# AICON: A field guide to the ideas Prepared by Acuity AI · 12 September 2026 An independent guide to core concepts, announced conference connections and selected background learning. Concept relationships are editorial synthesis. Sources are not endorsements. Video access and review limits are stated per resource. ## Contents - Knowledge readiness & authority - Retrieval-augmented generation - Knowledge graphs & connected retrieval - Entity resolution & data quality - Agents & delegated work - Human trust, oversight & decision support - Evaluation, observability & AI operations - Operational AI governance - Agent security & resilience - Data sovereignty, control & access - Adoption, diffusion & organisational change - AI across software delivery - AI in healthcare & clinical work - Creative authorship & human expression - Conversational voice & multimodal interaction - AI in the physical world ## Knowledge readiness & authority The work of making organisational knowledge discoverable, understandable, current and attributable to someone who can stand behind it. ### Why it matters An AI can retrieve a document without knowing whether it is the approved version. Useful organisational AI needs decisions about ownership, access, conflicting sources and updates as well as a capable model. ### Key distinctions - Information availability is different from authority: a readable file may be obsolete or unapproved. - A citation shows where a statement came from; it does not by itself establish that the statement is true. - A larger corpus improves coverage only when relevant material can be found and interpreted. ### A question for your organisation For one important business question, which source wins when two documents disagree, and who decides? ### Conference connections - AI Is Ready. Is Your Knowledge? Why the models we fixate on are rarely the real constraint in creating value from AI — programme topic link - The Importance of Data and Governance as the Foundations for AI Success — programme topic link ### Read and watch - [Liberty IT: our journey to responsible generative AI](https://www.liberty-it.co.uk/news-stories/liberty-it-journey-to-responsible-generative-ai) (vendor case study). See how reusable capabilities, use cases and governance were developed together. Provider-reported experience; outcomes have not been independently verified here. - [Tom Swann: observable data quality with Elementary and DataHub](https://tam-iii.medium.com/observable-data-quality-with-elementary-and-datahub-6fa5f92f2c81) (practitioner article). Connect quality checks with discoverable data context and ownership. Practitioner explanation or experience, not an independent controlled study. - [Instil: a practical guide to RAG strategies](https://instil.co/blog/a-practical-guide-to-retrieval-augmented-generation-strategies) (practitioner article). Understand how source material enters an answer through retrieval. Practitioner explanation or experience, not an independent controlled study. - [Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Enterprise Internal Knowledge](https://www.youtube.com/watch?v=LRGX-gTegVA) (video). Connect the conference knowledge demonstration to the challenge of making internal business knowledge useful to AI. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. - [Effective context engineering for AI agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) (practitioner article). Turn the knowledge-constraint argument into practical design choices. Public article reviewed. Background learning; no AICON appearance or endorsement implied. ### Related concepts - supplies trusted context to → Retrieval-augmented generation: Retrieval needs usable source knowledge, including its scope and authority. - requires ownership from → Operational AI governance: Someone must own source approval, refresh and access decisions. ## Retrieval-augmented generation Finding relevant material and supplying it to a language model as context for an answer, instead of expecting the model to remember everything. ### Why it matters RAG is a practical bridge between a model and organisational documents. Its quality depends on what is indexed, how documents are divided, how candidates are ranked and whether the final answer uses the evidence correctly. ### Key distinctions - Retrieval changes the context supplied at use time; fine-tuning changes model parameters. - Similarity is a useful search signal, not a guarantee of relevance or truth. - Reranking can improve candidate selection but adds work and latency. ### A question for your organisation Can the system retrieve the passage that would let a knowledgeable colleague answer this question? ### Conference connections - Zero ontology — programme topic link - AI Is Ready. Is Your Knowledge? Why the models we fixate on are rarely the real constraint in creating value from AI — editorial thematic connection ### Read and watch - [Instil: practical RAG