Demystifying clinical AI in mental health

Key terms related to AI and examples of their use:

Key considerations and questions for suppliers and your organisation when selecting an AI tool.

Products are medical devices when they meet the definition of a medical device as set out in the UK Medical Devices Regulations 2002 (SI 2002 No. 618, as amended) (UKMDR). AI and other forms of software that have a medical purpose will fall within this definition and qualify as a medical device.

Products classified as a medical device need to meet requirements of this regulation, which include (but not limited to):

• risk management system to ensure risks are reduced As Far as Possible (AFAP) as described in ISO 14971.
• clinical evidence to show benefits outweigh risks
• post-market surveillance system to ensure the device continues to work as intended.

When accessing and using patient data, all parties need to ensure this is used in a secure way that does not compromise patient safety. Further information on data privacy and security legislation can be found on the Gov.uk website.

Medical devices are classified based on risk and intended use and generally categorised into four main classes (Class I, IIa, IIb and III), with Class I being the lowest to Class III being the highest risk, according to the UK MDR 2002. 

• Class I: Requires manufacturers to self-declare conformity with the essential requirements of the UK MDR 2002 and registration on the MHRA portal
• Class IIa: Requires conformity assessment completed by a UK Approved Body and registration on the MHRA portal
• Class IIa, IIb: Requires conformity assessment completed by a UK Approved Body and registration on the MHRA portal
• Class III: Requires conformity assessment completed by a UK Approved Body and registration on the MHRA portal

Conformity assessments requires manufacturers to demonstrate compliance to the Essential Requirements set out in the UK Medical Devices Regulations (MDR) 2002 and must be completed by a UKAS-accredited UK Approved Body. 

A UKCA (or UKNI) or CE mark shows that a digital mental health device is compliant with relevant medical device regulations. MHRA registration shows the device can be placed on the GB market. This register is open to public view via the Public Access Registration Database.

Read further information on the regulation and evaluation of digital mental health technologies from the MHRA

In Central and North West London NHS Foundation Trust, a trust-level design authority has been set up to oversee and encourage the effective use of AI within the trust. The group meets bi-monthly and has representation from senior clinical and operational leaders from all the divisions and departments within the trust. 

In one trust, the finance director and chief clinical officer are both board sponsors for AI. This enables both a financial and technical view when considering the adoption of AI, and early financial involvement in conversations on AI roll out and testing. 

Bradford has transformed its talking therapies services by introducing AI-powered tools that increase access, personalise care and optimise clinical capacity. The trust’s talking therapies service implemented Limbic Access, an AI-powered clinical assessment tool (chatbot) and Limbic Care, a clinical AI companion app, which streamlined triage processes, reduced wait times and improved engagement. 

What was the challenge the organisation needed to address?

Bradford and Craven District talking therapies wanted to increase the number of referrals to their services and improve access for groups traditionally underrepresented by NHS Talking Therapy services including people from minority ethnic backgrounds, the LGBTQIA+ community and older adults. They already had an existing self-referral form on their website, but it led to numerous dropouts and incomplete submissions. 

Before implementing AI-powered clinical assessment tools, the service received approximately 6,900 referrals over six months. 

Bradford also wanted to further integrate clinical AI tools into their workflows to improve patient engagement in one-to-one sessions and groups. However, there was staff scepticism about clinical AI adoption and concern about loss of control of therapeutic sessions and whether the technology would help or complicate established practice.

“I started off quite hesitant …but actually over time it really does benefit the sessions… it guides the patient a lot easier and just gives that extra guidance outside of the sessions as well”. Psychological Wellbeing Practitioner, Bradford and Craven District Talking Therapies.

How did they address it?

Bradford District and Craven NHS Talking Therapies introduced two new AI tools. 

Limbic Access made it easier for people to seek help across the right services. Bradford sought to reduce dropout rates and reach groups such as people from minority ethnic backgrounds, the LGTBQIA+ community and older adults. 

Limbic Care was also introduced to deliver personalised therapeutic materials to patients. The tool acts as a personal clinical assistant engaging patients between therapy sessions and on waiting lists and complements therapy sessions, by helping people better understand their mental health difficulties, access resources in their own time and complete therapeutic exercises between appointments. 

What successes are they seeing so far?

