"Gemini is our most capable and general model, beating GPT-4 on 30 of 32 widely-used academic benchmarks used in LLM research." — Google DeepMind Blog, December 6, 2023
Every December, the predictions pieces come out. Most of them are either extrapolations of current trends or wishful thinking dressed up as analysis. I'm going to try to be more disciplined: here's what I think will actually matter for SMBs in 2024, anchored in what's already in the pipeline.
What I'm Confident About
1. Longer Context Windows Change What's Possible
GPT-4 Turbo, released in November 2023, has a 128,000 token context window — roughly 100,000 words. Anthropic's Claude 2 has 100,000 tokens. These aren't theoretical maximums; they're usable at production scale.
What this changes: you can feed entire contracts, full meeting transcripts, whole product documentation sets, or months of email threads into a single prompt. The "chunking and retrieval" problem that RAG exists to solve becomes less acute for some use cases. You can ask questions about a full document without worrying about retrieval precision.
In 2024, I expect to see a wave of applications built around "dump the whole thing in and ask questions" — simpler to build than RAG, powerful for document-heavy workflows. Legal review, financial analysis, research synthesis. Watch this space.
2. AI Agents Move From Demo to Production
"AI agents" — autonomous systems that can take multi-step actions, use tools, and work toward goals without continuous human input — got a lot of attention in 2023. AutoGPT was the headline grabber. It mostly disappointed in practice: too unreliable, too prone to getting stuck in loops or making bad decisions.
But the underlying patterns are maturing. Function calling (the ability to have LLMs trigger specific API calls or database queries) is working reliably. Chain-of-thought reasoning has improved. The tooling (LangChain agents, emerging alternatives) is getting more robust.
I expect 2024 to be the year that narrowly scoped agents — handling specific, well-defined workflows with clear guardrails — start shipping in production. Not general autonomous AI, but "run this specific 8-step workflow when triggered." Research agents, data collection agents, outreach agents, report generation agents. Narrow, reliable, valuable.
3. Model Commoditisation Accelerates Cost Reduction
GPT-3.5-turbo cost $0.002/1k tokens at the start of 2023. By November it was $0.001/1k for input, $0.002/1k for output on the Turbo model. Open source alternatives have further compressed the pricing pressure. The trend is clear: LLM API costs are declining rapidly.
This changes the economics of high-volume applications. Use cases that were marginal at 2023 pricing become clearly positive at projected 2024 pricing. I expect the cost-per-token to drop 30-50% again in 2024. Plan accordingly — don't let today's unit economics kill projects that will work fine in a year.
4. Multimodal Becomes Practical for Business
GPT-4 Vision is in early access. Audio in/out is available in the API. Multimodal capabilities are real and improving. In 2024, I expect to see production deployments of:
- Invoice and receipt processing using vision
- Product catalog management with image understanding
- Voice-interface AI assistants for field workers
- Automated analysis of charts, diagrams, and visual reports
These aren't the headline multimodal demos. They're the boring, high-value business applications.
5. The Data Infrastructure Gap Becomes Unavoidable
Every AI project we've worked on in 2023 hit data quality as a major bottleneck. In 2024, the businesses that invested in data infrastructure — clean CRM data, properly structured document management, accessible knowledge bases — will have a significant advantage over those that didn't.
I expect to see more businesses treating data infrastructure as AI-readiness investment. Not "data strategy" in the abstract, but specific projects to clean data, centralise knowledge, and build the infrastructure that makes AI applications better.
What I'm Less Certain About
Real-time AI in video/audio communication: The technology is advancing fast but I'm not sure the business applications will mature enough to matter in 2024. Possible, not certain. AI in legal and regulated industries at scale: The tooling is improving but compliance frameworks are still catching up. Progress in 2024, but production-scale in regulated industries will take longer. Open-source models catching GPT-4 quality: Llama 2 is impressive. Mistral (a new French model from late 2023) is remarkably capable at its size. But GPT-4's reasoning quality for complex tasks remains ahead. I think this closes significantly in 2024 but GPT-4-tier open-source models for complex reasoning are probably a 2025 story.What This Means For Your Planning
If I were advising an SMB on AI priorities for 2024, I'd say:
- Get your first RAG system running if you haven't. The foundation skill for most business AI.
- Evaluate long-context use cases for your document-heavy workflows. Simpler than RAG, worth exploring.
- Start thinking about agent workflows for repetitive multi-step processes. Not all of them are ready, but some are.
- Invest in data quality. Whatever you spend here multiplies the value of your AI applications.
- Build internal AI literacy. The organisations that will execute fastest are the ones where non-technical staff understand enough to identify AI opportunities in their own workflows.
The fundamentals don't change: start with problems, not technology. Measure what you build. Expand what works.
Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.