strategies](https://instil.co/blog/a-practical-guide-to-retrieval-augmented-generation-strategies) (practitioner article). Compare basic retrieval, reranking and more elaborate retrieval patterns. Practitioner explanation or experience, not an independent controlled study. - [Instil: why chunking matters](https://instil.co/blog/why-chunking-matters-the-key-to-smarter-rag-workflows) (practitioner article). Compare fixed, recursive, semantic and late chunking approaches. Practitioner explanation or experience, not an independent controlled study. - [GraphRAG local search](https://microsoft.github.io/graphrag/query/local_search/) (technical documentation). See how entities, relationships and source text can be combined for a specific question. Describes this implementation, not a comparative evaluation. - [Stanford CS25: V3 I Retrieval Augmented Language Models](https://www.youtube.com/watch?v=mE7IDf2SmJg) (video). Understand why retrieving relevant organisational knowledge is different from relying on model training. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. - [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) (research paper). Understand the original architecture behind answers grounded in retrieved material. Abstract reviewed; linked full paper not newly reviewed in this supplement. Background learning; no AICON appearance or endorsement implied. ### Related concepts - can be extended by → Knowledge graphs & connected retrieval: Relationships can supply context that isolated passages omit. - must be tested through → Evaluation, observability & AI operations: Assess retrieval coverage separately from answer correctness. ## Knowledge graphs & connected retrieval A representation of entities and explicit relationships that lets people or software follow how pieces of knowledge connect. ### Why it matters A graph can connect a topic to a session, a speaker to their work, or a claim to its evidence. Its value comes from the meaning and quality of those connections. A visually impressive network alone is not a useful knowledge model. ### Key distinctions - A graph visualisation displays relationships; a graph retrieval system uses them to select context. - Local questions concern particular entities; global questions ask for themes across a collection. - An editor-curated thematic connection is different from a relationship explicitly asserted in a source. ### A question for your organisation What useful question becomes answerable by following a relationship that a keyword search would miss? ### Conference connections - Zero ontology — programme topic link - AI Is Ready. Is Your Knowledge? Why the models we fixate on are rarely the real constraint in creating value from AI — editorial thematic connection ### Read and watch - [From Local to Global: a GraphRAG approach to query-focused summarization](https://arxiv.org/pdf/2404.16130) (research paper). Read the graph and community-summary approach to questions about whole collections. The captured full paper is a preprint; its evaluated global-question results are not a blanket advantage over all RAG systems. - [GraphRAG global search](https://microsoft.github.io/graphrag/query/global_search/) (technical documentation). Understand how community reports support collection-wide answers. Describes this implementation, not a comparative evaluation. - [GraphRAG DRIFT search](https://microsoft.github.io/graphrag/query/drift_search/) (technical documentation). Explore an approach that combines community context with more detailed retrieval. Describes this implementation, not a comparative evaluation. - [Tom Swann: exploring AWS Neptune](https://tam-iii.medium.com/exploring-aws-neptune-9925c175a945) (practitioner article). See a practitioner introduction to graph database ideas. Practitioner explanation or experience, not an independent controlled study. - [Intro to GraphRAG](https://www.youtube.com/watch?v=f6pUqDeMiG0) (video). See how relationships can support retrieval; pair with the research paper to understand the evidence and limitations. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. ### Related concepts - depends on consistent identities from → Entity resolution & data quality: Incorrectly merged or duplicated entities distort the graph. - makes relationships visible within → Knowledge readiness & authority: Connections can expose ownership, provenance and missing context. ## Entity resolution & data quality Deciding when two records refer to the same real-world thing, while preserving differences that matter. ### Why it matters A person, organisation or product can appear under several names. Joining those records can reveal connections; joining the wrong records manufactures connections. This is a central quality problem for any useful graph. ### Key distinctions - Matching names is not the same as identifying the same entity. - A false merge invents a relationship; a missed merge hides one. - A model-generated match should carry evidence and uncertainty rather than becoming