December 8, 2025
For most organisations, 2025 has really been the birth of this discipline — and it’s fair to say it hasn’t been without challenges.
The year began with a lot of pressure on technology leaders to answer the question, “What are we doing with AI?” While the question was ambiguous, the underlying tone wasn’t. What people were really asking was: “How is this going to affect the productivity of my teams?” and “What savings in people costs can we achieve next year?”
Most have focused on individual enablement — figuring out how to issue licences (not an easy task for many) and encouraging experimentation. Almost everyone has struggled with how to measure and report ROI on this work.
I think there are fewer cynical detractors now. Yes, there are still major questions around safety, ethics and inclusion, but only the most hardened sceptics are dismissing AI’s utility outright. Most customers I speak to say something like, “We don’t need convincing that it can work — we need help making it work.” The challenge now is one of adoption.
Our key learnings with AI are that we must concentrate (rather than replace) human capability; use AI to deliver real business value (not novelty); and prioritise quality and lead time to value (over speed or cost). Lean thinking remains an excellent partner to AI.
For someone who doesn’t like to make predictions, one I am confident in is that many in our industry are going to face a tough time in 2026 when their annual budget cycle — spent enthusiastically in 2025 on AI tooling roll-outs — comes up for renewal. The line item is not small, and justifying the ROI claims made last year is going to be difficult. Some will have no choice but to “roll back the roll-out”, which will only create further disruption and distraction.
Models and tooling will continue to improve, but for most large organisations the underlying tooling changes matter far less than the people and process changes required to use them effectively. We often say that we could happily stick with the tooling from March 2025 and still revolutionise how a pioneer team delivers value. The one thing I would like to see is more live, interactive, multi-modal tooling that enables cross-functional collaboration, reasoning and agreement — and does so earlier in the cycle. That would have a far greater impact than yet another round of fancier code generation capabilities.
Which brings me to my hope for 2026: that organisations stop obsessing about the individual and start focusing on the collective — and the system.
I’d love to see organisations create a small number of new pioneer teams: a group of the right-mindset people, tasked with solving a gnarly business problem (perhaps a legacy replacement or a critical component refresh), and empowered to streamline quality and value through the responsible application of AI. Pair this with cross-cutting team enablement — both discipline-focused (Product & Design, Data, QA) and platform-focused — to allow stream-aligned teams to stay focused on delivering business goals.
I started 2025 thinking the big models were getting close to a performance peak, and that making them bigger wasn’t necessarily leading to better results. I assumed things would stabilise, we’d get a moment to breathe, and we could focus on exploiting what we already had. How wrong I was. Almost immediately, DeepSeek dropped — proving you can catch up with the big models if you really want to, and that models can be genuinely good at reasoning. In fact, reasoning improved across the board thanks to techniques like chain of thought, mixture of experts and DeepThink. Multi-modality took a big step forward too. Progress continued on all fronts.
Fortunately, I’ve also started seeing real application — real value being delivered. Customer service and experience is a standout example. I’ve been genuinely impressed by some AI-enabled customer bots; the audio processing is amazing and the conversational handling feels very natural.
AI-supported software development has also become seriously impressive. I began the year a huge Cursor fan and I’m ending it a huge Claude Code fan. I’m sure next year it’ll be something different again — ideally something that can build a lasting understanding of the architecture, processes and tools I consistently use on a project. But I’ve also had to learn discipline with these tools. If “vibe coding” was the word of the year for 2025, then “Software Doom Loop” should definitely have been the phrase.
RAG also became mainstream in 2025. There are now plenty of out-of-the-box ways to deploy it, and they work well for some use cases. But for others, RAG alone is not enough. No matter how strict your system prompt is, models will still hallucinate or focus on the wrong thing. Critical layers like result re-ranking, agentic verification and reordering of results still need to be added.
