The boring AI features are the ones shipping

The flashy AI demos slowed down this quarter. What sped up is the boring stuff. Translation that works on a phone call. A tool that drafts a planning report. OCR that reads 50 languages. A companion that tells creators when to post. None of it is a moonshot. All of it just shipped into products people already use. Let me catch you up on what crossed the line from demo to default.

Translation stopped being a feature you turn on

The bar for voice translation moved fast. Gemini 3.5 Live Translate now does speech-to-speech in over 70 languages, stays a few seconds behind the speaker, and keeps their tone and pacing. No more turn-by-turn pauses. Google Meet jumped from 5 languages to 70, and from English-only to over 2,000 language pairs in one meeting.

The real signal is where it lives. Grab is testing it on driver-passenger calls, and those users make over 10 million voice calls a month. That is not a demo. That is plumbing inside a product at scale.

For your team, this means live translation is now an expectation, not a differentiator. If your product touches calls, support, or anything cross-border, someone will ask why you don't have it. Have an answer ready.

The tool does the grunt work, the human signs the name

The pattern that keeps showing up: AI does the heavy lifting, a person stays the decision-maker. The clearest case is the UK's AI planning prototype built with Gemini. Planning officers spend hours cross-referencing policy docs and PDFs. Householder applications are nearly 70% of all applications. The tool consolidates the data, flags relevant policies with exact citations, summarizes objections, and drafts the report.

But the officer reviews every line, edits the reasoning, and approves or rejects. The prototype records each step so there is an audit trail. The goal is to cut decision times by 50%, not to remove the human.

That split is your design template. The win is in the draft and the summary. The trust comes from keeping a person in the loop and showing the work. If your AI feature hides its reasoning, you are building the wrong half.

Grounded beats clever

The quiet fix this quarter was making AI stop guessing. Google's Agentic RAG in the Gemini Enterprise Agent Platform added a step that checks whether it actually has enough information before answering. If a question asks for three things and the search found two, it flags the gap, names what is missing, and searches again instead of bluffing.

The numbers back it. They report up to 34% better accuracy on factuality datasets, and 90.1% correct answers even when the system has to pick the right database out of four. Latency stayed within 3%.

The lesson for your roadmap: "sounds confident" is not the same as "is right." If your team is shipping anything that retrieves company data, the question to ask is whether it knows when it doesn't know. That gap is where users lose trust.

Specialized still wins on the edges

Not everything needs a giant model. PP-OCRv6 ships OCR in three tiers, from 1.5M to 34.5M parameters, with 50-language support and 83.2% recognition accuracy on the medium model. The tiny one runs on edge devices. It plugs into Transformers, ONNX, or Paddle backends, so you fit it to your stack.

Speed is getting the same treatment. DiffusionGemma generates blocks of text at once instead of word by word, hitting up to 4x faster output on a single GPU. The catch, stated plainly: quality is lower than standard Gemma 4, and the speedup only helps local, low-traffic use. For high-traffic cloud serving it can cost more.

Take the tradeoff seriously. A small, fast, specialized model often beats a frontier model for a narrow job. Match the tool to the task instead of reaching for the biggest thing on the shelf.

The deep cut

The split this quarter is about who AI replaces versus who it assists. The features that shipped clean all picked a side: assist. The UK planning tool drafts but the officer signs. The creator companion in Facebook's reimagined Creator Studio drafts comment replies in the creator's own tone, but the creator edits and approves before posting. And Deezer's Remix Lab let fans remix songs only with artist consent and pay, while competitors lean on AI-generated covers.

That is your actual decision. When you add an AI feature, are you removing a person or handing them a better draft? The assist version keeps trust, keeps the audit trail, and keeps the human accountable. The replace version is faster on the slide and harder to defend in the room. Pick on purpose, not by accident.

Three questions for your team

1. For our next AI feature, does the human stay the decision-maker, and can a user see the reasoning before they trust it? If not, what do we change before launch?
2. Where are we reaching for a big model when a small, specialized one (OCR, fast local generation) would do the job cheaper and faster?
3. Does anything we ship that retrieves data know when it doesn't have enough to answer, or does it just guess confidently?