an unquestioned fact. ### A question for your organisation Which ambiguous names or duplicate records would most distort our decisions if we matched them incorrectly? ### Conference connections - Zero ontology — programme topic link ### Read and watch - [Deep Entity Matching with Pre-Trained Language Models](https://arxiv.org/pdf/2004.00584) (research paper). Study Ditto and the use of language models for matching record pairs. Research findings apply to the studied setting; this is not a universal performance guarantee. - [Entity Matching using Large Language Models](https://arxiv.org/pdf/2310.11244) (research paper). Explore LLM matching and the effects of prompts and explanations. Research findings apply to the studied setting; this is not a universal performance guarantee. - [Observable data quality with Elementary and DataHub](https://tam-iii.medium.com/observable-data-quality-with-elementary-and-datahub-6fa5f92f2c81) (practitioner article). Place individual matching decisions in the wider discipline of monitored data quality. Practitioner explanation or experience, not an independent controlled study. ### Related concepts - establishes identities for → Knowledge graphs & connected retrieval: Edges need stable endpoints to mean what they appear to mean. - needs error analysis from → Evaluation, observability & AI operations: Test false merges and missed matches separately. ## Agents & delegated work AI systems that pursue a goal through a sequence of decisions and tool actions, sometimes coordinating with other agents. ### Why it matters Delegation can connect reasoning to real work, but every added action creates another place where an error can propagate. Designing the workflow includes defining tools, authority, stopping conditions and handover points. ### Key distinctions - A chatbot response and an action in a business system have different consequences. - Multiple agents add coordination choices; they do not automatically improve results. - A technically available tool is not the same as permission to use it for every purpose. ### A question for your organisation What can this agent do independently, what must it escalate, and what evidence shows it completed the job? ### Conference connections - The Agent at Work: Trust and Capability in the Age of Delegation — programme topic link - From AI Assistants to Agentic Delivery — programme topic link - Autonomy at the Frontier — programme topic link ### Read and watch - [Instil: making sense of multi-agent architectures](https://instil.co/blog/making-sense-of-multi-agent-architectures) (practitioner article). Compare network, supervisor and hierarchical coordination trade-offs. Practitioner explanation or experience, not an independent controlled study. - [Kainos: the promise of AI agents](https://www.kainos.com/insights/blogs/the-transformative-potential-of-AI-agents) (practitioner article). Explore proposed applications and delivery considerations from a provider perspective. Practitioner explanation or experience, not an independent controlled study. - [Enzai: agentic AI governance](https://www.enz.ai/product/agentic-ai-governance) (vendor case study). Examine a concrete product framing of autonomy, allowed actions and escalation. Provider-reported experience; outcomes have not been independently verified here. - [Tips for building AI agents](https://www.youtube.com/watch?v=LP5OCa20Zpg) (video). Decide where autonomous tool use earns its additional complexity, cost and oversight. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. - [Context Engineering for AI Agents with LangChain and Manus](https://www.youtube.com/watch?v=6_BcCthVvb8) (video). Learn practical ways to manage limited context instead of putting every document into every request. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. - [Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) (practitioner article). Choose the simplest architecture that improves measured outcomes. Public article reviewed. Page notes that tooling has changed since original publication; use patterns as background, not frozen implementation instructions. Background learning; no AICON appearance or endorsement implied. ### Related concepts - creates new oversight demands for → Human trust, oversight & decision support: Delegation changes what users must understand and supervise. - expands the action surface governed by → Agent security & resilience: Tool permissions and connected systems shape the consequences of errors. ## Human trust, oversight & decision support Designing AI-assisted work so people understand the system’s role, can challenge it and can intervene meaningfully. ### Why it matters A human approval button is weak oversight if the reviewer lacks time, context or the ability to stop an action. Trust becomes useful when it is calibrated to what the system can actually do and the evidence available. ### Key distinctions - Confidence in an interface is different from demonstrated reliability. - Human