Implementing Limbic Access has increased referral rates and saved clinician time:

• Since launch in 2023, more than 16,700 referrals have been processed. This includes rises in key demographics including people from minority ethnic backgrounds and LGBTQIA+ referrals.
• 39 per cent of these were made out of office hours, supporting the working unwell
• 88 per cent referrals were signposted to appropriate support, with early crisis detection
• There has been a 90 per cent conversion rate from referral to assessment
• Approximately 1,000 clinical hours saved
• 94 per cent of patients reported positive feedback

Implementing Limbic Care has shown measurable improvements across key areas: 

• Administrative efficiency with the support of the clinician dashboard that provides AI-generated summaries of activities the patient has completed, realising clinical admin time to allow for more targeted sessions.
• Enhanced therapeutic alliance with full visibility of a client’s homework completion and risk events. This allows practitioners to identify specific implementation barriers and tailor techniques to individual clients.
• Patient engagement beyond session with an AI companion exceeded expectations. Many report supportive and stigma-free conversations with the app which means patients are supported 24/7. 77 per cent of clients return after 7 days and 46 per cent remain active by day 30. The tool enables long term support with access for 12-months post discharge.
• 86.6 per cent of patient users report positive satisfaction with the tool.
• 70 per cent of clients attend multiple sessions, compared to 55.7 per cent not using the app.

Learn more about Bradford and Craven District talking therapies experience with AI tools. 

Rotherham, Doncaster and South Humber NHS Foundation Trust operates services in more than 100 locations across Rotherham, Doncaster and North Lincolnshire. The trust employs over 3,700 staff. 

What was the challenge the organisation needed to address?

The trust identified that clinicians spend 40-60 per cent of their clinical time on administration. 

How did they address it?

The trust tested AVT products within existing clinical pathways to summarise consultations in community mental health, neurodiversity services and inpatient mental health. The products went through a governance process supported by data protection impact assessments, regulation and digital clinical safety. An evaluation plan was also developed to enable real-world and controlled testing.

What successes are they seeing so far?

Early results from the pilot showed that:

• 82 per cent of colleagues feel they benefited from AVT
• 76 per cent of colleagues reported that using AVT saved them time
• participating clinicians strongly agreed to continuing to use AVT in clinical notes
• between 89 per cent and 96 per cent of patients gave consent for AVT use in their consultations (dependent on teams), though some patients declined use, citing concerns about the environmental impact of AI
• clinicians and patients benefit from the greater presence of the clinician at the consultation.

The pilot also identified some barriers to AVT roll out including infrastructure issues related to internet connection, sound recording and laptop battery issues and issues related to the tool’s understanding of accents. In some cases, the technology was seen to interrupt the therapeutic relationship, and some questions were raised over the accuracy of the output with 61-80 per cent of pilot participants reporting the output was correct without editing. The pilot highlighted the need for awareness of automation bias, and the need for clinicians to take full responsibility for checking what’s documented.

Key learnings for deployment from the pilot included the need for close working with an engaged supplier, the benefit of clinical champions to support adoption, and the need for gradual roll out to ensure supplier and organisational colleagues don’t spread their resources too thinly and any barriers to adoption can be addressed.

Living Well Consortium (LWC) comprises more than 30 mental health care organisations, charities and social enterprises in the UK. They are the leading provider of NHS Talking Therapies in Birmingham and Solihull. Given the breadth of their services, by 2023 Living Well Consortium was processing more than 11,000 referrals a year: a 400 per cent increase over 2019. 

What was the challenge the organisation needed to address? 

Living Well Consortium needed to increase capacity without overburdening their workforce. In reviewing their intake pathways, they found four key focus areas:

1. Increase capacity for demand- Increasing patient volume was resulting in longer and longer waiting lists. Meanwhile clinical staff were reporting that they felt rushed during their assessments.
2. Optimise data collection- LWC wanted to improve the quality and scope of data collection, since only 35 per cent of new patients on average responded to key demographic questions at intake.
3. Assess risk at self-referral- Administrative staff did not have clinical training and were therefore not taking risk assessments at the point of self-referral. As a result, there was a chance of at-risk patients sitting on waiting lists.
4. Improve client experience- Clients were frustrated with the long wait times to self-refer via the telephone, and the inconvenience of calling during work hours. LWC aimed to make the self-referral process easy and accessible, making clients feel supported from the start. 

How did they address it?

To help manage growing demand alongside existing phone and professional referral systems, LWC introduced an AI-powered clinical assessment assistant (chatbot) in October 2023. 

What successes are they seeing so far?

