Frequently asked questions

PromptSafe is an adversarial, pre-deployment testing platform for conversational AI in health and mental health. You run your agent through hundreds of hostile, realistic simulated conversations before launch, see exactly where it fails and why, get specific suggestions for what to change, and export a record of the testing for your governance, safety, and product teams.

Product, safety, clinical, and governance teams building or buying conversational AI in health and mental health. The people who get the most out of it are the ones closest to the agent's behaviour: product managers, behavioural scientists, clinical leads, subject-matter experts, and anyone responsible for what the agent says and how it says it. PromptSafe was designed so they can run testing themselves without needing engineering support.

The free trial takes a couple of minutes to sign up for, no card required. Your workspace comes with example agents, personas, and starter evaluator templates, so you can run your first simulation within minutes of signing up rather than starting from a blank page.

No. You build or describe your agent inside PromptSafe. You can also connect your own agent on request: OpenAI models are supported today on the paid tier, where setup may involve a fee and a short queue. Other models and fuller integration are enterprise.

Health and mental health. That is where our evaluators, benchmarks, and proof sit, so it is what we build around: patient-facing health assistants, mental health and wellbeing apps, digital health coaching, and clinical and behavioural agents. The platform itself works on any conversational AI, and it is built for health and mental health first and applicable to any high-stakes conversational AI, so if you are outside health and facing the same problem, talk to us.

PromptSafe is built on published behavioural science and evaluation methodology, not just features. Our founder co-authored a 2026 open letter calling for behavioural science to become core to AI evaluation, and a peer-reviewed framework (FAST) for evaluating conversational AI across fidelity, accuracy, safety, and tone. We are also developing the Evaluation Evidence Framework, a transparent way to quantify confidence in evaluation results. You can read more on our Scientific foundations page.

Seven days from when you verify your email.

No. The trial is fully self-serve and does not require a card. We absorb the running cost.

You get the full product, with 30 PS tokens per day. That is enough to run several simulations of varying complexity per day, exploring how PromptSafe works across personas, evaluators, and the Prompt Improver. We recommend starting with the guided product tour: it walks you through the whole workflow, from setting up an agent to reading your first evaluation, and surfaces features that are easy to miss on your own, such as the Prompt Improver.

You are redirected to a usage page where you can buy a Pack of tokens to keep going. Your account, agents, personas, and evaluators are preserved.

The trial is the trial, but if you have a specific reason to need more time before deciding, book a discovery call and we'll talk about it.

After the free trial, you buy Packs of usage. Each Pack is £99 and includes 2,000 PS tokens, valid for 90 days. If you buy another Pack before that period ends, your unused tokens roll over and, together with the new tokens, remain valid for a fresh 90 day period starting from your latest purchase. If a previous Pack has already expired, a new one gives you 90 days from the date of purchase.

Prices on this page are exclusive of VAT. UK customers pay 20% VAT on top of the listed price at the Stripe checkout step, so a £99 Pack will be charged at £118.80 in total. VAT is automatically calculated by Stripe based on the billing address you enter at checkout. A full VAT invoice is provided after purchase. Non-UK customers currently pay the listed price with no VAT added.

Not at launch. We deliberately started with Packs so teams can buy what they need, when they need it, without committing to a recurring spend. We may introduce subscriptions later based on what teams ask for.

A PS token is the unit PromptSafe uses to measure usage. A typical simulation uses around 5 to 10 PS tokens. Simpler runs cost less, more complex runs cost more, and usage varies with the combination and number of features you run, the number of personas, the number of evaluators, and how long the conversations get. The usage indicator inside the product shows you how many PS tokens each simulation has used, so you can build a feel for it over time. We will refine these estimates as more teams use the product.

A usage indicator inside the product shows your remaining tokens at all times. You can also see the current expiry date for your tokens, which moves forward each time you buy a new Pack.

All your PS tokens are pooled into a single balance, regardless of which Pack they came from. Each Pack is valid for 90 days, and buying another Pack before the current one expires rolls your unused tokens into a fresh 90 day period starting from your latest purchase. For example, if you buy one Pack on the 1st and another on the 20th, all your tokens expire together, 90 days after the 20th, not on different dates. If your previous tokens have already expired, a new Pack simply gives you a fresh 90 day balance from the date of purchase.