involvement is different from a human having meaningful control. - Decision support and delegated decision-making require different responsibilities. ### A question for your organisation At the point of consequence, can the responsible person understand, refuse or correct the AI recommendation? ### Conference connections - Why agents make designing for trust harder, yet even more critical — programme topic link - What Artists Can Teach Us About Building AI People Will Actually Trust? — programme topic link - Hercules — programme topic link - Regulating Intelligence — programme topic link ### Read and watch - [NIST AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1) (framework). Use the framework to connect context, measurement and management of AI risks. A framework for structuring work, not proof of compliance or effectiveness. - [NIST Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) (framework). Explore risks and suggested actions specific to generative AI. A framework for structuring work, not proof of compliance or effectiveness. - [AI and the future of psychiatry: qualitative findings](https://arxiv.org/abs/1910.09956) (research paper). Read how clinicians discussed empathy, collaboration and the impact of AI. This analyses responses from a 2019 physician survey; attitudes are not evidence of current model capability or patient outcomes. - [How to Construct Domain Specific LLM Evaluation Systems: Hamel Husain and Emil Sedgh](https://www.youtube.com/watch?v=eLXF0VojuSs) (video). Translate an abstract quality score into tests that reflect a particular product and its users. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. ### Related concepts - should be calibrated by → Evaluation, observability & AI operations: Measured performance helps people decide when to rely on a system. - needs accountability defined by → Operational AI governance: Oversight must have a named owner and a usable escalation route. ## Evaluation, observability & AI operations Testing whether an AI system does the intended job, and recording enough of its real operation to understand changes and failures. ### Why it matters Model, prompt and source changes can alter behaviour even when the application still runs. Teams need examples of acceptable performance, traceable versions and a way to detect regressions after release. ### Key distinctions - Observability records what happened; evaluation judges it against a purpose or criterion. - Offline tests and production monitoring answer different questions. - An LLM judge is another fallible measurement method, not ground truth. ### A question for your organisation Which representative cases must keep passing when we change the model, prompt or knowledge source? ### Conference connections - AI Adoption To AI Operations — programme topic link - Building a Real-Time Voice Agent that Feels Conversational, Not Transactional — programme topic link ### Read and watch - [Tom Swann: AI Adoption to AI Operations](https://tam-iii.medium.com/ai-adoption-to-ai-operations-4d5b58a66640) (practitioner article). Read a conference contributor’s account of moving from integrations to measurement and operational reliability. Practitioner explanation or experience, not an independent controlled study. - [Langfuse: evaluation of LLM applications](https://langfuse.com/docs/evaluation/overview) (technical documentation). Compare evaluation workflows and the roles of datasets and evaluators. Describes this implementation, not a comparative evaluation. - [Langfuse: LLM observability and application tracing](https://langfuse.com/docs/observability/overview) (technical documentation). Understand the trace information needed to investigate an AI workflow. Describes this implementation, not a comparative evaluation. - [Langfuse: link prompts to traces](https://langfuse.com/docs/prompt-management/features/link-to-traces) (technical documentation). Connect observed behaviour with the prompt version that produced it. Describes this implementation, not a comparative evaluation. - [How To Approach Your AI Evals](https://www.youtube.com/watch?v=DZxaPNYi_k0) (video). A quick entry point before the longer domain-specific evaluation talk; the description primarily provides presenter background. