security and appropriate use. Beam is responsible for storage and security of operational data during processing. CPS is the data controller and responsible for defining retention periods and data management.
4.4.5 - Data sharing agreements
Data sharing arrangements are in place for Beam Notes between the CPS and Beam. Data is shared only with Beam and its approved sub processors for the sole purpose of delivering the Beam Notes service. For the CPS deployment, sub processors include Google Cloud Platform (UK) for hosting, storage and databases; Azure (EEA) for fallback transcription and summarisation with limited 30 day retention for misuse detection; Fastly for content delivery with no data retention; Kinde (EEA) for user authentication for the duration of the contract; and OpenAI (EEA) for summarisation with zero data retention. All sub processors are contractually restricted to processing data solely for Beam Notes, with no use for model training or product development, and all processing occurs within the UK or EEA. Any changes to sub processors are subject to CPS approval, with a minimum of 30 days’ advance notice, and no data is shared under the Digital Economy Act.
Tier 2 - Risks, Mitigations and Impact Assessments
5.1 - Impact assessments
The Data Protection Impact Assessment (DPIA) for the CPS Beam Notes initiative was completed and approved on 16th December 2025. The DPIA identified privacy, security and AI ethics risks associated with processing evidential video content, and set out key mitigations including strict role‑based access controls, contractual limits on data use by Beam and its sub‑processors, and mandatory user training. The DPIA is overseen by the CPS Data Protection Officer (DPO), with ongoing monitoring in place and a full evaluation planned following the national deployment period. The DPIA is an internal document and is not published online.
5.2 - Risks and mitigations
Supplier Access - There is a risk that Beam engineers may access sensitive CPS information and that supplier staff may not always have the appropriate security clearances in place. This is mitigated through a strict CPS controlled access request process, with all core Beam personnel already security cleared to the appropriate levels, and any new joiners are prevented from accessing CPS data until their security clearance is complete. The risk is accepted but actively monitored to ensure supplier access remains appropriate, justified, and compliant with CPS security requirements.
Data Moving Outside the CPS Tenant - Beam Notes processing requires CPS data to leave the CPS tenant, creating risk around external handling. This is mitigated through continuous monitoring, completion of the CPS architecture review, removal of unnecessary system components, and a commitment to fully address hosting options in the long term business case. Additionally, security controls are in place for the supplier infrastructure, including Cyber Essentials Plus accreditation, contractual restrictions, and all data is processed in the UK/EEA.
Inaccurate or Unfair Outcomes - Incorrect transcriptions or summaries could result in miscommunication, unfair outcomes, or misrepresentation. In mitigation, Beam report 98% transcription accuracy which is currently being monitored in CPS, templates are co-designed and tested, and staff must check all outputs for accuracy before they are used.
Expansion risks: As Beam Notes expands, inconsistent use could increase the risk of data breaches or poor information quality; this is mitigated through defined governance and oversight, mandatory training, application of the CPS Responsible AI Policy, audit‑log monitoring, clear operational guardrails, and open feedback mechanisms to ensure compliance and continuous improvement.
Over reliance on Beam Notes: There is a risk that users may become overly reliant on the tool, particularly the summarisation feature, which could lead to misuse or reduced scrutiny of outputs. This is mitigated through robust training to build user confidence, close monitoring to identify errors or misuse, mindful friction embedded within the tool for reminders, reinforcing the CPS Responsible AI policy, ongoing evaluation through feedback and reinforcing the requirement for human oversight at all stages
Transcription Accuracy for Diverse Speech Patterns: There is a risk that transcription accuracy may be lower where recordings contain strong accents, English spoken as a second language, speech impediments, background noise, multiple speakers, or interpreter-mediated conversations. This could result in errors or omissions within transcripts and summaries. This risk is mitigated through mandatory human review of all outputs, the requirement for prosecutors to review the original video evidence, ongoing monitoring of transcription performance, user feedback mechanisms, and the ability for users to identify and correct inaccuracies before outputs are used in casework.