Yes, that's what the enterprise tier is for. Volume pricing, bring your own model API keys, agent-via-API integration, and custom support are all available. It is also where we design the test suite with you, which is the right route for governance-heavy or regulated deployments. Book a discovery call.

Conversational AI agents in health and mental health: patient-facing health assistants, mental health and wellbeing apps, digital health coaching agents, and clinical or behavioural agents. That is where our evaluators and benchmarks are tested. The platform itself works on any conversational agent, so if you are building outside health and facing the same problem, talk to us.

The platform can. The personas, the simulation, and the scoring all apply to any conversational AI, because the underlying patterns (how people behave in conversation, what makes an agent helpful or harmful) are not sector-specific. What is specific is where we have tested and can prove it, which today is health and mental health: that is where our tested evaluators and benchmarks sit. So PromptSafe is built for health and mental health first, and applicable to any high-stakes conversational AI. If you are building elsewhere and want testing you can stand behind, talk to us about what that would take.

LangSmith and Langfuse are engineering platforms focused on tracing, monitoring, and evaluating AI agents against test data or production traffic you bring them. They tell you what your agent did and where it slowed down or failed. PromptSafe sits upstream of that: it generates the conversations being tested, using realistic and adversarial personas, then scores observable behaviours across multi turn dialogue. PromptSafe is also designed for product, clinical, and governance leads to use alongside engineers, not engineers alone. Many teams who care about agent quality use both: PromptSafe pre-deployment to surface behavioural failures, observability tools after to watch the agent with real users.

No, and it is not designed to. Human review by domain experts is irreplaceable, especially in regulated or clinical contexts. PromptSafe scales the parts of testing that benefit from automation (running hundreds of conversations, scoring them consistently, surfacing patterns) so human reviewers can focus on the parts that need their judgement.

No. Nothing makes an agent safe. The team building the agent is responsible for its safety. What PromptSafe does is surface failures, help you fix them, and produce a record of the testing you did and the results you got.

No. PromptSafe is a pre-deployment testing tool for finding behavioural failures and producing a record of testing. It does not constitute clinical validation, regulatory approval, or compliance certification. Where your deployment context requires formal validation under FDA, MHRA, Health Canada, TGA, or equivalent frameworks, PromptSafe complements that work but does not substitute for it. The team responsible for the agent's deployment is responsible for ensuring the appropriate validation pathway has been followed.

ISO/IEC 42001 is the management-system standard for AI. PromptSafe produces a pre-deployment testing record for your conversational agent: what you tested, what the personas did, and how each evaluator scored it, with the conversation behind every result. That is the kind of evidence the standard expects you to keep. PromptSafe gives you that evidence; it does not certify your compliance. If you need testing designed to stand up to formal regulatory or audit scrutiny, that is an enterprise engagement, where we set up the test suite with you rather than leaving it to a self-serve trial.

If your agent is a high-risk system, Article 9 of the Act requires a documented, ongoing risk-management process, enforceable from August 2026. PromptSafe gives you test artefacts that contribute to that evidence requirement: what you tested, what the agent did, and where it failed. It is evidence for your file, not a conformity assessment. For high-risk systems we would recommend an enterprise engagement rather than self-serve testing, so the suite is designed around your risk profile.

DCB0129 and DCB0160 are the NHS clinical risk-management standards for manufacturers and deploying organisations. PromptSafe outputs, the transcripts and evaluator outcomes, can contribute to the clinical safety case your clinical safety officer assembles. PromptSafe is one input to that case, to be interpreted by your clinical safety officer alongside human review, and it does not replace clinical review. For clinical deployments we would normally recommend an enterprise engagement so the test suite is designed with you.

No. Frameworks like ISO/IEC 42001, the EU AI Act, and NHS DCB0129/0160 are process and risk-management standards. They tell organisations to test, document, and audit AI systems, but they don't specify a list of behavioural tests an agent must pass. The right tests depend on your agent, your users, and your risk profile. PromptSafe ships with a set of behaviour-focused starter evaluators across safety, scope, escalation, and clinical-risk dimensions, and you can author your own. The evidence those evaluators produce is what supports your framework documentation.

