Case studies

What Sacher AI work looks like in practice.

These examples show the kinds of problems Sacher AI helps solve: retention pressure, behaviour change design, AI safety, and the need for credible evidence that stands up beyond demos. Each one reflects a different way we help: strategy, delivery, evaluation, and research.

01

Decide

Twenty AI ideas, narrowed to the six worth doing

An international GLP-1 care provider had heavy support demand and was losing patients early. We helped it narrow more than twenty AI ideas to the six worth building, with a phased roadmap.

The problem

  • Hundreds of thousands of support tickets a month overwhelming operations
  • High early churn before patients reached full benefit
  • Manual processes creating clinical and compliance risk
  • Retention identified as a major commercial growth lever
  • More than twenty candidate AI ideas across different teams, with no shared basis for choosing between them

The approach

  • Mapped high-impact intervention points across the patient journey
  • Combined behavioural science with AI system design
  • Focused on proactive support, retention, and clinical consistency
  • Scoped use cases aligned to operational and regulatory realities
  • Tested each candidate for behavioural value, technical feasibility and regulatory alignment, which is what took the longlist down to six
  • Worked closely with the provider's regulatory affairs and quality assurance teams, so the roadmap met their requirements in each country where it operates.

Why it mattered

The commercial problem was not separate from the care problem. Patients needed better behavioural and clinical support earlier, while the provider needed a model that could scale without simply adding more manual operations. The expensive mistake was never going to be a weak model. It was going to be spending a year building the wrong thing, in an organisation where several plausible options all had internal support. Deciding what not to build was the work.

What this enables

A comprehensive strategic AI discovery, grounded in 20 cross-functional interviews and large-scale analysis of support-ticket and patient-survey data. It distilled more than twenty candidate ideas into six prioritised, safety-governed AI use cases and a phased roadmap, giving the provider a clear, de-risked path to improve retention, clinical safety, and operational efficiency, with the governance foundations to build on. This is the Before You Build method applied to a portfolio rather than to a single decision: the same uncertainties, asked across a set of competing options.

02

Design and deliver

Personalised AI health coaching deployed across obesity care

For Allurion, our team designed the behavioural architecture of Coach Iris, a personalised AI health coach for obesity and GLP-1 care. It launched in August 2023, within a year of ChatGPT's public release, and was deployed internationally.

The problem

  • Patients needed safe, personalised, around-the-clock support across behaviour change, nutrition, activity, and motivation
  • Support had to stay clinically appropriate and escalate to humans when needed
  • It had to scale across many clinics, countries, and treatment types

The approach

  • Behavioural and clinical architecture for anti-obesity medications, bariatric surgery, and intragastric balloons
  • Personalisation to treatment type, dietary needs, and interaction history
  • Human and automated safety monitoring, with referral to clinicians for medical questions
  • Worked closely with Allurion's regulatory affairs and quality assurance teams, so the coach met their requirements across the countries where it was deployed.

Why it mattered

At this scale, the behavioural layer is what turns a medication or device into sustained outcomes. The quality, safety, and personalisation of support directly shape adherence, retention, and patient experience.

Outcome

Patient-facing AI coaching deployed across 80 obesity clinics in 22 countries. In a patient survey presented at IFSO-EC 2024, covering 133 patients at seven centres, 94% were satisfied with the quality of support and 89% rated it safe and trustworthy.

03

Design and deliver

Safe, personalised patient-facing AI for people on weight-loss medication

Embedded within Numan's team from the very beginning, and working with them for more than two years, Sacher AI helped create Nu, an AI health assistant for patients on anti-obesity medications, alongside Aegis, a monitoring, safety, and escalation system.

The problem

  • Patient-facing AI at scale needs real-time personalisation without compromising safety
  • Clinical oversight has to be built in from the start, not bolted on later
  • Risks in language-model systems surface gradually across many interactions

The approach

  • Designed the AI Health Assistant for patients on anti-obesity medications
  • Built Aegis: real-time monitoring, safety, and escalation with clinical oversight
  • Combined rigorous clinical governance with real-time personalisation
  • Worked closely with Numan's clinical, regulatory affairs and quality assurance teams, so Nu and Aegis met their requirements.

