"Gemini 2.0 Flash achieves better performance than 1.5 Pro across key benchmarks at twice the speed, while Sora's public launch signals that video generation has crossed the production threshold." — Google DeepMind / OpenAI, December 2024
End of year. Time for the "predictions for next year" piece that everyone writes and almost nobody gets fully right, including me.
I'm going to try to be disciplined about this: I'll stick to things I'm reasonably confident about based on what's clearly in the pipeline, and I'll clearly label the more speculative parts. I'll also be practical — if something is coming but won't be ready enough for most SMBs to use reliably in 2025, I'll say so.
The Things I'm Confident About in 2025
1. AI Agents Become the Main Conversation
We spent 2024 talking about chatbots and RAG systems — these are real and valuable. In 2025, the conversation shifts to agents.
The tooling has matured significantly. LangGraph (for complex stateful agents), newer frameworks like Vercel's AI SDK, and the increasing capability of function calling make it viable to build production agents that handle multi-step workflows reliably. Not for everything, not with no human oversight, but for specific, well-scoped workflows: the research agent, the data pipeline agent, the report generation agent.
For SMBs, the highest-value agent applications will be the ones that automate recurring, multi-step operational workflows — things your team does repeatedly that involve gathering information, processing it, and producing a structured output.
My recommendation: If you don't have a basic RAG system or AI assistant running yet, do that first. If you do, start mapping your operational workflows for agent candidacy.
2. Open-Source Models Get Competitive With GPT-4
The trajectory of open-source model quality has consistently beaten expectations. Mixtral 8x7B was a revelation in late 2023. The models coming in early 2025 — derivatives of Meta's next Llama generation and from other labs — will push the frontier significantly.
My expectation: by mid-2025, there will be open-source models capable of handling the majority of SMB AI use cases with quality comparable to GPT-4, deployable on manageable cloud infrastructure. This is significant for the data privacy constraint — more businesses will be able to run capable AI without sending data to third-party APIs.
For businesses with real data sovereignty requirements, evaluate open-source deployment options seriously in H1 2025.
3. Multimodal Becomes Standard, Not Special
In 2024, vision capabilities felt like a special feature you added to an AI system. By the end of 2025, I expect sending documents, images, and audio to models to be as natural as sending text.
The business applications: automated invoice processing that can handle any format, product catalog management where images are first-class data, voice interfaces for field workers, video content analysis for media-heavy businesses.
If your business generates or processes significant amounts of visual or audio content, multimodal AI deserves a spot in your 2025 roadmap.
4. AI Cost Continues Declining
Every year, the same story: models get more capable and the cost per token drops. 2025 will continue this trend. My rough expectation: cost per token for frontier model quality will drop 40-60% again in 2025.
This is important for planning. Projects that are currently on the edge of positive ROI will become clearly positive at 2025 pricing. Applications that require high-volume inference — processing every customer interaction, every document, every transaction — become more viable.
Don't let 2024 pricing kill projects that will work fine in 2025. Model your unit economics at projected prices, not current prices.
The More Speculative Things
AI coding tools become central to SMB development: Tools like Cursor, Claude's expanded coding capabilities, and future products will significantly accelerate software development. The implication: custom software becomes cheaper to build. This could mean smaller businesses building more custom AI tooling than they currently consider viable. Real-time voice AI in customer-facing applications: GPT-4o's voice capability is impressive but the API access for third-party development is still limited. If OpenAI and others open this fully in 2025, real-time AI voice in customer service becomes practical. I'm confident this direction is right; less confident about the 2025 timing. Reasoning models become affordable: o1 is expensive and slow. If the cost trajectory of previous model generations continues, reasoning model quality becomes accessible at significantly lower cost. This changes what's viable for complex analysis and decision-support applications.What SMBs Should Actually Do in Q1 2025
If you're in planning mode, here's what I'd prioritise:
- Audit your AI readiness. Do you have your data in order? Clean, accessible, structured? This is the foundation everything else builds on.
- If you have nothing running yet: Pick one high-frequency workflow and build a focused RAG or automation system. The learning is as important as the output.
- If you have something basic running: Evaluate its performance rigorously. Is it actually being used? Are users finding it helpful? Fix what's not working before expanding scope.
- Map your agent opportunities. Look at your most time-consuming recurring operational workflows. Which ones are AI-agentable? This is your 2025 roadmap.
- Invest in AI literacy across your team. The organisations that will execute fastest on AI in 2025 are the ones where more employees understand what AI can do and can identify opportunities in their own workflows. This doesn't require technical training — it requires conceptual literacy.
2024 was the year AI went from experiment to serious business tool for early adopters. 2025 is the year it goes mainstream. Position accordingly.
Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.