Limbic Access: An AI Self-Referral Entry Point for NHS Talking Therapies

QiaoLink Rating · Public Report

The QL Score is shown as a range because the current evidence does not allow us to pin it to a single number. The ends of the range are rounded to whole numbers. The rating is set by comparing the midpoint of the range (62.4) against fixed thresholds; the midpoint is calculated from the unrounded raw scores. The midpoint is only used to assign the rating band; it is not a "true, precise score".

Current recommendation

• For NHS Talking Therapies services: it can be used as one of several referral routes, provided that the service:

  • keeps other routes open, such as phone and online forms;

  • does not treat it as a risk-assessment or crisis-triage tool;

  • does not allocate treatment directly on the basis of its "diagnosis predictions";

  • takes part in the collection of outcome data that NICE requires.

• For people seeking help: you can use it to refer yourself. It is easier than phoning, and it is available in the evening too. But it is not treatment, and it is not a way to get help in a crisis.

Key Rating Drivers

• NICE, the body that assesses health technologies for the NHS in England, allows the NHS to use it under conditions for 3 years while more evidence is collected.

• Current evidence shows that it helps more people take the first step towards getting help, and that people who use it are highly satisfied.

• However, there are still no independent controlled studies of how accurate the information it collects is, how much time it actually saves, or whether it is worth the money once indirect costs are included.

• Some of the vendor's marketing claims, such as the claim that it doubles recovery rates, come from the developer's own observational study, and the authors acknowledge that other factors cannot be ruled out.

What it is

In England, if you are struggling with anxiety or depression, you can ask for help from the NHS's free psychological therapy service, NHS Talking Therapies. Until recently, this usually meant phoning, filling in a form, or asking your GP to refer you.

Many services now show a chat window on their website. That is Limbic Access. It asks you questions step by step, much like a text conversation:

1. whether you are eligible for the service, for example your age and which GP practice you are registered with;

2. how to contact you, and what you are struggling with;

3. several standard mental health questionnaires commonly used in the NHS.

Based on your answers, it works out which common problem you may have, from 8 categories including depression, anxiety, panic, obsessive-compulsive disorder (OCD) and post-traumatic stress disorder (PTSD), and then adds the relevant questionnaires. Finally, it produces a report for the service, which a clinician uses as background at your first assessment.

It is certified in the UK as a Class IIa medical device, a relatively high class among comparable chatbots. To be clear:

• it is registered with the Medicines and Healthcare products Regulatory Agency (MHRA);

• but the MHRA states that registration "does not represent any form of accreditation, certification, approval or endorsement by the MHRA";

• the Class IIa conformity assessment is carried out by a UK approved body.

According to NICE, it is "used by about 40% of NHS Talking Therapies services".

Is the problem worth solving?

Yes. Many people get stuck at the first step of asking for help: they do not want to talk about their situation over the phone, and they cannot reach the service during working hours. Demand for psychological therapy in England is high and waits are long. Patient experts who took part in the NICE assessment said that delays "can negatively affect relationships, employment, and mental wellbeing (for example, increasing suicidal thoughts)".

An entry point that is available at any time, and does not require speaking to anyone, genuinely lowers the barrier.

How strong is the evidence that it works?

Bringing more people forward for help: supported by evidence, and recognized by an independent body.

• A 2024 study by the developer, published in Nature Medicine, compared 14 services that used it with 14 similar services that did not. Referrals rose by 15% in the services using it, against 6% in the others.

• The increase was larger in some minority groups: 179% for non-binary people and 29% for people from ethnic minority backgrounds. However, the paper's supplementary data show no significant difference when the results are broken down by sexual orientation.

• The independent assessment group that evaluated the technology for NICE found that "the strongest evidence related to accessibility and overall satisfaction" with Limbic Access; satisfaction among users who completed a referral was 89% or higher. However, the referral data it relied on came from the same developer study described above.

• An independent academic called it "a well-considered study, with convincing findings", but noted that the reasons behind the results need further research.

• A US digital psychiatry expert who was not involved in the study pointed out that an interactive chatbot and a static web form are very different ways of collecting information. If it becomes easier to bring people in, "it does beg the question of what type of services are we going to be offering people, and how do we allocate those services?"

Saving clinical time: early signs, but not yet certain.

• Another study by the developer found that, for people referred through it, the first assessment took 12.7 minutes less on average (about a quarter shorter).

• One NHS trust reports saving about 15 minutes of staff time per referral on average.

• In a 2026 survey by the developer, 131 clinicians rated their own experience and said assessments felt less demanding when they had the information it collects. However, the survey was funded by the developer and relies on clinicians' self-reports, and the authors acknowledge possible recall bias.