A full export of each simulation: the agent name and version, the persona used, the evaluator applied with criteria, the score and summary statement, what the agent did well and where it failed, improvement suggestions, the full conversation log, and the date the report was downloaded. At launch, exports are in CSV format, easy to analyse, summarise, or turn into a polished report using your AI tool of choice outside PromptSafe. You can hand the export to your governance team, your AI safety lead, your auditor, or your regulator.

Governance teams, AI safety leads, regulatory contacts, and product teams who need to demonstrate testing happened. We deliberately design the artefact for those readers, not for board summary slides.

Yes, the exports are yours to share. They are designed to contain the elements an auditor would typically expect to see: what was tested, what the agent said, and how it was scored. Bear in mind that results are probabilistic and vary between runs, so they should be presented as a record of testing carried out and read alongside human review, rather than as a certification. Where the evidence needs to withstand formal external scrutiny, we would recommend an enterprise engagement.

PromptSafe runs your agent against synthetic personas in multi turn conversations. The personas are not scripts. They behave dynamically based on a behavioural science framework, changing direction, asking unexpected questions, and pushing back the way real users do.

No. Personas are synthetic and grounded in behavioural science. They model real patterns of behaviour, beliefs, and conversational dynamics, but they are not derived from any identifiable individual. The persona schema covers identity and background, beliefs and behavioural orientation, agent interaction style, behaviour history, and additional context including neurodiversity and emotional regulation.

PromptSafe uses a number of carefully chosen models under the hood, each tuned for the specific work it does: generating personas, simulating conversations, evaluating results, and producing improvement suggestions. At launch, simulations run on GPT-5.4. The simulation settings show other models in a dropdown but they are greyed out by default. To unlock alternative models for simulation, you need the enterprise tier, or you can contact us to discuss what you'd like.

Yes, on the enterprise tier. Bring your own keys lets you run your testing through your own model accounts. This is gated behind enterprise because it adds complexity that most self-serve users do not need.

Your workspace is private to you and your teammates. We never share your content with other customers and we do not sell it. Running a simulation does send the conversation, including your agent's instructions, to the AI model providers that power the testing (currently OpenAI, Anthropic and xAI), who process it as our service providers in order to run the test. We treat the things that describe your own AI system, your agent prompts, system instructions, uploaded documents and configuration (we call these Customer Product Assets), as confidential, and we do not use them for research or to improve our models by default. We may use Evaluation Assets, the personas, evaluators, simulation transcripts, scores and metrics produced by the testing process, with appropriate safeguards, to improve PromptSafe and for research. Nothing is used to train third party or general purpose AI models, and any research we publish will contain only aggregated findings, never anything that identifies you or your organisation. Full details are in our privacy policy.

Your workspace is private to you and any teammates you have invited. No other customer can see it, and Sacher AI does not share workspace content with third parties for their own purposes. Two things are worth knowing. The AI model providers that power the simulations (currently OpenAI, Anthropic and xAI) process conversation data on our instructions in order to run the tests. And on enterprise engagements, Sacher AI may set up and work inside the workspace as part of the agreed scope.

Sometimes, yes. Sacher AI may add curated example personas or evaluators to your workspace where they are relevant to your sector, so you have more useful starting material to work with. Any items Sacher AI adds are clearly labelled as examples, so you can always tell what you created and what was added. Sacher AI never modifies or deletes anything you have created.

PromptSafe does not use your data to train third party or general purpose AI models, and it does not use your Customer Product Assets (your agent prompts, system instructions, uploaded documents and configuration) to improve our models by default. We may use Evaluation Assets (the personas, evaluators, simulation transcripts, evaluator outputs, metrics and benchmarks produced during testing), with appropriate safeguards, to improve PromptSafe itself, including our evaluation methodologies, benchmarking systems, safety classifiers, and scientific research. We never publish customer identifiable information and never sell customer data. The principle is simple: PromptSafe improves by learning how AI systems behave during evaluation, not from the design of your AI system.

PromptSafe runs on secure cloud infrastructure with customer data hosted in the European Union. If you have specific data residency requirements, the enterprise tier can accommodate them.

Yes. Persona Builder has an Import a batch button that creates multiple personas at once from a single CSV file, rather than building them one at a time in the UI. It is the fastest way to populate a workspace if you already know the range of users you want to test against.

