Quick Navigation
- Separating signal from noise
- Multimodal AI — beyond text
- On-device AI — local, private, fast
- Open source and democratisation
- Protocols and the agent economy
- Regulation and governance evolution
- AI-native organisations
- The horizon scan in one diagram
- Cheat sheet
Separating signal from noise
Most AI predictions are wrong. Not because the technology doesn't advance — it does, rapidly — but because people predict specific capabilities rather than directions.
The AI space is full of hype. Every week there's a "breakthrough" that will "change everything." Most won't. Separating signal from noise is the most important skill for navigating AI's future.
A simple framework helps:
| Question | If yes → Signal | If no → Noise |
|---|---|---|
| Is it already happening? | Working products exist | Only demos or papers |
| Are multiple players pursuing it? | Independent validation | Single company claim |
| Is there a clear trajectory? | Each version better than last | No clear progression |
| Does it solve a real problem? | Clear use case | Solution looking for a problem |
Signal is already happening, with multiple independent sources and a clear trajectory. Noise is a single announcement with no independent validation and a vague timeline.
Most AI technologies move through a predictable cycle: technology trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment, plateau of productivity. Being too early or too late both cost you. The key is to invest during the slope of enlightenment — when real value is emerging but before widespread adoption.
This guide focuses on trends that are already past the trough.
Try it yourself — The signal test
Think of three AI trends you've heard about recently. Run each through the signal test. How many are signal? Focus your preparation on those.
| Trend | Already happening? | Multiple players? | Clear trajectory? | Solves real problem? | Signal or Noise? |
|---|---|---|---|---|---|
Multimodal AI — beyond text
Most AI interaction today is text-based — type a question, get a text answer. But the world is multimodal. People communicate through text, images, audio, video, and gestures. AI is catching up.
Multimodal AI models can understand and generate across multiple formats — text, images, audio, video, and code — in a single system.
| Capability | Status | Impact |
|---|---|---|
| Image understanding | Production (GPT-4V, Gemini, Claude) | Document analysis, visual QA |
| Image generation | Production (DALL-E, Midjourney, Stable Diffusion) | Creative, design, prototyping |
| Audio understanding | Production (Whisper, Gemini) | Transcription, voice commands |
| Audio generation | Production (ElevenLabs, OpenAI) | Voice assistants, accessibility |
| Video understanding | Early production | Content moderation, search |
| Video generation | Early (Sora, Runway) | Content creation, simulation |
The architecture underneath: a multimodal encoder converts all inputs — text, image, audio — to a shared representation. The core LLM reasons across modalities. The output decoder generates in any modality. This is what enables a single model to read a document, look at a chart, listen to a call, and produce a summary.
Multimodal AI expands what agents can do — agents can now see and hear, not just read. It creates new data challenges — images and audio need different pipelines than text. And it expands use cases dramatically: document understanding, visual inspection, voice interfaces.
Try it yourself — Your multimodal opportunity
For your organisation, which modalities matter? Your highest priority is where to experiment first.
| Modality | Current use case | Future opportunity | Priority |
|---|---|---|---|
| Image understanding | Low / Medium / High | ||
| Image generation | |||
| Audio | |||
| Video |
On-device AI — local, private, fast
Most AI today runs in the cloud — your data is sent to a server, processed, and results returned. This creates latency, privacy, and cost concerns. On-device AI runs locally, on your hardware.
Models are being optimised to run on phones, laptops, and edge devices — without cloud connectivity.
| Trend | What's happening | Impact |
|---|---|---|
| Smaller models | Distilled and quantised models run on phones | AI everywhere, not just cloud |
| Apple Intelligence | On-device AI built into iOS/macOS | Hundreds of millions of devices |
| Google AI Core | On-device AI on Android | Privacy-first AI features |
| Edge computing | Models on IoT devices, cameras, sensors | Real-time AI without cloud |
The techniques that make it possible: quantisation reduces precision (32-bit to 8-bit) for smaller, faster models with slight accuracy loss. Distillation trains a small model to mimic a large one — much smaller, retaining most capability. Pruning removes unnecessary parameters. Compilation optimises for specific hardware.
On-device AI enables the privacy-first approach discussed in the Responsible AI training. On-device inference is critical for agents — agents need local tool access. Smaller models trade some capability for accessibility — the right trade-off depends on the use case.