• But NICE concluded that "the net time saved remains uncertain". The independent assessment group also stated: "No relevant information on quality and accuracy was identified" for the information collected. Any time saved would go into better assessments, and would not mean more patients seen per clinician.

Doubling recovery rates: the evidence does not support this claim.

• The claim comes from an observational study by the developer: the recovery rate was 58% for people referred through it, against 27.4% for other routes.

• But the study authors themselves acknowledge that people were not randomly assigned, so other factors cannot be ruled out. For example, people willing to complete the questionnaires patiently may be more likely to recover anyway.

• NICE also states that technologies of this kind "are not expected to directly affect service user outcomes".

"92% diagnostic accuracy": outside the scope of the independent assessment.

• Under the NHS Talking Therapies manual, "problem descriptors are not allocated until a full clinician-led assessment has taken place".

• NICE therefore specifically excluded the "diagnosis prediction" function from its assessment. The accuracy figure comes only from the vendor.

Points to note

• The authors of the effectiveness studies above are employed by, or paid consultants to, the developer, and hold shares in the company. The studies were peer reviewed, but they are not independent research.

• NICE also noted that most of the available evidence had no comparison group and could not be pooled, and stressed "the importance of independent, externally reviewed research".

• The developer registered a randomized, double-blind trial comparing versions with and without AI. A 2025 NICE document states: "The study ended in December 2024, but there are no publicly available results yet." No results have been published to date.

• A separate large observational study in Essex was completed in 2025, and its results have not been published either.

• The absence of published results does not mean the results are poor; they simply cannot be seen yet.

Is it safe? Could it miss someone in danger?

It is not a crisis tool or a risk-assessment tool. The vendor's Instructions for Use say so clearly.

• It can recognize commonly used phrases such as wanting to harm yourself, immediately show helpline numbers (999, 111, Samaritans and others), and alert the service.

• But the Instructions for Use acknowledge that "the device's capacity to identify risk is limited and relies on you self-identifying as at risk".

• Risk assessment remains the responsibility of the service's clinicians.

• The NICE committee concluded that technologies of this kind are "low risk to use with further evidence generation". But it also warned that AI "may selectively present information to healthcare professionals", which in some circumstances could increase the risk of an inaccurate diagnosis.

We found no reports of harm caused by the product. The vendor's website lists a Field Safety Notice numbered FSN-2025-001, but the document is set so it cannot be downloaded, and we were unable to read its contents. A Field Safety Notice is the formal document a medical device manufacturer uses to tell users about a safety-related corrective action; its existence does not mean that harm has occurred.

If you are in crisis: do not seek help through a chatbot. Call 999, NHS 111 (option 2), or Samaritans on 116 123.

What happens to my information?

• Your answers go to the NHS service you are referring yourself to, and are then handled under that service's data rules.

• The vendor acts as a "data processor". Its UK privacy policy (May 2026 version) states that:

  • the chatbot is hosted on Amazon Web Services, with chat data stored in MongoDB Atlas;

  • Sentry is used to track errors, and Mixpanel to understand how people use the chatbot;

  • "your data may be stored or accessed outside the UK or EU", subject to safeguards;

  • data is deleted once no longer needed, "usually within 30 days";

  • "We don't currently use automated decisions that affect your care or legal rights."

• It uses a large language model: the vendor says so in its FAQ, and independent academics also describe it as an "LLM-enhanced" application. But the privacy policy does not name any large language model provider, and we could not confirm which one is used or how data flows to it.

• Two statements do not match:

  • on the government procurement platform, the vendor lists data storage in the "United Kingdom" and the "European Economic Area (EEA)", while the privacy policy says data may be held outside the UK and EU;

  • the privacy policy says it is for people "over the age of 13", while both the Instructions for Use and NICE's scope cover people aged 16 and over.

• The Instructions for Use advise care when using it in public, so that others cannot see your screen, and warn that if you close the browser tab before submitting, the information you entered will be lost.

Is it worth using compared with the alternatives?

Public NHS procurement records show 4 contracts with Limbic:

• values range from £120,895.50 (3 years) to £199,101 (about 3.5 years);

• this works out at roughly £40,000 to £80,000 per service per year.

NICE's independent assessment group calculated that saving less than 3 minutes per assessment would be enough to offset the license fee.

But that calculation covers only the license fee. NICE also notes that there is little data on indirect costs such as training, promotion and clinical safety assurance, which may be substantial, and that the later benefits cannot yet be quantified. So whether it pays off once all costs are counted is not yet clear.

