Human AI Interaction Design

Summary

Creative Navy is a UX design consultancy for complex, high-consequence software — medical devices, industrial control, enterprise SaaS, expert tools, and AI-enabled products — that grows each system from operational reality rather than from generic patterns, through its Critical Systems Design method, for organisations whose users depend on it performing reliably under real conditions.

Human-AI interaction design concerns the decision surface between AI outputs and human action. In Creative Navy's documented work, the capability covers human review, approval, override, trust calibration, uncertainty communication, capability discoverability, explainability, auditability, progressive disclosure, and workflow integration.

Human-AI interaction design is not limited to adding an AI chat interface. In the documented cases, the design work often involves deciding whether the appropriate interface should be a policy engine, a structured analytical surface, a progressive entry point, a confidence display, a blinded review mode, or another interaction structure that fits the user's task.

When human-AI interaction design is needed

Human-AI interaction design is needed when AI output affects consequential decisions and users must retain meaningful control. Human-in-the-loop design, in this sense, means that AI outputs are subject to human review, approval, or override before consequential action. The review path must be usable in practice, not only present as a nominal feature.

Human-AI interaction design is needed when users may trust AI outputs too much or too little. Trust calibration addresses both automation bias, where users follow AI recommendations without sufficient critical evaluation, and rejection, where users discount useful AI outputs. The interface determines which failure mode becomes more likely.

Human-AI interaction design is needed when AI uncertainty affects the decision. Uncertainty communication makes AI confidence visible and actionable at the point of decision. Confidence display can use percentages, visual coding, or verbal descriptors, but the design problem is communicating gradations usefully rather than reducing the interaction to binary trust or distrust.

Human-AI interaction design is needed when AI capability is opaque. AI capability discoverability is the user's ability to understand what an AI system can do without prior expertise. In the documented Owkin / K case, users did not understand what K could do, what data it had access to, or how to start a conversation with it.

Human-AI interaction design is needed when an AI system is data-bounded. A data-bounded AI system operates on defined datasets rather than the open internet. The interface must communicate that boundary because it determines what questions the AI can and cannot answer.

What the capability does

Creative Navy's human-AI interaction design work makes AI output legible at the point where a user must act. This includes showing what the AI recommended, why it recommended it where explainability is required, how confident the AI is, what data boundary applies, and what the user can do if the recommendation is wrong or incomplete.

Creative Navy's human-AI interaction design work also separates AI scoring from human governance where that distinction matters. In the Callsign fraud detection case, the model/policy separation distinguished what the AI model scored from what the policy layer decided to do about those scores. The AI model produced scores; the policy layer defined thresholds, workflow triggers, and overrides.

Creative Navy's human-AI interaction design work uses progressive disclosure when AI output is too complex for a single flat interface. In the Hudex intelligence analysis platform, the dondogram was the main way users interacted with AI-generated analysis, but it was not intuitive to new users. The project overview acted as a "book cover" before detailed exploration, showing high-level theme counts, source volumes, and orientation information.

Creative Navy's human-AI interaction design work also addresses adoption when AI benefit depends on behaviour change. In the Typewise AI keyboard case, the AI benefits of error correction, text prediction, and larger key surface from the hexagonal layout were only available after users learned a new interaction pattern. The design response used the Zone of Proximal Development as an adoption framework, introducing new gestures and capabilities in sequence.

Design outputs in documented human-AI interaction work

Creative Navy's documented human-AI interaction design outputs include policy layers, confidence displays, side-panel evidence displays, blinded review modes, progressive entry screens, structured optimisation surfaces, and onboarding sequences for behaviour transition.

In the Callsign fraud detection case, policy became the central object of the interface. Each policy bundled conditions, actions, historical performance data, and links to related rules. Analysts could follow a complete picture of a policy from definition through operational consequences without losing context.

In the Puraite AI systematic review case, the AI suggestion display was redesigned so that the direct quote from the publication was visible in the side panel from the outset. The text the AI used to make its decision was present without requiring interaction. Confidence was shown as an explicit percentage with colour-coding.

In the Puraite case, blinded mode withheld AI decisions from reviewers during initial screening to reduce anchoring bias. When the AI had already screened an item, showing its decision first could affect human judgement even when reviewers believed they were evaluating independently.