Download the persona import template (CSV). The CSV has 24 columns covering identity, demographics, behavioural traits, and interaction style. It includes the column headers and one filled example row showing the expected format. Open it in Excel, Google Sheets, Numbers, or any text editor. Fill one row per persona and save as CSV. Some columns are free text and some map to dropdowns in the product UI, where the CSV value needs to match the dropdown label exactly. The example row in the template shows the correct format for each.

This is for you to decide. You know your users better than we do. Persona design depends on your product, your user base, and the failure modes you most want to test against. That said, the schema is designed to work well with a clear brief and an AI assistant. A practical approach: write a short brief describing the kind of users your agent serves, the sub-segments that matter, and the test cases you care about (typical users, edge cases, adversarial users). Share the brief with an AI tool of your choice, along with the import template, and ask it to draft a CSV that fits the schema. Review and refine the output. Adjust any persona that feels generic, exaggerated, or off-brief. Save the CSV and import.

If you would rather have personas designed for you, we offer professional services for workspace setup. Book a discovery call to discuss your testing goals.

A typical ten-turn simulation runs in one to three minutes end to end. Longer conversations take proportionally longer. You can keep working in the app while a simulation runs and come back to the results when it is done.

No. A single score on one conversation is not a reliable measure on its own. Large language models judge each conversation, and like any probabilistic system, they can return a slightly different score when the same conversation is scored more than once, especially on nuanced, judgement-based criteria. PromptSafe is designed to be read as trends, not single results. Run your agent across many conversations and personas, then look at how the scores move over time. That tells you whether a change to your agent has made things better, worse, or no different. The pattern across many conversations tells you more than any single score does.

Each evaluator describes a specific behaviour, for example "recognises and signposts clinical risk" or "stays within its scope of authority". It scores a conversation from 1 to 5, shown as a percentage, and a higher score means the agent met that criterion more fully. The score is an overall judgement of how well the agent met the criterion across the conversation, taking into account both what it did well and where it fell short, not a count of how often it did something. Scores are produced by a large language model reviewing the conversation against your criteria, and each score comes with an explanation and the quoted evidence behind it, so you can see the reasoning and judge it yourself. Scores can vary slightly when the same conversation is scored more than once, particularly on nuanced criteria, so read them alongside their explanations and look at the trend across many conversations rather than relying on a single result.

PromptSafe uses a large language model to score each conversation against your evaluators. Large language models are probabilistic, which means they can return a slightly different score when they assess the same conversation more than once, particularly when the criteria call for judgement. If you see meaningful variation when you rescore the same conversation, it often means the evaluator criteria are not precise enough. Tighten the wording to describe the behaviour you want the agent to show, then rescore. The variance should narrow. Some variance will always remain because of the probabilistic nature of large language models. This is why we recommend reading scores as a trend across many conversations and personas, not as a fixed measurement of any single conversation.

Click the down arrow on the evaluator card to expand it. You will see the reasons for the score, with examples from the conversation. Click Prompt Improver to get suggestions for updating your agent's prompt or instructions. Make the change, then run a new simulation across a few personas to see how the trend moves.

Five to ten turns is the sweet spot for most testing. Long enough for the agent to show how it responds when a persona keeps pushing, changes direction, or asks something unexpected. Longer simulations are useful when you are stress-testing a failure mode that only appears later in the conversation.

Once you have updated your agent, run a new simulation with the same personas and evaluators. You will get a new conversation, which means the scores can differ from the previous run regardless of whether your changes helped. Run a few simulations across different personas before deciding whether the change worked. Comparing a single new conversation to a single old one is not enough to tell. The reliable read is the trend across the set.

Yes. Every simulation produces an exportable CSV report you can hand to your team, AI safety lead, auditor, or regulator. For the full field list, see "What does the audit trail actually contain?" in the Governance group.

Realistic personas behave like the users your agent is built to serve. They have plausible needs, beliefs, and patterns of behaviour drawn from your sector. Adversarial personas are designed to stress-test a specific evaluator, pushing, deflecting, and probing in the ways most likely to make the agent break that particular check. Use realistic personas to find out how your agent performs in normal use, and adversarial ones to find where it can be broken.