Try it yourself — Your on-device case
For your organisation. Your "yes" answers tell you where on-device AI is worth evaluating.
| Question | Your answer | Implication |
|---|---|---|
| Do you have privacy-sensitive use cases? | If yes: on-device is compelling | |
| Do you have latency-sensitive use cases? | If yes: on-device is compelling | |
| Do you have offline requirements? | If yes: on-device is essential | |
| What devices do your users have? | Determines what's feasible |
Open source and democratisation
A few companies control the most powerful AI models. Open-source alternatives are rapidly closing the gap — and changing the economics of AI.
| Model family | Creator | Notable for |
|---|---|---|
| Llama | Meta | Strong general-purpose models, widely adopted |
| Mistral | Mistral AI | Efficient European models |
| Gemma | Lightweight, on-device capable | |
| Qwen | Alibaba | Strong multilingual capabilities |
| DeepSeek | DeepSeek | Reasoning and code capabilities |
| Open Source | Closed (API) | |
|---|---|---|
| Cost | Infrastructure only | Per-request pricing |
| Customisation | Full fine-tuning possible | Limited customisation |
| Privacy | Full control | Data leaves your infrastructure |
| Capability | Rapidly closing gap | Currently leading edge |
| Support | Community | Vendor support |
| Operational burden | You manage everything | Vendor manages everything |
Open-source models change the operating model decision — build vs buy looks different when the model is free. Open source enables on-device AI — models can be deployed anywhere. Fine-tuning open-source models connects to the Data and AI training — your data is your advantage.
Try it yourself — Your open-source evaluation
For your top AI use case. Count each column. The winner is your model strategy.
| Factor | Points to open source | Points to closed |
|---|---|---|
| Cost sensitivity | ||
| Need for customisation | ||
| Data privacy requirements | ||
| Need for cutting-edge capability | ||
| Available technical expertise |
Protocols and the agent economy
Without standards, every AI system is an island. Protocols like MCP and A2A are creating an interoperable ecosystem where tools and agents can work together across organisational boundaries.
Two open protocols are shaping the future:
MCP (Model Context Protocol) creates a universal tool connection standard. Write a tool once, use it with any MCP-compatible model. Vendor-agnostic. The ecosystem is growing rapidly.
A2A (Agent-to-Agent Protocol) enables agent discovery and collaboration. Agents expose capability cards, other agents find them and send tasks. This enables the "agent economy" — agents as services for other agents.
| MCP | A2A | |
|---|---|---|
| Scope | Tool and data connection | Agent-to-agent communication |
| Creator | Anthropic (2024) | Google (2025) |
| Transport | JSON-RPC | HTTP + JSON |
| Standardisation | Open | Open |
| Ecosystem | Growing rapidly | Emerging |
The agent economy vision: agents offering specialised services to other agents. MCP servers as a marketplace of reusable tools. Cross-organisational AI collaboration without custom integrations. AI capabilities as composable building blocks.
Protocols create ecosystems. Tools built for MCP work with any MCP-compatible model. A2A enables agents to discover and collaborate across organisational boundaries. Early adoption creates competitive advantage.
Try it yourself — Your protocol readiness
For your organisation. Your "yes" answers tell you which protocols to evaluate.
| Question | Yes / No | What this means |
|---|---|---|
| Do you build custom AI integrations? | If yes: MCP could simplify | |
| Do you need AI to connect to multiple tools? | If yes: MCP is relevant | |
| Do you need agents to work together? | If yes: A2A is relevant | |
| Do you work across organisational boundaries? | If yes: A2A is compelling |
Regulation and governance evolution
AI regulation is no longer theoretical. The EU AI Act is now in force as the first comprehensive AI law, and other jurisdictions are following. Organisations that are compliant — or preparing to be — have a competitive advantage over those still treating regulation as a future concern.
| Jurisdiction | Framework | Status |
|---|---|---|
| EU | AI Act | Enacted, phasing in through 2026-2027 |
| US | Executive Order on AI, state-level laws | Fragmented, evolving |
| UK | Pro-innovation, sector-specific approach | Principles-based |
| China | Algorithm regulation, deepfake rules | Active enforcement |
| India | DPDP Act, AI advisory | Evolving |
The EU AI Act risk tiers:
| Tier | Definition | Obligation |
|---|---|---|
| Unacceptable risk | Banned | Cannot be deployed |
| High risk | Allowed with strict obligations | Conformity assessment, documentation, monitoring |
| Limited risk | Transparency required | Disclose AI involvement |
| Minimal risk | No specific obligations | Voluntary codes of conduct |
What organisations should prepare for: AI inventory (know what AI you have and where it's used), risk classification (classify each use case by risk tier), documentation (maintain records of training data, model design, testing), human oversight (ensure humans can override AI decisions), bias monitoring (regular testing for fairness and bias).
The rules are coming. Prepare now, not when they arrive.
Try it yourself — Your regulation readiness
For your organisation. Any "no" is a regulatory gap to close.
| Preparation | Done? | Gap |
|---|---|---|
| AI inventory — know what AI you have | ||
| Risk classification — classify each use case | ||
| Documentation — records of training data, model design | ||
| Human oversight — humans can override AI decisions | ||
| Bias monitoring — regular testing for fairness |
AI-native organisations
Most organisations today are adding AI to existing structures. The end state is organisations designed around AI — where AI is not an add-on but the foundation.
| Dimension | Traditional | AI-Native |
|---|---|---|
| Processes | Designed for humans | Designed for human-AI teams |
| Data | Byproduct of operations | Core strategic asset |
| Decision-making | Human with data support | AI-augmented with human oversight |
| Organisation | Functions organised by skill | Functions organised by AI capability |
| Culture | AI is an IT concern | AI is everyone's concern |
The AI-native stack has three layers: culture (experimentation, continuous learning, data-driven decisions), process (AI-augmented workflows, governance by design, value measurement), and technology (AI platform, data infrastructure, MLOps).
