My 8 AI Predictions for 2025 — With Receipts From What I Got Wrong Last Year
Every December, AI Twitter fills with predictions. Most are either so safe they're useless ("AI will continue to grow!") or so wild they're meaningless ("AGI by summer!"). I try to land somewhere in between — specific enough to be proven wrong, practical enough to be useful.
But before I make new predictions, let me be honest about my track record. Last year, I predicted enterprise RAG adoption would be "widespread by Q3 2024." It wasn't. Most enterprises are still in pilot phase. I overestimated adoption speed and underestimated how hard data preparation is. (To be fair, I wrote an entire post about hidden costs — I should have listened to my own advice.)
What I got right: small models getting competitive with large ones, and the rise of agentic patterns. More on both below.
Here are my 8 predictions for 2025. I'm assigning confidence levels so you can calibrate how seriously to take each one.
1. Multimodal AI Becomes Boring (90% confidence)
By mid-2025, "our AI handles text, images, audio, and video" will be as unremarkable as "our app has a mobile version." GPT-4V, Claude with vision, Gemini — multimodal is already here. The novelty is wearing off.
The business implication: stop being impressed by multimodal demos. Start asking what specific multimodal capability actually solves a business problem. A model that can describe images isn't useful unless you have a workflow where image description creates value.
Where I see real multimodal ROI: document processing (invoices, contracts, medical records that combine text, tables, images, and handwriting), quality inspection (visual + sensor data), and customer support (screenshots + text descriptions of problems).
What to do now: Audit your workflows for places where humans currently translate between modalities — reading a chart and typing a summary, looking at a product photo and writing a description. Those are your multimodal opportunities.
2. AI Agents Move from Demo to Disappointing Reality (75% confidence)
The term "AI agent" is overused to the point of meaninglessness. In 2025, the hype will meet reality, and the result will be... mixed.
I think we'll see a clear split: narrow agents that do one thing well (like my customer support agent) will prove their value. Broad "autonomous" agents that try to handle complex, multi-step business processes will mostly disappoint.
Why? Because the real-world environment is messy. Agents need tools, and tools need integrations, and integrations need maintenance, and maintenance needs people. The "AI that runs your business" pitch requires solving dozens of gnarly integration problems that have nothing to do with AI.
My prediction: the MCP (Model Context Protocol) ecosystem will help a lot. Standardized tool interfaces mean less custom integration work. But even with MCP, expect 2025 to be the year agents go from "wow, look at this demo" to "okay, this works for specific tasks but it's not magic."
What to do now: Pick one specific, well-bounded agent use case. Build it. Learn from it. Don't try to automate your entire business with agents in 2025.
3. Small Models Get Dangerously Good (85% confidence)
This is the prediction I'm most confident in. Models in the 1-8B parameter range (Mistral, Phi, Llama 3) are getting good enough for most production tasks. And they're getting good fast.
The business implication is enormous: you can run these models on a single GPU, self-hosted, for a fraction of API costs. For tasks like classification, extraction, summarization, and simple generation, a fine-tuned 7B model often matches GPT-4's performance.
I already use small models in production for my clients — see my cost optimization case study where quantized Mistral 7B handled 24% of all requests at $0.0003/request vs $0.03 for GPT-4.
What to do now: If you're spending >$5K/month on LLM APIs, benchmark your workloads against small models. You might be surprised. Start with classification and extraction tasks — those are where small models shine brightest.
4. AI Regulation Gets Real and Fragmented (95% confidence)
The EU AI Act is being implemented. The US is doing... well, the US thing (sector-specific guidance, no comprehensive law). China has its own framework. This fragmentation is going to be a headache for any company operating globally.
I don't think regulation will slow AI adoption — but it will increase compliance costs, especially for companies in healthcare, finance, and HR. Budget for it. See my responsible AI post for how to turn this into an advantage rather than just a cost center.
What to do now: If you operate in the EU, start your AI Act classification process. If you're in healthcare or finance anywhere, review your existing AI systems for bias testing and documentation requirements. This isn't optional anymore.
5. The "Intelligence Plateau" Creates Opportunity (70% confidence)
Here's my contrarian take: model intelligence improvements will slow down in 2025. We won't see the same leap from GPT-3.5 to GPT-4 repeated. Models will get cheaper, faster, and more efficient — but not dramatically smarter.
If I'm right, this is actually good news for builders. It means the competitive advantage shifts from "who has the best model" to "who builds the best system around the model." Prompt engineering, RAG pipelines, evaluation frameworks, user experience — these become the differentiators. See my prompt engineering patterns and RAG guide for where the value moves.
What to do now: Stop waiting for the next model release to solve your problems. Build systems with the models you have. The gap between "model capability" and "deployed value" is where all the money is.
6. Enterprise Adoption Splits into Haves and Have-Nots (80% confidence)
The gap between companies that are genuinely using AI in production and those still running pilots will widen significantly in 2025. The haves will be 2-3 years ahead in data infrastructure, tooling, and institutional knowledge.
This isn't just about technology — it's about organizational muscle memory. Companies that shipped AI products in 2024 learned lessons about data quality, deployment, monitoring, and user adoption that you can't get from a McKinsey report. Those lessons compound.
What to do now: If you're still in pilot phase, the most valuable thing you can do is ship something to production. It doesn't have to be ambitious. A simple internal tool that uses an LLM. The organizational learning from going through the full lifecycle — from prototype to production to maintenance — is worth more than the tool itself.
7. "Trust" Becomes a Product Feature (75% confidence)
Users are getting more skeptical of AI outputs. "This was AI-generated" is starting to carry a negative connotation. In 2025, I expect to see "verifiable AI" — systems that show their sources, explain their reasoning, and admit when they're uncertain — become a differentiator.
This aligns with what I've seen in hallucination prevention: the systems that say "I don't know" when they're uncertain are trusted more than the ones that confidently make things up.
What to do now: Add source citations to your AI outputs. Show confidence levels. Make "I'm not sure" a valid response. Your users will thank you.
8. The Real AI Skills Gap Isn't Technical (90% confidence)
Everyone talks about the shortage of ML engineers. That's real, but it's not the binding constraint anymore. The actual bottleneck in 2025 will be people who can:
- Identify which business problems are actually AI-shaped
- Manage AI projects (which behave differently from software projects)
- Evaluate AI vendor claims critically
- Bridge the gap between technical teams and business stakeholders
I call these "AI-adjacent" skills, and they're much harder to hire for than pure technical skills. A company with three good ML engineers and no one who understands the business problems will underperform a company with one ML engineer and a team that knows exactly what to build.
For more on this, see my talent strategy analysis.
What to do now: Invest in AI literacy for your leadership team. Not "learn to code" — learn to evaluate, prioritize, and manage AI initiatives. $20K in education saves $200K in bad decisions.
What I'm Actually Doing About All This
Predictions are cheap. Here's what I'm doing in my own practice based on these beliefs:
- Investing heavily in small model expertise. I'm building tooling for fine-tuning and deploying sub-10B models because I believe that's where the cost-performance sweet spot will be.
- Building agent systems with MCP. Not grand "autonomous AI" dreams, but specific, narrow agents with well-defined tool sets.
- Helping clients build data infrastructure, not just AI models. Because the data bottleneck is the binding constraint.
- Pricing consulting around deployed value, not model sophistication. The question isn't "how advanced is the AI?" but "how much business value does it create?"
Ask me in December 2025 how I did. I'll be honest about what I got wrong.
Planning your 2025 AI strategy? I help teams cut through the hype and focus on what actually delivers value. Let's talk.