• Reduced clinician workload: In their first year of activation, the clinical assessment assistant processed 42 per cent of Living Well Consortium’s self-referrals. This enabled LWC to increase capacity while reducing demand on frontline staff. Prior to adopting the AI tool, LWC couldn’t offer self-referrals outside of working hours. Now, 40 per cent of referrals happen after hours, indicating that LWC is lowering barriers to access.
• Richer patient data: Where only 35 per cent of new patients responded to key demographic questions prior to the roll out of the clinical assessment tool, that figure now stands at 98 per cent. Clinicians have a more complete picture of each individual before their first session. And LWC now better understands the communities it serves.
• Faster and more appropriate care: LWC developed a tailored screening and signposting process that flagged 12 per cent of incoming patients for risk, prioritising them in the system, while also directing ineligible patients to suitable services much earlier.
• Improved patient experience: Shorter waiting lists, earlier prioritisation of at-risk patients, better data capture, faster and more personal clinical assessments, after-hours pathways into care. In the first six months, LWC has seen a 14 per cent decrease in patient dropouts.

Find out more about the work at Living Well Consortium.

A multi-site evaluation explored the use of contactless patient monitoring in reducing patient harm across 29 inpatient mental health wards in six NHS mental health trusts. Contactless patient monitoring employs computer vision, interpreting visual information with advanced algorithms to provide proactive information to staff so they can intervene earlier to prevent harm occurring. Results varied by ward, but most showed decreases in event rate following adoption of contactless monitoring with meta-analyses indicating notable reductions across all five outcome measures ranging from -21.1 per cent to -28.9 per cent. (events included self-harm on acute wards; falls on older adult wards and assaults and restraints in PICUs) 

What is contactless patient monitoring? 

Contactless patient monitoring supports clinicians to deliver better, safer inpatient care. Advanced signal processing algorithms can derive medical-grade vital signs and sleep insights using the same underlying techniques, photoplethysmography and actigraphy, as contact monitors. Machine learning and computer vision predictive AI techniques can also be used to translate patient situational information into proactive safety alerts and ambulatory trends. For example, identifying when a patient is leaving bed to help head off falls, or when a patient has spent an extended period of time in a higher risk area.

South London and Maudsley NHS Foundation Trust (SlaM) provides the widest range of NHS mental health services in the UK. The organisation serves a local population of 1.3 million people in South London, as well as specialist services for children and adults across the UK and beyond. The organisation employs over 7,000 people.

SlaM is taking an organisation-wide approach to AI, exploring and testing the potential of AI for both organisational processes and clinical care. The initiative combines a focus on creating a connected digital ecosystem to ensure organisational readiness for adoption, with the introduction of intelligent tools like Microsoft Copilot agents, automation and exploring innovative technologies such as ambient voice technology. The aim of this approach is to make everyday tasks easier, streamline processes, and free up time for what matters most: delivering excellent services and outcomes for those that use the services. 

A number of use cases are being tested:

To streamline internal processes, SlaM is developing an organisation-wide digital ecosystem with the creation of a single seamless user interface for staff that brings together internal systems and ensures staff have streamlined access to all internal processes such as finance, HR, estates and facilities, and learning and development . This will also include AI agents and automation solutions to support use of the system and reduce pressure on staff. 

Work is also underway at SlaM to explore the role of agentic AI and pilot this ahead of a further roll out. This includes a pilot, in collaboration with Microsoft, focusing on HR and finance functions to build, test and deploy agents such as job evaluation, report writing, finance and risk assessments in live environments. Early work is also underway to consider the role of AVT within community, CAMHS and inpatient settings. 

SlaM has also been using Microsoft 365 Copilot for over two years as an early adopter and as part of a national Microsoft pilot to automate routine tasks such as summarising emails, meetings and sense-checking documents while ensuring that the human always remains in the loop and responsible for outputs using the technology. In December, the trust launched its AI and automation policy setting out its ethical approach to AI adoption. In March, SlaM launched an organisation-wide training programme on using Microsoft Copilot chat, delivered jointly by Microsoft, communications and digital services. Over 1,000 colleagues took part in one of 28 training sessions across foundational, advanced and drop-in formats and staff feedback showed strong interest in practical applications of AI tools and enthusiasm for continued support. Alongside online delivery, the communications and Microsoft team engaged 116 ward-based staff across all four hospital sites, helping to reach colleagues who may not regularly access email-based communications. Ten training videos have been published on the intranet with further content being reviewed for release.  