AI-native is the destination that the entire curriculum points toward. Every training addresses one aspect of the AI-native stack. The journey from traditional to AI-native is the transformation addressed in the AI Strategy and AI Adoption trainings.
Try it yourself — Your AI-native gap
For each dimension, where is your organisation? The gap between "where are you" and "AI-native" is your transformation roadmap.
| Dimension | Traditional | AI-Native | Where are you? |
|---|---|---|---|
| Processes | Designed for humans | Designed for human-AI teams | |
| Data | Byproduct of operations | Core strategic asset | |
| Decision-making | Human with data support | AI-augmented with oversight | |
| Organisation | Functions by skill | Functions by AI capability | |
| Culture | AI is IT's concern | AI is everyone's concern |
The horizon scan in one diagram
flowchart TD
NOW["Today\nAI added to existing structures"]
NOW --> TREND1["Multimodal\nAI sees, hears, creates"]
NOW --> TREND2["On-Device\nAI moves to the edge"]
NOW --> TREND3["Open Source\nAI becomes accessible"]
NOW --> TREND4["Regulation\nRules catch up"]
NOW --> TREND5["AI-Native\nOrganisations designed around AI"]
TREND1 --> FUTURE["Tomorrow\nAI embedded everywhere"]
TREND2 --> FUTURE
TREND3 --> FUTURE
TREND4 --> FUTURE
TREND5 --> FUTURE
What to prepare for:
Separate signal from noise — already happening plus multiple sources plus clear trajectory equals signal, and the framework is simple and it works. The organisations that prepare now will adapt fastest to whatever comes.
Cheat sheet — all the key terms
| Term | Plain English | Where it fits |
|---|---|---|
| Signal vs Noise | Already happening + multiple sources + clear trajectory | Evaluating any AI trend |
| Multimodal AI | AI that sees, hears, and creates — not just reads and writes | Beyond text |
| On-Device AI | Privacy, speed, offline, cost — AI moves to where the user is | Edge computing |
| Open Source | Cost, customisation, privacy, transparency — increasingly competitive | Model selection |
| MCP | USB for AI — standardised tool connections | Tool interoperability |
| A2A | Agent discovery and communication protocol | Multi-agent coordination |
| EU AI Act | First comprehensive AI law — risk-based approach | Regulatory compliance |
| AI-Native | Designed around AI from the start — not adding AI to legacy structures | End state |
How to know if this landed
You'll know this has landed when someone has a horizon scanning process — regularly reviewing AI trends and implications. When they can apply the signal vs noise framework to any new AI announcement without help. When multimodal capabilities have been assessed for their business — they know where vision, audio, or video AI fits. When open-source models have been evaluated against their current closed models — they know the trade-offs. When regulatory preparation is underway — AI inventory, risk classification, documentation. When leadership receives regular briefings on AI trends and strategic implications. And when they can describe what AI-native means for their organisation in concrete terms.
What horizon scanning actually does for a team
The signal vs noise exercise is the one that shifts behaviour most immediately.
I bring 15 claimed AI trends to the workshop. Some are signal. Most are noise. The team works through the framework together — already happening, multiple sources, clear trajectory, real problem solved. By the end, they've identified 5 trends that require preparation and dismissed 10 that don't. The conversation changes from "should we be worried about this?" to "here's how we evaluate anything new."
The multimodal assessment tends to surface use cases nobody had considered. Teams that have been thinking exclusively in text-based AI suddenly see applications in document understanding, visual inspection, voice interfaces. The technology isn't changing what they do. It's making visible the applications that were always there.
The open-source evaluation usually produces the most immediate cost conversation. Teams paying premium prices for closed API usage discover that open-source models can handle 60-80% of their current workload at infrastructure cost only. The conversation changes from "which API should we use?" to "which workloads belong on open-source and which need closed models?"
The regulatory readiness assessment tends to surface tensions that were already there but unnamed. Teams that have been treating regulation as a "future concern" suddenly see that the EU AI Act is already in force and phasing in. The AI inventory exercise — simply listing what AI they have and where it's used — is usually the first time anyone has a complete picture. That's not bureaucracy. That's the infrastructure for moving faster with less risk.
Book a Workshop
Ready to see around the corner — without falling for the hype?
or
Half-day workshop includes signal vs noise analysis of current AI trends, multimodal AI assessment for your business, open-source model evaluation for your use cases, regulatory readiness gap analysis, AI horizon scan with strategic implications, and preparation roadmap for the next 12-18 months.