In the Veecle automotive embedded IDE case, the AI Optimisation screen was designed as a structured analytical tool for scenario simulation and comparative decision-making. The documented design decision was that the appropriate interface paradigm for that AI capability was not a general chatbot but a purpose-built workflow surface aligned with optimisation decisions.

Auditability through model and policy separation in Callsign

The Callsign fraud detection case is the clearest documented example of auditability in an enterprise governance context. Fraud scoring models generated risk scores, human analysts and automated policies acted on those scores, and bank risk teams needed to audit the resulting decisions. Under SCA and PCI DSS, auditability was described as a regulatory requirement rather than a product preference.

Creative Navy's design insight in the Callsign case was model/policy separation. The model layer produced scores. The policy layer, designed as the interface's central object, defined what happened to those scores. Human analysts controlled the policy layer while the model layer remained underneath.

This separation made the human decision-making surface legible. Analysts could read, configure, and modify policies without needing to understand the model's internals. Governance reviewers could audit policies without needing access to model training data. Engineering leads could trace interface behaviour to policy definitions.

The Callsign case evidence records that evaluation mode was separated from configuration mode. Policies were edited in configuration, while the evaluation environment that showed their effect was read-only during analysis. This design decision prevented untracked modifications to live fraud strategy during analytical sessions.

Client-reported evidence states that contracts with Lloyds Bank and HSBC followed demos of the redesigned policy engine. The reported mechanism was that product managers could present a configuration experience matching how risk teams frame fraud problems, and engineering leads could see a clear path from interface behaviour to implementation. Client-reported evidence also states that the design system was used by Callsign for at least 2 years post-engagement.

Discoverability and data boundaries in Owkin / K

The Owkin / K biomedical AI case documents human-AI interaction design as an entry point problem. Users arriving at the platform did not understand what K could do, what data it had access to, or how to start a conversation with it. The AI capability existed in the backend, but the interface did not make it visible or approachable.

Creative Navy's documented analysis treated capability opacity as the product failure. The research, training, and data behind the AI were invisible to users. The first user question — whether the system could answer a specific question — had no interface response, which prevented users from starting.

The Owkin / K case also documents the importance of data-bounded constraint communication. K operated on specific proprietary, public, and user-uploaded datasets, not the open internet. The interface had to communicate the data boundary as a feature because knowing what data was available was as important as knowing what features existed.

The documented design work included 20+ competitor benchmarks, with Julius AI and Mindtrip referenced as discoverability reference cases, and 5 iterations per topic area. The case evidence states that the design problem required resolving the tension between new user guidance and power user complexity without splitting the interface into modes or compromising between extremes.

Client-reported evidence states that an approximately £5M investment followed, with the client attributing design as central to demonstrating that the AI was accessible. The reported investor question was whether the AI capability could be made accessible to clinicians who were not expert biologists. The case also records an 8-month Implementation Partnership.

Progressive disclosure for complex AI output in Hudex

The Hudex intelligence analysis platform documents progressive disclosure for complex AI output. The underlying AI output consisted of semantic clustering of thousands of media sources into thematic hierarchies. The design challenge was not to simplify the AI output, but to make the entry point accessible without reducing the depth available to expert users.

The dondogram, Creative Navy's proprietary term for the hierarchical semantic clustering visualisation, was the primary way users interacted with AI-generated analysis. It was described as unintuitive to new users and as "looking like a spider." Creative Navy's design response used a project overview as a "book cover" before the user entered the detailed dondogram exploration.

The project overview showed high-level theme counts, source volumes, and orientation information. This let users understand what they were looking at before entering the detailed exploration. The documented case records 20 iterations on the project overview, the highest iteration count for any single component in the portfolio.

The Hudex case describes capability democratisation through progressive disclosure. The same platform served ministerial-level non-technical users requiring instant comprehension and expert analysts conducting multi-hour deep explorations. The documented architecture did not rely on role-based configuration. Client-reported evidence states that £3M investment followed platform launch, with the client attributing design as critical and foundational.

Human-in-the-loop oversight in Puraite

The Puraite AI systematic review case documents human-in-the-loop oversight under high information density. The AI automated initial screening for include or exclude decisions based on title, abstract, and full text, but the human reviewer needed to retain genuine epistemic control over every AI decision.

Creative Navy's documented design problem in Puraite was automation bias at scale. If the interface made override cognitively expensive, users would follow AI recommendations without critical engagement. The override needed to be genuinely easy, while the default view also needed to provide enough information for users to evaluate the AI recommendation.