Overcoming barriers to adoption 

To ensure responsible deployment of tools and alignment with patient and staff needs, SLaM is focused on creating an organisation-wide digital ecosystem through data readiness to ensure interoperability between data systems, staff training, governance and assurance processes and patient engagement. This approach includes: 

• Building staff confidence: To manage varying levels of staff confidence around AI, SLaM recently held an organisation-wide digital takeover week focused on AI training and development focussed on supporting staff to use Copilot chat.
• A robust approach to governance: To ensure a robust approach to governance, SLaM has set up an AI governance group. This brings together colleagues from across the organisation including clinical and non-clinical roles. The AI governance group feeds in to the digital board which reports into the finance, performance and investment committee and then on to the trust board.
• Data readiness and interoperability have been identified as key enablers of effective AI roll out. The organisation is undertaking a programme of work on data readiness, to address existing challenges with unstructured data and to ensure interoperability between systems and data. This programme of work will allow other use cases to be explored. 

Trusts are taking different approaches to staff upskilling including weekly ‘copilot cafés’ to share experience of using Copilot, weekly training sessions and questionnaires at the end of pilots to gather feedback. 

One trust advised that showcasing the benefits of AVT changed clinical perception over its use and addressed concerns about increased workloads in adopting new tools.

At Bradford District Care NHS Foundation Trust, staff training on new tools helped overcome uncertainty on the benefit of AI tools.

At South London and Maudsley NHS Foundation Trust, ensuring the patient and carer view in AI decisions is a crucial part of building confidence in AI tools. Representatives from the patient and carer reference group have been invited to attend the AI reference group and then feed back to their wider group. This ensures patient and carer insights are reflected in decision-making and their concerns and expectations about AI technologies are considered. The trust will continue to work with patients and carers throughout the AI transformation journey. 

Evaluation from an AVT pilot in another trust considering patient consent found that whether consent is given or not is affected by the confidence of the clinician in explaining the technology. This also found that while there were variable reasons for declining the use of AVT which require further research, some patients are choosing to decline as a result of concerns about the environmental impact of AI.  

Everyturn Talking Therapies is part of the NHS Talking Therapies service and provides specialist services on behalf of the NHS including talking therapies, crisis support, dementia care, services for children and young people, specialist nursing and hospital step-down, and community wellbeing support. 

What was the challenge the organisation needed to address?

In 2018, the Government Equalities Office issued a report exploring the experience of LGBTQ+ people in the UK. Over 108,000 LGBTQ+ individuals completed the survey. With a focus on mental health, the survey found that 24 per cent of respondents had access mental health services in the 12 months preceding the survey and 28 per cent of those who had accessed or tried to access services, said it had not been easy at all. The most commonly cited reason was waiting lists, given by 72 per cent of respondents and around a fifth (22 per cent) said their GP was not supportive. 

How did they address it?

Everyturn Talking Therapies used an AI-powered clinical assessment assistant as a digital front door for people self-referring into the service. It supports patients to be signposted or referred to services sooner in the process, reducing duplication of assessments and delays for more complex or high-risk patients.

What successes are they seeing?

‘Charlie’ (pseudonym), a non-binary individual, embarked on their talking therapy journey using the clinical assessment assistant within the Everyturn Talking Therapies service. They first encountered the clinical assessment assistant on the Everyturn website, and it allowed them to self-refer by expressing their emotions and thoughts freely. 

At the point of referral, Charlie indicated they were struggling with ‘extreme stress and anxiety’ and chose anxiety as their primary therapeutic focus. The unique insights and information gathered enables the service to review initial PHQ9 and GAD7 scores before assessment, helps to monitor any potential risk concerns promptly, and allows patients to choose and book their assessment at the end of the referral. This has positively increased access to the services and created a streamlined pathway for patients.

The clinical assessment assistant supports patients to be signposted or referred to other services sooner in the process, reducing duplication of assessments and repeating their stories. It reduces delays for more complex or high-risk patients accessing appropriate care via triage from the internal risk team within the service who liaise with local crisis services. This also means assessors can focus on talking therapy appropriate assessments, which can reduce burnout for staff. 

Charlie’s primary presenting problem and other information gathered at referral were then confirmed in their assessment with a psychological wellbeing practitioner (PWP). The information allowed the practitioner to focus on Charlie’s specific issues from the start. 

