"2025 was the year AI went from pilot to production. Enterprise AI investment exceeded $100 billion globally, and for the first time, the majority of Fortune 500 companies reported measurable ROI from AI deployments." — Sequoia Capital AI Report, December 2025
2025 was the year AI stopped being a future thing and became a today thing.
Not for everyone — the gap between organizations that are actually deploying AI and those still "exploring" it widened significantly this year. But for the businesses we worked with at BlokAI, 2025 was the year of real production systems, real ROI, and real lessons.
Here's what we built, what surprised us, and where we're headed.
What We Built
The projects we're most proud of this year weren't the most technically complex — they were the ones that actually changed how businesses operate.
A multilingual customer support agent for a Tokyo-based e-commerce company handling inquiries in Japanese, English, and Korean. The system deflects 65% of tickets without human involvement and has a satisfaction score within 3 points of their human team. It took three months to get there, and two of those months were about Japanese language quality, not AI capability. An automated proposal generation system for a professional services firm. Consultants used to spend 4–6 hours per proposal. The AI drafts in under 10 minutes; consultants review and customize in about 45. The firm has increased proposal volume by 3x without adding headcount. An internal knowledge agent for a manufacturing company facing aggressive retirements. Senior engineers with 20–30 years of institutional knowledge were leaving. We built a system that captured their expertise through structured interviews and documents, then made it queryable by junior engineers. It won't replace what's walking out the door, but it means the next generation doesn't start from zero.These aren't moonshots. They're practical, and they work.
What Surprised Us
Speed of model improvement outpaced implementation timelines. We started several projects with one capability set and finished them with a significantly better one — sometimes mid-project. This is a good problem, but it requires staying current in a way that's genuinely demanding. Data quality is still the bottleneck. We knew this intellectually. Experiencing it repeatedly across different clients drives the point home harder. The limiting factor in almost every project was not the AI model — it was the state of the client's data, documentation, and systems. The businesses that had invested in data hygiene in prior years deployed AI significantly faster. "Change management" is not a soft skill. Two projects in 2025 were technically successful and operationally unsuccessful because the human systems around them didn't adapt. Tools don't transform organizations. People do. We're building change management explicitly into every engagement now, not treating it as optional. Japanese enterprises move slower to decide and faster to commit. Once a Japanese enterprise made the decision to proceed, they were extraordinarily organized and thorough. The challenge is the pre-decision phase, which takes longer than comparable decisions in Western markets. We adjusted our approach accordingly and saw better results.What Didn't Go as Expected
We'll be honest: not everything worked.
One client deployed an AI customer service agent before we thought it was ready — they felt competitive pressure and moved faster than the testing protocol warranted. The agent made mistakes in front of real customers. We rolled back, iterated for six more weeks, and redeployed. It's in good shape now. The lesson: competitive pressure is real, but rushing a production AI deployment creates the kind of visible failures that set back internal trust for months.
We also underestimated the appetite for custom model work among mid-market clients. Several clients wanted fine-tuned models for specialized terminology when off-the-shelf RAG implementations would have served them just as well at lower cost. We've improved our diagnostic process for when fine-tuning is actually warranted versus when it's an attractive complexity trap.
What's Coming in 2026
Three themes are shaping our focus heading into next year:
Agentic systems at scale. The shift from single-turn AI tools to multi-step agents is accelerating. We're investing heavily in our ability to design, deploy, and govern complex agent systems for enterprise environments. AI governance and risk. As AI touches more business-critical processes, the questions around reliability, auditability, and responsible operation are becoming central — not compliance afterthoughts. We're building structured governance frameworks into every engagement. The Japan market. We're doubling down on bilingual capability and Japan-specific implementations. The market is real, the timing is right, and there's significant unmet need for AI consulting with genuine cultural and linguistic fluency.2025 was the year the work got real. 2026 is where it gets scale.
Want to implement AI? Book a call at blok-ai.com