Google Cloud Next 25: nine customer deployments show automation shifting from drafting to doing

The tell is in Intuit's numbers, not the adjectives

Most of the nine stories Google collected trade in enthusiasm, but Intuit's brings figures. In tax year 2023, TurboTax processed 44 million U.S. returns and $107 billion in refunds on Intuit's GenOS. For this season, the company folded Google Cloud's Doc AI and Gemini into GenOS specifically to widen its "done-for-you" autofill across the 10 most common U.S. tax forms — the 1099 and 1040 variations.

That scoping matters. The automation isn't framed as answering any tax question; it's document extraction bounded to the highest-volume, most standardized forms. Alex Balazs, Intuit's CTO, described it as delivering on the promise to "do the hard work for them." The pattern here is narrow and repeatable data entry, backed by domain expertise, rather than open-ended reasoning — which is exactly the surface where automation tends to hold up at scale.

Deutsche Bank and Seattle Children's automate the synthesis step

Two deployments target the same bottleneck from different industries: turning scattered source material into a usable answer fast. Deutsche Bank's DB Lumina compresses analyst research that "used to take hours or even days" into minutes, while claiming to hold the financial sector's data-privacy requirements. Seattle Children's Hospital built Pathway Assistant on top of clinical standard pathways covering more than 70 diagnoses, letting the tool synthesize text, images and medical literature into a conversational answer for clinicians.

Both are automating retrieval and summarization over a curated, trusted corpus the organization already owns — a decade of clinical pathways in Seattle's case. That grounding is the safeguard. The systems aren't inventing the underlying knowledge; they're accelerating access to material that was already vetted, which is a more defensible use of generative models in high-stakes settings.

The constraints are as revealing as the capabilities

Reddit and L'Oréal both name explicit limits. Reddit's Pali Bhat said Reddit Answers is built to be "grounded in Reddit's existing posts and conversations, so it shows you more of what real humans think versus creating unverifiable perspectives on its own." L'Oréal's CREAITECH lab uses Imagen 3, Veo 2 and Gemini but sets one hard rule: no generating images of people for advertising, "to remain true to human beauty."

These are guardrails written into the product, not afterthoughts. They point to where teams have decided the automation should stop — verifiable grounding for Reddit, human depiction for L'Oréal. The interesting design work in both cases is the boundary, not the model.

Regulated environments become the next automation frontier

"Now, if you can't come to the cloud, Google Cloud will bring AI to you." — Jensen Huang, co-founder and CEO, NVIDIAMontana Labs

The NVIDIA partnership is the piece with the widest reach. Bringing Gemini models on NVIDIA's Blackwell platform to on-premises data centers through Google Distributed Cloud targets the industries and jurisdictions with digital-sovereignty rules that have kept public-cloud gen AI out of reach. For applied teams, this is the constraint most worth watching: the deployment target, not the model, is what unlocks automation in banking, healthcare and government.

The specific implication across all nine cases is that the durable automation wins are the bounded ones — Intuit's ten forms, Seattle's 70 pathways, Reddit's own corpus, and now regulated data staying on-prem. Google's showcase is less a story about model capability than about how tightly each customer scoped where the automation is allowed to operate.