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. - [Demystifying evals for AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) (practitioner article). Test the entire agent interaction rather than treating one plausible answer as proof of quality. Public article reviewed. Background learning; no AICON appearance or endorsement implied. ### Related concepts - provides operating confidence for → Adoption, diffusion & organisational change: Production use needs repeatable evidence, not a successful demonstration. - diagnoses failures in → Retrieval-augmented generation: Separate missing evidence from unsupported answer generation. ## Operational AI governance The people, decisions and routines that determine how AI is introduced, used, monitored and changed. ### Why it matters A policy document cannot on its own keep track of deployed systems or resolve a difficult use case. Governance becomes operational through inventories, accountable owners, risk assessment, review and follow-through. ### Key distinctions - An inventory tells you what exists; an assessment examines its risks and context. - A voluntary framework is different from a legal obligation. - A vendor’s compliance feature is not evidence that a customer’s implementation complies. ### A question for your organisation Can we name every material AI use case, its owner, its purpose and the next review decision? ### Conference connections - The Quiet Risk: Why AI Governance Does Not End With the Policy — programme topic link - Building an Enterprise AI Governance Programme in Practice — programme topic link - The Importance of Data and Governance as the Foundations for AI Success — programme topic link ### Read and watch - [NIST AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1) (framework). Use Govern, Map, Measure and Manage to structure the work. A framework for structuring work, not proof of compliance or effectiveness. - [NIST AI RMF Playbook](https://airc.nist.gov/docs/AI_RMF_Playbook.pdf) (framework). Move from framework headings to practical suggested actions. A framework for structuring work, not proof of compliance or effectiveness. - [Enzai: how to build an AI system inventory](https://www.enz.ai/blog/how-to-build-ai-inventory) (practitioner article). Consider the data and ownership a working inventory needs. Practitioner explanation or experience, not an independent controlled study. - [Liberty IT: responsible generative AI journey](https://www.liberty-it.co.uk/news-stories/liberty-it-journey-to-responsible-generative-ai) (vendor case study). See an organisation’s account of embedding governance in development. Provider-reported experience; outcomes have not been independently verified here. ### Related concepts - defines delegation limits for → Agents & delegated work: Agent actions need owners and decision boundaries. - assigns responsibility for → Knowledge readiness & authority: Source ownership is part of operational control. ## Agent security & resilience Protecting the information, tools and operational boundaries an AI system depends on, and preparing to contain failures. ### Why it matters An agent’s risk depends on the systems it can reach and the authority it has. A useful security discussion follows the complete workflow, including external content, tool execution, logs and recovery. ### Key distinctions - A model refusing a request is different from a system enforcing a permission boundary. - Privacy, cybersecurity and service availability overlap but are not interchangeable. - A product security claim is not an independent assurance assessment. ### A question for your organisation If an agent follows the wrong instruction, which actions can the surrounding system prevent or reverse? ### Conference connections - Autonomy at the Frontier — programme topic link - The Agent at Work: Trust and Capability in the Age of Delegation — programme topic link ### Read and watch - [NIST Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) (framework). Review generative-AI risk categories and proposed mitigation actions. A framework for structuring work, not proof of compliance or effectiveness. - [Enzai: agentic AI governance](https://www.enz.ai/product/agentic-ai-governance) (vendor case study). Inspect a provider’s approach to allowed actions and escalation. Provider-reported experience; outcomes have not been independently verified here. - [Options: PrivateMind](https://www.options-it.com/products/privatemind/) (vendor case study). Compare claims about private deployment, confidential computing and observability. This is a product description; no penetration test or independent verification of its security claims is included. ### Related concepts - constrains tool use by → Agents & delegated work: Permissions limit what delegated work can change. - overlaps with control questions in → Data sovereignty, control & access: Both require examining access and dependencies, not just model behaviour. ## Data sovereignty, control & access Understanding who controls data and AI infrastructure, who can access them and which dependencies shape that control. ### Why it matters A data-centre location is only one part of an AI deployment. Operational access, provider dependencies and the ability to move or run a workload also affect the control an organisation retains. ### Key distinctions - Residency describes where data is stored; sovereignty asks wider questions about control and exposure. - Open weights do not by themselves establish ownership of training data or remove operating responsibilities. - Private infrastructure still requires access management and operational assurance. ### A question for your organisation Beyond the hosting location, who