Get tomorrow's brief in your inbox
Today: Google launched the Gemini Decisions API for low-cost classification and routing, priced at $0.10 per million input tokens with no output-token charge. Holiday Extras reports faster app development and higher booking conversion using Codex, while United's restrictions show why AI travel assistants still need a human checkout step.
Google's new decision API targets routine business sorting, not chatbot conversations. Gemini Decision 1 chooses among predefined options, answers yes/no questions, and scores inputs against a rubric. Available through the Gemini API in Google AI Studio, it accepts text, images, audio, video, and PDFs, making it relevant to developers building support-ticket routing, document checks, or other bounded workflows. Google announcement
The price makes narrow automation worth testing. Google lists $0.10 per million input tokens, with no output tokens billed, and supports up to 1,000 independent questions against a shared input. It reports 150-250 millisecond end-to-end latency in tests covering contexts up to 16,000 tokens, although the API accepts inputs up to 256,000 tokens. Simon Willison also tested its video-question capabilities, providing an example beyond Google's demonstrations. Independent coverage
Start with one decision you can check objectively. If you run a service business, test whether incoming messages belong in "quote request," "existing job," "billing," or "needs review," using previously labeled examples before connecting live systems. The API returns probabilities and confidence scores, but those are not guarantees; choose review thresholds from your own test results and keep consequential actions behind approval. The opportunity is cheaper, faster triage, not replacing judgment with a probability.
Google Tabletop builds shared planning tools from descriptions. Add links, files, and ideas, then prompt @tabletop or choose a template to create interactive itineraries, polls, expense splitters, and similar group tools. Google says access is rolling out through a U.S. waitlist for adults 18 and older; treat it as an experiment for a team outing or informal event, rather than a replacement for your operational systems. Its potential benefit is less coordination overhead, but evaluate access and sharing controls before adding business information. Google announcement · The Verge
Databricks Genie expands into multistep business analysis. Databricks says its updated agent can carry out extended assignments using context from information stored on its platform; TechCrunch's demonstration included analyzing employee-survey weaknesses and summarizing a lost sales opportunity. If you already use Databricks, ask for a pilot against a specific management question, require links back to underlying records, and investigate the usage limits available through Unity Gateway. The value is quicker access to operational problems, but the company's claim of unsupervised completion needs testing against your own data and permissions. TechCrunch
LLM 0.37.2 fixes an API-key handling problem. Simon Willison reports that certain third-party OpenAI-compatible endpoint configurations could accidentally receive the user's configured OpenAI API key, including when a named provider key had not been stored. If you use the LLM command-line tool, update to 0.37.2, review the advisory-linked conditions, and check whether affected requests occurred; if a key was sent unintentionally, revoke and replace it. Using the tool alone does not establish exposure, but checking the configuration can help prevent credential misuse and unexpected charges. Release details
AWS's local emulator offers developers another testing option. Willison describes an official emulator that lets developers test against local versions of AWS services without provisioning a paid API key; it currently requires Docker on Windows or macOS. If your developer maintains AWS integrations, try it against one existing test suite and verify that it supports the services and behavior you need before relying on it. Local testing could reduce paid-service use during development, but it should complement validation against the real environment. Coverage
The practical lesson is how responsibilities were divided: customer-facing experts shaped improvements, while engineers handled reusable packages, architecture, APIs, security, and development workflows. Try that division on one contained internal tool or customer-requested feature, then measure delivery time, defects, and the relevant business outcome. The opportunity is translating operational knowledge into useful software faster, without assuming this company's results will transfer to yours.
For your next promotion, create one approved template with fixed logos, colors, and claims, plus editable fields for offers, locations, and formats. Adobe positions Firefly Creative Production workflows as one option; assess integration effort, learning time, and current pricing before adopting it. Measure production hours and corrections per campaign, because producing more variations only helps if they remain accurate and recognizable.
Compare total monthly cost before choosing: hosting, staff maintenance time, recovery work, and any paid subscription or usage charges. Confirm current plan prices and limits directly with the vendor, and remember that this comparison comes from competitor Zapier, whose convenience claims are sales positioning rather than neutral findings. If you already have technical capacity, self-hosting may fit; if you do not, the cheaper invoice can conceal a more expensive workload.
If you are setting up an AI travel workflow, let it compare options and prepare an itinerary, then hand checkout to the traveler through an authorized channel. For your own business, decide which agent actions you will permit before enabling account changes or purchases. Clear handoffs can prevent failed transactions and support work caused by promises an assistant cannot fulfill.
Create a reusable brand-check prompt. Based on Adobe's consistency findings, try: "Compare this draft with the approved brand guide and offer details below. Flag unsupported claims, incorrect prices, tone mismatches, and unclear promises. Quote each problem and suggest a correction. Do not invent facts." Give a person final approval, especially for pricing and commitments; the immediate benefit is a more repeatable review process. Background
Run a small, read-only routing test. Take 30 historical customer messages with personal details removed, assign the correct categories yourself, and compare those labels with Gemini Decision 1's choices. Record mistakes and confidence scores before choosing when a human must review, and do not let the pilot send replies or modify accounts. This gives you an initial accuracy check before investing in integrations. API details
The useful pattern is narrower automation with clearer boundaries: classify a request, adapt an approved asset, or build a contained feature. This week, prioritize measured pilots and explicit human handoffs over promises of fully autonomous operations; low model prices do not remove integration, review, or maintenance costs.