NICE considers the overall quality of evidence for the two products to be "broadly comparable". Therefore:

• compared with phone and online forms, it offers fairly credible help in bringing more people forward for help, and is worth considering;

• compared with Wysa DRA, it has more evidence but costs more; the choice depends on a service's referral volume and how much evidence it requires.

Who it suits, and who it does not

Suits

• NHS Talking Therapies services that want to increase self-referrals, reduce manual data entry and have more complete information before the first assessment, provided they meet the conditions set out in the "Current recommendation" above.

• Adults who do not want to phone, or who want to seek help outside working hours.

Less suitable for

• Services that want to use it in place of risk assessment or crisis triage.

• Services that want to allocate treatment directly from its "diagnosis predictions".

• Organizations that mainly serve people who do not speak English: the Instructions for Use state that using it in another language will "negatively impact the accuracy of the device", and the vendor does not guarantee the accuracy of automatic translation.

• People in crisis: please call an emergency number or crisis helpline directly.

• Anyone under 16.

Practical advice for services

• Before signing, ask the vendor to confirm in writing whether it calls an external large language model, where data is stored, and for how long.

• Ask the vendor for the contents of Field Safety Notice FSN-2025-001.

• NICE will review its guidance after 3 years, and contract length should take this into account (this is NICE's own advice).

• Keep an eye on the vendor's ability to keep operating:

  • the company's UK parent had net assets of only about £46,000 at the end of 2025, with accumulated losses of about £11 million, and relies on equity funding and loans to operate;

  • the company's accounts disclose a further equity raise of about US$20 million in March 2026.

Why this rating

• The conditions for using it are fairly clear, and fairly numerous: NICE allows its use only while evidence is being collected, and services must also keep other routes open, not use it as a risk tool, and not allocate treatment directly from its predictions. That is why this is "adoption under conditions".

• Some of its benefits have been recognized by an independent body, mainly in bringing more people forward for help and in user satisfaction.

• The most important questions still lack independent evidence: how accurate the collected information is, how much time it actually saves, and whether the total cost is worth it. The developer's only randomized controlled trial has not yet published results. NICE has allowed 3 years to collect this evidence; if independent research in that period shows that it does improve the quality of information and save clinical time without missing risks, there would be grounds for an upgrade.

• Marketing claims go further than the approved use: "diagnostic accuracy", doubled recovery rates, and an NHS trust's 2023 account that "UKCA auditors … found" a "53% improvement in recovery rates" all go beyond what the independent assessment recognizes, and some cannot be traced to a source.

Where the evidence comes from

• Directly verified by QiaoLink (E2):

  • the vendor's Instructions for Use (August 2025 version);

  • the privacy policy (May 2026 version) and terms of use;

  • the MHRA medical device registration record;

  • clinical trial registry records;

  • published NHS procurement contracts;

  • the product listing on the government procurement platform (G-Cloud);

  • the company's 2025 accounts filed at Companies House;

  • the vendor's marketing website;

  • the Field Safety Notice web page (existence confirmed; the document itself cannot be downloaded).

• Independent external evidence (E3):

  • NICE early value assessment HTG756 (published July 2025): recommendations, technology descriptions, committee discussion and prices;

  • the full report of the independent assessment group for NICE (Health Technology Assessment, June 2026);

  • usage data published by an NHS trust (Surrey and Borders);

  • an independent review by Professor Robert Meadows, University of Surrey;

  • independent comments in the media by John Torous, digital psychiatry specialist at Beth Israel Deaconess Medical Center (BIDMC) in the US;

  • commentary by independent academics in npj Digital Medicine;

  • public statements by NHS organizations including Oxford Health and Bradford;

  • trade media reports on the company's funding.

• Mainly from the developer (E1):

  • the 2024 Nature Medicine study and the 2023 JMIR AI study (all authors are employed by, or paid consultants to, the company, and hold shares);

  • the 2026 clinician survey in Frontiers in Digital Health (funded by the developer; the authors are employed by, or paid consultants to, the company);

  • figures on the vendor's website for accuracy, share of services using it, recovery rates and similar.

• Still missing (E0):

  • an independent controlled evaluation of the quality of the information collected and its effect on clinical decisions;

  • actual net time saved and total cost;

  • real-world performance in identifying risk;

  • which large language model is used;

  • the contents of the Field Safety Notice;

  • the results of the developer's randomized controlled trial.

• Verification limits: we did not test the product hands-on for this rating. It is embedded in NHS service websites, and using it would submit a real referral, so QiaoLink has not directly verified the conversation experience or the quality of the reports it produces.

QiaoLink did not register, pay or contact the vendor in the course of this rating. Evidence is current as of 2026-10-22.