Graham Gillen
At the start of the year, our CTO Phil Lewis published his annual predictions blog — seven trends he expected to define AI and search in 2026. With the year now at its midpoint, it’s a good time to look back and ask: how is he doing?
The short answer: pretty well. But the nuances are worth unpacking.
1. MCP Becomes Standard Infrastructure — Hit
This one landed cleanly, and arguably faster than Phil expected. Model Context Protocol (MCP) has become the de facto integration layer for AI-native applications in H1 2026. Developer adoption accelerated dramatically, major platforms shipped MCP support, and the ecosystem of connectors and tooling grew from a niche technical curiosity into something closer to a standard.
Phil’s caveat about memory footprint and scalability challenges for very large systems was also fair — that’s still a real engineering consideration. But the directional call was right: MCP is infrastructure now, not a beta experiment.
To put that in concrete terms: I’m writing this post having spent the morning using Claude with MCP connections into HubSpot and ZoomInfo — Claude as the natural language interface, structured business data as the fuel. That would have been a custom integration project six months ago. Today it’s a morning workflow.
2. Agentic AI Moves Into Production — Directionally Right
Agentic AI is genuinely in production. Real enterprises are running real workflows on agentic systems — not just POCs and demos. The shift from “assistant that responds” to “system that plans and executes” is happening.
That said, “autonomous teammates” still overstates where most deployments actually are. What we’re seeing with our own customers is a deliberate, phased approach — start with AI-powered search, then layer in agentic capabilities like generating presentations from research results. Cautious and incremental, not autonomous. Phil’s qualifier — “still maturing” — was the right hedge. Correct on trajectory, not yet on destination.
3. Multimodal Search Becomes the Default Interface — Partially There
The multimodal capability story is strong. Text, voice, image, and document inputs are all well-supported by leading AI platforms, and consumer experiences have moved quickly. But “default interface for search” is still a stretch for most enterprise environments at the midpoint.
With the exception of specialized use cases — automotive diagnostics being one example — the majority of our customers’ implementations remain firmly text-based, sometimes extending to numeric analysis or structured data. Most knowledge workers are still typing queries into text boxes — just smarter ones. This prediction may prove correct by year end, but it hasn’t landed yet.
4. AI Governance and Provenance Become Mandatory — Strong Hit
Probably the most validated prediction in the list. The combination of EU AI Act compliance requirements, high-profile hallucination failures in enterprise deployments, and growing board-level scrutiny of AI risk has made provenance and explainability genuine procurement criteria — not just compliance checkbox items.
In practice, this is one of the most common stumbling blocks we see when organizations try to move from pilot to production. Governance and security get treated as an afterthought during the sandbox phase — and then become a hard blocker when it’s time to go live.
If you’re selling AI to a regulated enterprise in 2026 and you can’t answer “where did that answer come from?”, you’re losing deals. Phil saw this coming.
5. The Great AI Reality Check — Strong Hit
This is playing out exactly as predicted. CFO-level pushback on AI spend intensified in Q1 and Q2. The “what’s the ROI?” conversation is everywhere. Enterprises that invested heavily in AI experimentation over the past two years are now being asked to show results — and many are finding that hard to do.
A common mistake we see is executives leaning on falling token prices as a proxy for control over costs. It’s a false comfort. As agentic AI scales, token consumption scales with it — Goldman Sachs Research forecasts a 24-fold increase in token consumption by 2030. What starts as a manageable line item can quietly become a significant operating cost if consumption isn’t actively governed. ROI conversations need to account for that.
The tolerance for undefined value and poorly scoped AI projects has dropped sharply. Practical use cases embedded in core workflows are winning; speculative capability showcases are getting cut. Phil called this one well.
6. Small, Specialized Models Overtake General-Purpose LLMs — Too Early to Call
The underlying trend is real. Domain-specific fine-tuned models are gaining serious ground for well-defined enterprise tasks — legal document review, technical support, structured data extraction — and open-source models are seeing broader enterprise adoption, particularly where data control and governance are priorities.
Some observers are making a far more radical claim — that networks of smaller models have already overtaken frontier AI on speed, accuracy, and cost, and that centralized AI has permanently lost the lead. It’s a provocative argument worth reading.
But for most enterprises, the question is more practical than structural. Phil’s prediction was always about adoption patterns, not industry architecture — and at the midpoint, the evidence is directionally correct but not yet conclusive. Worth watching in the second half.
7. RAG Evolves to Agentic and Knowledge-Driven Architectures — Hit
This one is well on track. The Retrieval-Augmented Generation implementations that matter in 2026 look very different from basic vector search pipelines. Multi-step reasoning, structured data integration, knowledge graph augmentation, and agentic retrieval planning are all moving from leading-edge to mainstream practice among serious enterprise deployments.
The “skilled researcher rather than simple search tool” framing holds up well. This is where the industry is heading, and it’s where the real differentiation lives. Our CEO, Kamran Khan, makes a related case in his recent post — that the content processing layer beneath RAG is what separates production-quality AI from demo-quality AI — and it’s proving out in practice.
The Verdict
Five clear hits, two still playing out — that’s a strong track record for any technology predictions blog at the midpoint of the year. The overarching theme of the post — that 2026 would be the year AI moved from experimentation to execution — is holding up well.
The two open questions (multimodal as the default interface, and small models overtaking large ones) aren’t wrong so much as they’re early. Both trends are real; the timing is the only thing in question.
If this were a report card, it’d be a solid B+. We’ll pick this up again when Phil publishes his 2027 predictions.
In the meantime, if you want to talk through what any of these trends mean for your own AI and search programs, get in touch. That’s exactly what we’re here for.
Phil Lewis is CTO of Pureinsights. His original 2026 predictions can be found here.