And of course, 2025 was the year everyone started talking about agents. The idea has been around for years, but maybe with LLMs their time has finally come. “Agent” is a very flexible term, so in the second half of the year I made it my mission to understand what people actually meant. From what I’ve seen, depending on who you ask, agents can be:
And unsurprisingly, plenty of vendors have started selling “agent frameworks.”
In 2026, companies will realise that a clean, business-friendly data layer isn’t a luxury — it’s a prerequisite for high-value AI use cases. Many of the agent-driven workflows organisations want to build are impossible without coherent, accessible data underneath. To adopt AI at scale, organisations will need to provide AI friendly data with:
In 2025 the front runner has been the MCP protocol — essentially an LLM-friendly API layer that standardises how AI systems access tools and data. Early adopters have already begun experimenting, and the potential is huge. MCP could become the foundation for reusable, trustworthy AI-ready data assets. But there are still deficiencies – will MCP mature to meet these or will a better protocol take its place?
If 2024 was the year AI went mainstream, 2025 became the year open-weight models rewired the entire landscape. The “Sputnik moment heard around the world” arrived in January with DeepSeek-V3, a breakthrough that challenged Western assumptions about scale, investment and dominance. Suddenly, the frontier no longer belonged exclusively to billion-dollar labs — ingenuity and optimisation mattered more than raw compute.
Then came another shock – OpenAI released something ‘Open’: GPT-OSS:20B and GPT-OSS:120B. The 20B model made high-performance local inference feel trivial; the 120B model showed that frontier-class capability could be both open and deployable. The combination with NVIDIA’s pocket-scale DGX SPARK created something unimaginable even 18 months earlier: serious AI running on your desk, not in a hyperscale datacentre. AMD’s Strix Halo joined the party by proving consumer hardware could now field impressive models without melting, marking the first real shift in the silicon balance of power in a decade.
Meanwhile, AI shifted from a tool to an interface. AI browsers — Perplexity, Commet, OpenAI’s Atlas, and Arc’s agent-first experiments — triggered the first credible threat to Google’s search dominance since 2004. For the first time, users weren’t browsing; they were delegating.
Then DeepSeek dropped yet another bombshell: DeepSeek-OCR, nominally an OCR model but in reality a blueprint for dense, AI-native information exchange — compressing structure, semantics and reasoning into a single pass.
2025 will be remembered as the year agents appeared everywhere and value appeared nowhere, but the foundations were laid.
If 2025 was chaotic, experimental and occasionally incoherent, 2026 will be the year AI becomes tangible, valuable and cultural. Businesses will finally push agents into production, not as flashy PoCs but as measurable systems — reducing operational drag, accelerating document workflows and quietly replacing entire classes of manual tasks. The “agent stack” will become as normal as the LAMP stack once was.
But the real story will unfold on the consumer side. The battle for attention — ignited by AI browsers — will sharpen dramatically. Interfaces built around search → conversation → action will become the new homepage of the internet. Traditional browsers will feel archaic next to agent-first experiences where the system reads, reasons and acts on your behalf.
This raises the biggest question of all: will 2026 be OpenAI’s iPhone moment? Rumours of a personal device — something between a SPARK-class palm computer and a wearable agent — won’t go away. If it ships, the centre of gravity in consumer tech could shift overnight. Apple, long viewed as untouchable, now looks uneasy as Siri Next and VisionOS race to remain relevant.
On the frontier, quantum computing will finally escape the lab. Not the sci-fi version — but hybrid AI-quantum optimisers for logistics, finance, and materials, enough to trigger analyst speculation about a “super-bubble” where AI and quantum feed each other’s momentum.
And underpinning everything is the open-weight surge. Model quality will plateau, optimisation will explode, and 2026 may be the year when open beats closed not ideologically, but practically.
All business models have their pros and cons. But, when you consider the type of problems we help our clients to solve at Equal Experts, it’s worth thinking about the level of experience and the best consultancy approach to solve them.
If you’d like to find out more about working with us – get in touch. We’d love to hear from you.
