25 June 2026: AI tools start asking to act (AM)

This morning’s AI story is less about who made the loudest announcement and more about which tools are getting close enough to real work to matter. The shift is showing up in software that can click through tasks, design tools that now output code, and infrastructure that aims to make AI cheaper to run and easier to control.

Google has added computer use to Gemini 3.5 Flash, pushing one of its faster models closer to practical task automation. In its official post, Google says the feature lets Gemini interact with on-screen interfaces by understanding what is visible and deciding what to click or type next. That matters because the most commercially interesting AI tools now need to do more than answer questions. They need to move through real software without turning every step into a manual handoff.

For small businesses, the useful question is not whether an AI model can navigate a screen in a demo. It is whether the task is narrow enough to verify. Booking data into a CRM, checking a form for missing fields, or moving through a support workflow are very different from giving a model open-ended control over a laptop. Cristoniq’s explainer on what an AI agent is matters here because the practical version of agent use is still bounded work, with review, permissions and clear failure points.

This is also where product maturity will be judged. If a model can use tools but cannot explain what it did, undo mistakes or respect approval gates, it is still a lab feature dressed up as workflow software.

OpenAI says it is working with Broadcom on a custom inference chip called Jalapeño, aimed at running large language models more efficiently at scale. The announcement comes from an OpenAI post, so the performance framing is vendor-reported and should be treated that way. Even so, the direction is clear: major AI companies want more control over the hardware that serves model responses, not just the models themselves.

The practical impact is cost and availability. If inference becomes cheaper or less dependent on the most constrained GPU supply, AI products can become easier to price, easier to scale, and potentially easier to offer in lower-cost tiers. That does not mean a magic drop in subscription prices next week, but it does mean the AI market is moving from general-purpose cloud dependence toward more specialised infrastructure choices. That tends to shape what ordinary users end up paying far more than a benchmark graph does.

Figma is adding code layers, motion support and more AI features, which is a sign that design tools are increasingly expected to hand work off directly to production. TechCrunch reported that the update introduces code layers, animations and AI-assisted plugin creation. That matters because the boundary between mock-up and working interface is getting thinner. Designers are being asked to produce outputs that engineering teams can use more directly, while AI fills in some of the repetitive scaffolding.

There is a useful caution here. When a design platform starts generating more implementation material, teams need clearer review rather than less. Outputting code faster is only a gain if somebody still checks structure, accessibility and long-term maintainability. If you are rolling AI into creative or product workflows, the best use case is still acceleration on repetitive tasks, not blind acceptance of whatever a model suggests.

NVIDIA’s NeMo AutoModel fine-tuning workflow on Hugging Face points to a quieter but important AI shift: more teams want to adapt models without building a full research stack around them. The Hugging Face post presents a route to fine-tune transformer models with NVIDIA tooling in a way that is meant to be more accessible to developers and ML teams. This is the sort of infrastructure story that rarely leads the news cycle, but it often matters more in practice than a splashier launch.

For businesses experimenting with internal copilots or specialist assistants, adaptation is where the real work starts. A model that is generally capable still needs workflow-specific data, guardrails and evaluation. Cristoniq’s guide to what MCP is and its explainer on AI governance both connect to this point: the challenge is not only model quality, it is how safely the model fits into the systems around it.

Businesses are also starting to discover that AI budgets can be burned on low-value tasks surprisingly quickly. TechCrunch reported that companies are trying to rein in usage as employees spend tokens on small jobs that do not justify the cost. That may sound less exciting than a product launch, but it is one of the most honest indicators of where enterprise AI is heading. The era of asking whether teams will use AI is giving way to the era of asking which uses are worth paying for.

That is a healthier phase for the market. Once cost controls, review trails and usage policies start appearing, AI stops being a novelty budget and starts being treated like software that has to earn its place. The thing to watch next is whether the best AI products make those controls visible enough for ordinary users, because the winners from here will not only be the smartest tools. They will be the ones that can act, but also show their workings clearly enough to be trusted.

Here is everything else worth knowing from today’s AI news.

• Memory-chip earnings are still the clearest commercial read on AI demand, but the revenue story stayed out of the main sections because it is more about market structure than reader action.
• Google’s researcher departures to rivals are notable, but they do not tell readers what changed in products they can use today.
• Washington’s chip-war spillover into Europe remains important background, though it belongs in context rather than as a lead AI Daily story for this slot.
• A new Infosys-linked startup and other funding news suggest services firms still see room to rebuild around AI, but the details were too vague to justify top billing.
• Claims about model extraction or misuse on named companies need a higher verification bar than a fast morning AI Daily run allows, so those items were left out of the main set.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning and evening.