In collaboration with Charlie, a treatment plan was agreed, and they were offered low intensity computerised cognitive behavioural therapy (cCBT), supported by a PWP, through the Space from Generalised Anxiety Disorder programme on Silvercloud. 

Charlie completed their cCBT with regular reviews with a PWP. Everyturn’s cCBT review calls and email content is always personalised to the specific issues and content shared by the patient. All correspondence with Charlie was personalised and focused on their specific challenges with a focus on supporting change and acceptance. They were discharged from service in recovery. Charlie’s scores changed from moderate PHQ9 score and Severe GAD7 score at referral to non-clinical scores on both PHQ9 and GAD7 at the end of treatment. They reported that the ‘therapy was starting to help to get down to the root of my problems, so I feel less anxious about situations than usual.’ At the end of their treatment. Charlie was able to apply for jobs and go to job interviews which they had not been able to achieve before. They also had ongoing access to their cCBT programme for 12 months following the completion of treatment. 

All of this was achieved with Charlie referring and completing their assessment within seven days. This allowed them to access appropriate therapy promptly, minimising the escalation of their symptoms and helping them to stay focused on their goals. They were discharged from the service within two months of referral in recovery.

Further information on this case study can be found in NHS Talking Therapies for Anxiety and Depression: LGBTQ+ Positive Practice Guide (2024) (p54).

In a real-world study published in Nature Medicine involving 129,400 patients across 28 NHS Talking Therapies services in England, it was observed that services implementing Limbic’s clinical assessment assistant experienced a significant 179 per cent increase in non-binary individuals seeking support (Habicht et al., 2024). Using qualitative feedback from individuals, the study found that non-binary individuals value the human-free aspect of the tool more, highlighting the importance of alternative and inclusive self-referral methods for LGBTQIA+ individuals. (Habicht et al. 2024).

• The AI Ambassador’s Network is a forum for healthcare professionals and those interested in AI in healthcare to learn about and share practical applications of AI. The network supports learning, capability building and delivery of AI across health and care systems, whilst encouraging safe, inclusive spaces for members to be involved and influence AI embedding locally and nationally. The network is led by the Joint Policy Unit, NHS England and DHSC and supported by NHS Horizons. 

The National AI and Mental Health Community of Practice provides an opportunity to come together to learn about and shape the responsible, ethical and effective use of AI in mental health. 

• Cersi-AI: The AI Readiness Checklist - The Centre of Excellence for Regulatory Science and Innovation in AI and Digital Health (CERSI AI) in collaboration with University of Birmingham and University Hospitals Birmingham has developed an AI Readiness Checklist which aims to support safe, effective and equitable AI adoption across the UK Health and Care system.
• Health Innovation Oxford and Thames Valley – Artificial Intelligence and health inequalities: a summary of the evidence.
• MHRA: Regulating Medical Devices in the UK- Guidance on medical device regulation in the UK.
• MHRA: National Commission into the Regulation of AI in Healthcare- this page will provide further updates on the work of the Commission to provide recommendations for a new regulatory framework for AI in healthcare.
• MindEd Medicines and Healthcare Products Regulatory Agency (MHRA): Digital Mental Health Technologies (DMHT). A series of resources for the public, parents, carers and professionals who use or recommend digital mental health tools.
• NHS England: A Buyer’s Guide to AI in Health and Care, developed by NHS Transformation Directorate in 2020 contains additional questions and detailed answers. The majority of the information is still relevant in the 2026 environment; however, elements of the regulatory environment section may not reflect the current situation. This document is due to be updated.
• NHS England: All Adopters' Guidance lists all the guidance and regulations that apply to adopters of digital technologies in health and social care.
• NHS England: AI Knowledge Repository brings together all available resources, guidance, and case studies on AI adoption.
• NHS England: Adopting ambient scribing products in health and care settings. Guidance and resources to support the adoption of AVT in the NHS including the NHS England Ambient Voice Technology Supplier Registry, guidance for NHS executives and boards, guidance for CIOs and CCIOs and links to communities of practice and additional support.
• NHS England: Understanding regulations of AI and digital technology in health and social care. Provides information on regulatory requirements for adopters and developers.
• NHS Providers: Navigating AI- practical questions for NHS leaders to ask- A toolkit to support conversations at each stage of the journey from considering using AI through to implementing solutions into business-as-usual workflows.
• UK Parliament POST – POSTnote 737 AI and mental health care: opportunities and delivery considerations. A literature review of AI in mental health care, bringing together use cases and considerations for AI adoption.