Why it mattered

Combining safe, scalable, and personalised AI is one of the hardest problems in healthcare. The safety and escalation layer is what makes patient-facing AI trustworthy enough to deploy at scale.

Outcome

In a real-world safety evaluation of 26,654 conversations with Nu over 12 months, presented at ObesityWeek 2025, 99% were rated safe and 99% accurate, with 4% (1,066) escalated to clinicians.

04

Test and prove

Stress-testing an AI health assistant before it reached patients

A team building a patient-facing AI health assistant asked Sacher AI for an independent view on conversation quality before wider deployment. Manual transcript review had looked reassuring, but simulated testing found weaknesses it had missed.

The problem

  • The team was reviewing transcripts by hand across product, clinical, and operations
  • At first glance the assistant looked fine: polite, mostly accurate, nothing obviously alarming
  • But manual review misses the messy edge cases that only appear at scale

The approach

  • Ran the assistant through PromptSafe, our behavioural and clinical simulation framework
  • Simulated hundreds of full conversations from start to finish, not isolated prompts
  • Used synthetic patients designed to push the system: anxious, confused, angry, convinced the medication was harming them, or refusing to accept the answer

Why it mattered

These are the interactions that actually happen once a system goes live, and they rarely surface in a handful of hand-reviewed transcripts. Testing at scale, the way real people behave, is what reveals them.

Outcome

The simulations surfaced real weaknesses before launch: the assistant drifted from its own instructions, guardrails were easier to bypass than expected, and in some cases it reassured patients when it should have escalated to a healthcare professional. None had shown up in manual review, and all were fixable, because they were found in testing rather than after deployment.

05

Decide

Finding the right AI use cases and building the roadmap

Sacher AI worked with Stride to identify high-value AI use cases and turn competing ideas into a practical, prioritised roadmap the team could act on.

The problem

  • Many competing ideas for where AI could help, without a clear priority
  • A need to align product, clinical, and commercial thinking
  • A roadmap that could guide delivery and support investor conversations

The approach

  • Structured discovery to define the AI vision and business goals
  • Use-case identification and prioritisation across the patient journey
  • A phased, actionable roadmap with clear ownership and decision points

Why it mattered

The hardest part of AI strategy is not generating ideas, it is choosing the right ones and sequencing them so they can actually be built and adopted.

Outcome

A clear, prioritised set of AI use cases and a roadmap the team is now using to guide delivery.

06

Design and deliver

An intuitive eating coach for a healthier relationship with food

As Holly Health's AI partner within its UK Research and Innovation Women in Innovation programme, Sacher AI designed and tested an LLM-based AI agent MVP and smarter behavioural nudges to help people build a more sustainable, compassionate relationship with food, rather than another restrictive diet.

The problem

  • Relationship with food is a sensitive area where tone and framing matter as much as content
  • The programme needed an LLM-based coach that was behaviourally credible and safe
  • New features had to be genuinely helpful in daily life and tested with real users

The approach

  • Designed and tested an LLM-based AI agent MVP for goals around relationship with food and weight management
  • Introduced smarter nudges grounded in behaviour change science
  • Applied COM-B and structured behaviour change techniques, with PromptSafe testing for safety and alignment
  • Worked as an extension of the Holly Health team, alongside their Hunger-Fullness Scale and Insights Path features

Why it mattered

Relationship with food is high-stakes and easy to get wrong. The behavioural design and safety of the coach determined whether it felt compassionate and genuinely useful to people.

Outcome

Following roll-out, Holly Health reported a meaningful, tangible shift in users' relationship with food, with a user evaluation of the new features underway.

Where Sacher AI adds value

When the challenge sits across product, behaviour, operations, and evidence rather than inside one narrow function.

What tends to go wrong without this layer

Teams build the tool, but not the adoption logic, evaluation model, or safeguards needed for real-world use.

What the work is aiming to create

Products that are more trusted, more useful, easier to scale, and easier to defend with clinical, behavioural, and commercial stakeholders.

Next step

If your team is trying to turn AI ambition into a stronger product, this is the kind of work we help with.

Book a discovery call