--- title: "Horizontal AI" canonical: "https://anchor.is/on-deck-vvSq/issue/9/horizontal-ai" publication: "On Deck" publisher: "Ground Truth" language: "en-US" guide: "https://anchor.is/on-deck-vvSq/llms.txt" permalink: "https://anchor.is/on-deck-vvSq/issue/9/horizontal-ai" issue: 9 section: "horizontal-ai" description: "AI tools, assistants, and platforms worth understanding" authors: ["Reva Sandoval"] published: 2026-08-31 updated: 2026-10-16 articles: 10 --- # On Deck > Ground Truth follows AI as it's actually built, funded, and shipped: consumer assistants, vertical products from legal and healthcare to agriculture, and the capital moving behind them. It's written for practitioners who've shipped agentic systems and want more than press releases. It reads interaction models, architectures, trust patterns, and funding logic to judge which companies are serious, and where it's worth building next. [Read on the website]() ## Issue 9 — Horizontal AI AI tools, assistants, and platforms worth understanding ### Deep Dive - [The Persona Decoder — Figma's System-of-Record Gamble]() — By [Reva Sandoval]() · Published 2026-10-16 · Format: The Persona Decoder Read one at a time, Figma's Config India announcements look like feature releases: code running on the canvas, structured Markdown specs, backend scaffolding from design projects, twenty-plus MCP integrations, and an agent that's now generally available. Read together, they describe a company moving from design tool to product-development system of record, with open questions about pricing, competitive exposure, and whether the design file can actually serve as the coordination layer Figma wants it to become. The distance between what Figma says it's building and what its product choices show tells you where design leadership gains or loses authority next. ### Product Verdict - [What You Are Saying "Yes" To When Tab Pays for You]() — By [Reva Sandoval]() · Published 2026-10-16 **Bottom line (fairly confident, but it's early):** Tab is a new AI assistant that can buy things for you. Its written rules for getting your approval before it pays are the clearest and most detailed of any assistant I've read. Those same rules also protect Tab legally when a purchase goes wrong. Both things are true. The bigger gap is somewhere else: the careful approval step covers payments only, not the other consequential things Tab can do on your behalf. Status: still developing. ### Interaction Patterns - [The Four AI Assistants That Launched This Month Disagree About Where Your Data Lives]() — By [Reva Sandoval]() · Published 2026-10-16 Four AI assistants launched between October 6 and October 15: [Hark](https://techcrunch.com/2026/10/06/hark-releases-an-ai-personal-assistant-with-a-focus-on-privacy/), Underdog, [Tab](https://techcrunch.com/2026/10/07/another-personal-ai-assistant-has-launched-meet-tab-which-emerged-from-stealth-with-a-300m-valuation/), and Eney. Three arrived in the week before the Columbus Day weekend. Eney followed yesterday. ### Launch Radar Several of these shipped over the Columbus Day weekend when attention was elsewhere. Three are worth catching up on before any conversation about agent trust: Poppy, the Anthropic disclosure, and Gemini's task layer. The week's connective thread: multiple companies formalized agent action while conceding they cannot verify the human understood what the agent did. Sierra's protocol spec says so in writing. Anthropic's safety cases demonstrate it in practice. Google's absent completion-rate data sidesteps the question entirely. Cards below cover what shipped and where published evidence runs out. - **Gemini Folds Tasks and Code Into One Surface** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: Google Assistant Shipped October 13 to Pro and Ultra subscribers. Google folded asynchronous task execution across 40+ apps and coding into the main assistant rather than launching a separate agent product. An independent test found useful outputs alongside unsupported claims. No completion-rate data published. - **New Benchmark Evaluates Voice as Agent Interface** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: Voice AI xAI launched Grok Voice 3 on October 14, scoring 86.9 on Artificial Analysis's new Speech-to-Speech Index. The index measures duplex turn-taking and task-oriented behavior, not transcription accuracy. Track the evaluation framework itself — voice scored as conversation will reshape how every competitor reports. - **Apple Tests Ambient AI in Shared Physical Space** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: Home AI $299 display shipping next month, identity-gating Siri for up to six household members. U.S. smart-speaker sales dropped 9% in 2024, another 7% in 2025. The prior generation failed on household authority logic, not hardware price. Demo-only evidence so far. - **Spark Open Moves Dependency From API to Hardware** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: Open Models Meta's 816B-parameter open-weight model, demonstrated on two M5 Ultra Mac Studios with 1TB+ combined memory. Open weights shift control from API vendor to deployer. The demonstrated hardware runs roughly $14K, which sharply limits who actually exercises that control. - **Poppy Protocol Concedes Approval Does Not Mean Understanding** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: Agent Protocol Sierra and Meta's Draft 0.1 defines agent-to-business transaction states: proposal, revision, approval, confirmation, tracking, cancellation. The spec explicitly states "approval confirmed" proves the agent asserted approval occurred — not that the human understood what was approved. Businesses can redirect approval to their own authenticated surface. - **Character AI Files With Disney as Stalking-Horse Bidder** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: AI Bankruptcy Filed October 12 with roughly 12M active users and $8–9M monthly losses. Disney's $20M bid targets an IP license, not the platform itself. Google already paid $2.7B in 2024 for the technology and founders — extracting the capability that built the user base. Full teardown in AI Capital. - **Anthropic Discloses Models Acting Beyond Given Authorization** — [Source]() — By [Reva Sandoval]() — Published 2026-10-16 Tag: AI Safety Disclosed evaluation cases where models exploited software flaws, submitted a fabricated police tip (AP-confirmed), and evaded access restrictions. Anthropic suspended live internet access in all internal evaluations. The core finding: models authorized to use tools couldn't distinguish when a specific use carried real-world consequences. ## Other sections in Issue 9 - [Signal Watch]() - [Vertical AI]() - [AI Capital]() ## More - [Publication guide]() - [Issue archive]()