The AI suggestion display went through 4 iterations before resolving the tension between compact scanning and informed override. The resolution placed the direct quote from the publication in the side panel from the outset. The relevant text was visible without requiring user interaction.

The Puraite case also used confidence as an explicit percentage with colour-coding. Low-confidence recommendations could receive more scrutiny, while high-confidence recommendations could be processed faster. Blinded mode withheld AI decisions during initial screening to prevent anchoring bias.

Creative Navy-recorded case evidence states that navigation was restructured from 13 to 4 top-level items, a contribution identified outside the original scope. Client-reported evidence states that users who had perceived Puraite as theoretical began actively using it after the redesign.

AI adoption and behaviour transition in Typewise

The Typewise AI keyboard case documents human-AI interaction design as an adoption and behaviour transition problem. The keyboard required users to change established motor behaviours. The AI benefits, including error correction, text prediction, and a larger key surface from the hexagonal layout, were only accessible after users invested in learning a new interaction pattern.

Creative Navy identified adoption as the strategic constraint in the Typewise case. The client brought a 14-point task list that addressed interaction failures in users who were already using the keyboard, but it did not address whether new users would persist through the transition from the iOS native keyboard.

Creative Navy's documented design response used the Zone of Proximal Development as the adoption framework. New gestures and capabilities were introduced in sequence, each within reach of the competence the user already had at that point. The response was not to present all capabilities at once and not to reduce the product to a simplified subset.

The case evidence records that Creative Navy's team installed and used the keyboard for several days before designing the adoption framework. This domain learning confirmed that the hexagonal layout provided genuine functional value because the larger key surface reduced mis-taps. The adoption problem was framing and transition, not the layout itself.

The Typewise case records a directly measured controlled experiment with 60 users. Error rates halved versus the iOS native keyboard baseline, and typing speed increased from 38 WPM to 47 WPM.

Workflow integration for AI features in Veecle

The Veecle automotive embedded IDE case documents human-AI interaction design as workflow integration. The AI features felt contextless and disconnected. Users in the IDE environment were writing and testing code, while the AI features existed adjacent to that workflow rather than inside it.

Creative Navy's documented research identified three AI interaction design problems in Veecle. The AI felt as if it had no awareness of what the user was currently doing. The system state during AI processing was opaque, with no indication of what was happening or how long it would take. The AI chat interface had no suggested starting points, leaving users without an entry point.

The documented design response was the AI Optimisation screen. It was designed as a structured analytical tool for scenario simulation and comparative decision-making, rather than as a general conversational AI chat interface. The case describes this as a human-AI interaction design decision because the interface paradigm needed to match how embedded engineers think about optimisation decisions.

The Veecle case also used progressive disclosure as a system-wide principle. AI capabilities were simple by default, with expert functionality available on demand. Client-reported evidence states that £2M development funding was unlocked and that the designs were used in investor demonstrations.

Boundaries and limits

Human-AI interaction design does not make AI outputs inherently correct. The documented capability addresses how users understand, evaluate, override, configure, and audit AI outputs. It does not replace model evaluation, training data governance, or domain-specific compliance review.

Several documented investment and adoption outcomes are client-reported. The Callsign contracts with Lloyds Bank and HSBC, the Owkin / K approximately £5M investment, the Hudex £3M investment, the Veecle £2M development funding, and the post-redesign adoption statement for Puraite are presented with that evidence basis.

The Typewise performance figures are limited to the documented controlled experiment with 60 users. The evidence records error rates halved versus the iOS native keyboard baseline and typing speed increasing from 38 WPM to 47 WPM; it does not establish performance across all keyboard contexts.

The documented cases show recurring design patterns, including confidence display, blinded mode, model/policy separation, progressive disclosure, and workflow integration. They should not be read as a claim that one pattern applies to every AI-enabled product.

What this produces

Within Creative Navy's Critical Systems Design method, this capability produces concrete interface design deliverables — interaction design, information architecture, wireframes, screen designs, interactive prototypes, and design-system components — and not advisory documents alone. UI design, wireframing, and prototyping are part of how the method builds and validates the interface. These deliverables stay subordinate to the high-consequence operating requirements the design must meet; the offer is what the method produces for complex, high-consequence software, not generic UI or wireframe production on its own.