can access the data and what happens if we must change provider? ### Conference connections - Residency Isn't Sovereignty: Own Your Exposure — programme topic link - The AI Capability Divide — programme topic link - Story Engine: AI for a Civic Heritage Destination at Scale — programme topic link ### Read and watch - [Options: PrivateMind](https://www.options-it.com/products/privatemind/) (vendor case study). Use a concrete private-deployment proposition to formulate questions about control. Provider-reported experience; outcomes have not been independently verified here. - [AI for NI: a strategic overview](https://www.ulster.ac.uk/__data/assets/pdf_file/0009/1735992/AIforNI-AStrategicOverviewFullDocument.pdf) (policy report). Read the infrastructure, data and skills context behind regional capability. Contextual reading; not proof of a particular product outcome. - [Open AI meets open notes](https://jme.bmj.com/content/50/2/84) (research paper). Examine how access to information can interact with privacy risks. An ethics analysis concerning patient records; it is not a technical assessment of a particular deployment. ### Related concepts - requires controls from → Agent security & resilience: Control claims need an account of actual access and operational protections. - shapes deployment choices in → Adoption, diffusion & organisational change: Infrastructure and access constraints influence which uses are feasible. ## Adoption, diffusion & organisational change Turning AI capability into sustained use across teams and organisations, with the skills, workflows and ownership needed to make it useful. ### Why it matters A promising pilot can stall when it reaches existing processes. Diffusion asks how benefits spread beyond an expert team; adoption asks what changes in day-to-day work. ### Key distinctions - Tool access is different from effective use. - A faster task is different from an improved end-to-end business outcome. - An organisation’s case study is context, not a promised return for another organisation. ### A question for your organisation Which workflow, owner and measure of success will change if this pilot becomes ordinary work? ### Conference connections - Diffusion Wins — programme topic link - The Adoption Gap — programme topic link - AI Readiness Workshop — programme topic link ### Read and watch - [AI for NI: a strategic overview](https://www.ulster.ac.uk/__data/assets/pdf_file/0009/1735992/AIforNI-AStrategicOverviewFullDocument.pdf) (policy report). Explore the regional combination of data, trust, infrastructure and skills. Contextual reading; not proof of a particular product outcome. - [Liberty IT: responsible generative AI journey](https://www.liberty-it.co.uk/news-stories/liberty-it-journey-to-responsible-generative-ai) (vendor case study). Read an account of experimentation becoming reusable organisational capabilities. Provider-reported experience; outcomes have not been independently verified here. - [Tom Swann: AI Adoption to AI Operations](https://tam-iii.medium.com/ai-adoption-to-ai-operations-4d5b58a66640) (practitioner article). Follow the operational work that appears after initial AI features ship. Practitioner explanation or experience, not an independent controlled study. - [Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Enterprise Internal Knowledge](https://www.youtube.com/watch?v=LRGX-gTegVA) (video). Connect the conference knowledge demonstration to the challenge of making internal business knowledge useful to AI. Publisher title and description only; video not watched and transcript not reviewed. Background learning; no AICON appearance or endorsement implied. ### Related concepts - is constrained by → Knowledge readiness & authority: Teams need usable business knowledge as well as tools. - needs evidence from → Evaluation, observability & AI operations: Measures help distinguish sustained value from initial enthusiasm. ## AI across software delivery Using AI in the chain from requirements and design through implementation, testing and operation. ### Why it matters Code generation is one part of delivery. Changes to requirements, review, system design and quality control determine whether faster local work becomes a better delivered product. ### Key distinctions - Producing more code is different from delivering a useful change. - An assistant helping a developer is different from an agent coordinating a delivery workflow. - A legacy-code explanation is not evidence that a proposed migration preserves behaviour. ### A question for your organisation Where is the real delivery bottleneck, and what evidence will show that AI improves the whole flow? ### Conference connections - Beyond Coding — programme topic link - From AI Assistants to Agentic Delivery — programme topic link ### Read and watch - [Kainos: AI Augmented Delivery](https://www.kainos.com/insights/articles/ai-augmented-delivery) (practitioner article). Read a provider’s argument for connecting speed to trusted delivery outcomes. Practitioner explanation or experience, not an independent controlled study. - [Instil: LLM-assisted code modernisation](https://instil.co/blog/llm-assisted-code-modernisation) (practitioner article). Explore the use of language models in understanding and modernising legacy code. Practitioner explanation or experience, not an independent controlled study. - [Instil: agentic software delivery lifecycle](https://instil.co/what-we-do/agentic-product-engineering/agentic-sdlc) (vendor case study). Inspect a proposed end-to-end delivery approach rather than a single coding tool. Provider-reported experience; outcomes have not been independently verified here. ### Related concepts - can delegate parts of delivery through → Agents & delegated work: Tool-using systems extend AI beyond drafting code. - needs regression evidence from → Evaluation, observability & AI operations: Generated changes must preserve required behaviour. ## AI in healthcare & clinical work Applying AI to healthcare tasks while examining the evidence, responsibilities and patient context of each particular use. ### Why it matters Healthcare discussions often jump from capability demonstrations to claims about replacing clinicians. Useful analysis separates tasks, professional attitudes, actual clinical performance and patient outcomes. ### Key distinctions - A survey of clinicians measures opinions, not model accuracy. - Administrative assistance and clinical decision support have different purposes and consequences. - An abstract, preprint and journal version of the same study are not three independent studies. ### A question for your organisation What specific task is being proposed, and what kind of evidence would show that it helps in the intended setting? ### Conference connections - Dr Bot: Why AI Is Coming for Doctors' Jobs? — programme topic link - AI in Health: What Getting It Right Actually Looks Like — programme topic link ### Read and watch - [AI and the future of psychiatry: insights from a global physician survey](https://www.charlotteblease.com/_files/ugd/510dea_6f6fad768dd046ff842e71aa2923f67c.pdf) (research paper). Explore how psychiatrists expected different tasks to change. Survey of professional opinions collected in 2019; not a clinical effectiveness study or a measure of present AI capability. - [AI and the future of psychiatry: qualitative findings](https://www.charlotteblease.com/_files/ugd/510dea_82e3f0bed6b940f1973863cdac061764.pdf) (research paper). Read the themes in clinicians’ written responses about empathy and collaboration. A qualitative analysis of the same global survey programme, not independent replication of clinical outcomes. - [Open AI meets open notes](https://jme.bmj.com/content/50/2/84) (research paper). Examine the intersection of patient access, online resources and privacy. Ethical analysis rather than a clinical trial. ### Related concepts - makes the consequences concrete for → Human trust, oversight & decision support: Clinical work exposes the difference between help, reliance and responsibility. - requires context-specific assessment through → Operational AI governance: Intended use and affected people shape the review needed. ## Creative authorship & human expression Using generative tools within creative work while keeping sight of who sets the intent, makes choices and speaks through the result. ### Why it matters Creative AI raises questions beyond output quality: authorship, participation, provenance and whose perspective is represented. The conference connects this to both production workflows and trust. ### Key distinctions - Rendering an output is different from setting the creative intent. - A generated reconstruction is different from a preserved first-person account. - A provider’s account of a creative project is not an independent audience-impact study. ### A question for your organisation Which creative decisions remain with people, and how can an audience tell whose voice they are encountering? ### Conference connections - Generative AI is the Renderer Not the Author — programme topic link - From AI Production to AI Platform — programme topic link - What Artists Can Teach Us About Building AI People Will Actually Trust? — programme topic link ### Read and watch - [Hamilton Robson: Listening to Belfast](https://hamiltonrobson.com/blog-page/listening-to-belfast-ai-powers-a-new-chapter-in-storytelling/) (vendor case study). Examine an AI interviewing project’s stated intention to preserve participants’ own stories. Provider-reported experience; outcomes have not been independently verified here. - [Hamilton Robson: Powering Belfast’s Human Narrative with AI](https://hamiltonrobson.com/blog-page/powering-belfast-human-narrative-with-ai/) (vendor case study). Explore the relationship between technical delivery and human narrative. Provider-reported experience; outcomes have not been independently verified here. - [NIST Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) (framework). Place creative use within a wider discussion of generative-AI risks and human context. A framework for structuring work, not proof of compliance or effectiveness. ### Related concepts - reveals design questions in → Human trust, oversight & decision support: People need to understand their agency and the system’s role. - needs provenance from → Knowledge readiness & authority: Creative source material and human contributions need attributable context. ## Conversational voice & multimodal interaction AI interaction that combines speech, turn-taking and sometimes other inputs with retrieval or tool use. ### Why it matters A conversational voice experience is a workflow, not just a natural-sounding voice. Interruptions, timing, tool responses and evidence all affect whether it helps the user complete the task. ### Key distinctions - Speech recognition, speech generation and speech-to-speech interaction are different components or approaches. - A fluent voice is not evidence of a correct answer. - A product guide explains implementation; it does not establish superiority over alternatives. ### A question for your organisation What should happen when a person interrupts, changes their mind or asks for evidence while a tool is running? ### Conference connections - Building a Real-Time Voice Agent that Feels Conversational, Not Transactional — programme topic link ### Read and watch - [Amazon Nova Sonic: core concepts](https://docs.aws.amazon.com/nova/latest/nova2-userguide/sonic-core-concepts.html) (technical documentation). Understand bidirectional streaming and the interaction model. Describes this implementation, not a comparative evaluation. - [Amazon Nova: voice conversation prompts](https://docs.aws.amazon.com/nova/latest/nova2-userguide/sonic-system-prompts.html) (technical documentation). Explore how instructions shape spoken interaction. Describes this implementation, not a comparative evaluation. - [AWS: Live Meeting Assistant with Transcribe, Bedrock and Strands](https://aws.amazon.com/blogs/machine-learning/live-meeting-assistant-with-amazon-transcribe-amazon-bedrock-and-strands-agents/) (technical documentation). Read a worked architecture connecting live speech with an agent workflow. Describes this implementation, not a comparative evaluation. ### Related concepts - provides an interaction channel for → Agents & delegated work: Voice requests can initiate tool-using workflows. - needs interaction tests from → Evaluation, observability & AI operations: Correctness alone misses timing, interruption and recovery failures. ## AI in the physical world AI systems whose perception or decisions interact with equipment, manufacturing processes or physical environments. ### Why it matters Physical applications connect software behaviour with sensors, hardware and operating conditions. A useful discussion must ask how an idea survives real constraints and how failure is detected and contained. ### Key distinctions - A digital demonstration is different from operation in a changing physical environment. - Perception, planning and physical action are distinct parts of a system. - General AI risk guidance does not substitute for domain-specific engineering validation. ### A question for your organisation Which physical constraint or failure mode would invalidate a convincing software-only demonstration? ### Conference connections - AI in the Physical World: Prompts are Cheap, Parts Aren't — programme topic link ### Read and watch - [NIST AI Risk Management Framework 1.0](https://doi.org/10.6028/NIST.AI.100-1) (framework). Start with intended context, measurement and risk management. A framework for structuring work, not proof of compliance or effectiveness. - [RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control](https://arxiv.org/abs/2307.15818) (research paper). See how vision-language pretraining is combined with robot trajectories and action tokens to connect language instructions to physical control. Abstract and official project explanation reviewed; reported research settings do not establish industrial safety or general deployment readiness. Background learning; no AICON endorsement implied. - [Open X-Embodiment: Robotic Learning Datasets and RT-X Models](https://robotics-transformer-x.github.io/) (research project). Explore how standardised robot datasets and shared policies support transfer across different robot platforms. Official project methods/results text reviewed; experimental transfer is not proof of dependable operation in every environment. Embedded demonstrations not watched. Background learning; no AICON endorsement implied. ### Related concepts - needs environment-specific testing from → Evaluation, observability & AI operations: Software tests alone cannot establish performance in a physical setting. - needs responsibilities from → Operational AI governance: Operational ownership and intervention arrangements matter where actions have physical effects.