Introducing ai_decide: make fast decisions on your governed data
https://www.databricks.com/blog/introducing-aidecide-make-fast-decisions-your-governed-data📌 Databricks 推 ai_decide，捨棄 LLM 做結構化決策
TL;DR：ai_decide 用決策模型取代 LLM，做結構化決策更快更省。
「這張客服工單該分到哪一類？」「這份文件需不需要人工複審？」這類問題根本用不到 LLM 的完整推理與生成能力，但企業卻常常為了方便直接丟給 LLM 處理，結果在大規模場景下付出不必要的延遲與成本。Databricks 這次推出的 ai_decide，就是衝著這個痛點而來。
🤔 不是每個決策都需要「生成文字」
Databricks 觀察到，許多團隊把 LLM 用在不需要複雜推理與文字生成的任務上，像是工單分類、文件是否需要人工複審、該把某個 prompt 交給哪個模型處理。這些看似簡單的結構化決策，卻撐起客戶端不少高規模的企業工作流程，從處理數百萬份文件到控制即時應用程式邏輯都包含在內。問題是，用 LLM 處理這類小型結構化決策，意味著要為額外的延遲與成本買單，而且這個成本會隨企業規模不斷累加。
🧩 決策模型：不生成文字，直接吐出答案與機率
Databricks 指出，像 TypeSafe AI 的 Jev 這類「決策模型」（decision model）正好補上這個缺口：它不生成文字，而是接收非結構化輸入與一組問題，直接回傳決策結果與對應機率。ai_decide 正是由這類決策模型驅動的新 Databricks AI Function。它能在一瞬間針對一段文字評估一或多個問題，每個問題可以回傳機率、從指定選項中挑出一個類別，或是在一個有序量表上給出分數。由於針對快速決策做了最佳化，ai_decide 在類似任務上比起 LLM 有更低的延遲與成本。
🧩 怎麼用：SQL 一行搞定，也能即時互動
ai_decide 目前以 Beta 版本在 Databricks 上線，可以直接從 SQL 呼叫，用來對治理下的資料（governed data）大規模做結構化決策；也能透過 REST API 呼叫，用於即時應用程式與 agent。對已經在使用 TypeSafe AI API 的團隊，ai_decide 直接相容。官方也列出幾個可以立刻嘗試的情境：用一條 SQL 查詢，為每則產品評論標註主要問題類型與退貨意願，再依產品與月份彙總，追蹤重複出現的抱怨；針對 AI 助理收到的各式 prompt，用 ai_decide 判斷所需推理層級與難度，再據此把請求路由到對應的模型；或是在新版本上線前，用 ai_decide 當作裁判，檢查回答是否符合退款政策、完整度評分。
📊 用 Snake 遊戲展示即時決策能力
為了展示 ai_decide 在即時場景的能耐，Databricks 做了一個部署在 Databricks Apps 上的 Snake 遊戲展示：每個 tick 都把當前棋盤狀態送進 ai_decide，詢問蛇該往哪個方向移動。因為整個決策迴圈能在一瞬間內完成，官方表示這代表 ai_decide 可以放在任何即時決策的背後，例如 agent 該呼叫哪個工具、該走哪個分支、下一步該做什麼。
💡 背後有一整個開放生態系
除了 Databricks 自己的 managed function，官方也提到背後存在一個「廣泛且持續成長的開放權重決策模型生態系」，這些模型都能直接在 Databricks 上以 SQL serve 與執行。
🎯 實務啟示
如果你的資料管線裡有不少「分類、評分、決定下一步」這類輕量任務，卻一直用 LLM 硬做，ai_decide 提供了一個值得評估的替代路徑：把這些高頻、低複雜度的判斷從昂貴的生成式呼叫中抽離出來，交給專門的決策模型處理，理論上能同時省下延遲與成本。目前處於 Beta 階段，建議先從非關鍵路徑的分類或路由場景試水溫。
🔗 來源
- 標題：Introducing ai_decide: make fast decisions on your governed data
- 作者／機構：Databricks
- 連結：https://www.databricks.com/blog/introducing-aidecide-make-fast-decisions-your-governed-data
#Databricks #aidecide #DecisionModel #LLM #MLOps #DataGovernance #AIagent #SQL #EnterpriseAI #TypeSafeAI
原始資料 Databricks · 收集於 2026-09-30
摘要原文
Today, we’re launching a new Databricks AI Function ai_decide that makes fast decisions over your governed data. We see a lot of teams use LLMs for tasks that do not require complex reasoning and text generation. Which category does this support ticket belong to? Does this document need human review? Which model should handle this prompt? These questions power many of our customer’s high-scale enterprise workflows from processing millions of documents to controlling real-time app logic. However, using LLMs for these smaller, structured decisions means paying for additional latency and cost that compounds at enterprise scale. A decision model like TypeSafe AI's Jev closes that gap: instead of generating text, it takes unstructured input and a set of questions and returns decisions and their probabilities directly. ai_decide is a new Databricks AI Function powered by a decision model. It evaluates one or more questions against text in a fraction of a second, and for each question, it returns either a probability, a choice from named criteria, or a score on an ordered scale. Because ai_decide is optimized for fast decisions, it yields lower latency and cost than an LLM on similar tasks. ai_decide is now available on Databricks in Beta. Call it from SQL to make structured decisions at scale on your data or through the REST API for real-time applications and agents. For teams already building with the TypeSafe AI API, the ai_decide function is directly compatible. Not sure where to start? If you have unstructured data and need to classify it, score it, or decide what happens next, ai_decide is a good fit. Here are a few use cases you can try today: You receive thousands of product reviews each month. Customers describe everything from confusing setup instructions to incorrect billing charges. With ai_decide , you can tag each review with its primary issue and whether the customer is looking to return the product in one SQL query. Aggregate those tags by product and month to track recurring complaints and see which issues appear most often alongside reported returns. Let’s say you’re a software company building an AI assistant. Your users ask the assistant to do everything from rewriting a sentence to designing a database architecture. You have several models available with different costs and capabilities and want to choose an appropriate model for each request. Use ai_decide to determine the prompt reasoning level and difficulty. Then route the request to the corresponding model. An example result response.answers : Let’s say you run an online retailer with an AI support assistant. Before releasing a new version, you want to evaluate its answers against your refund policy. You have a set of customer questions, generated answers, and reference policies. Use ai_decide as a judge to check whether an answer follows the policy and score how completely it addresses the customer’s request. Here is an example resulting row, showing response.answers : Because ai_decide returns a decision in a fraction of a second, it's fast enough to drive an application in real time, where each decision depends on the state of the moment before. To demonstrate, we built a live demo of Snake, hosted on Databricks Apps. ai_decide plays the game: on every tick, the app sends the current board to ai_decide and asks, which direction should the snake move next? Because the whole decision loop closes in a fraction of a second, you can put ai_decide behind any real-time decision an agent or application needs to make: which tool to call, which branch to take, which action comes next. ai_decide is available now in Beta. Use it in SQL to classify, score, and make decisions across governed data at scale, or call the REST API to add fast decisions to real-time applications and agents. Beyond the managed function, there's a broad and growing ecosystem of open-weight decision models, and you can serve and run any of them directly in SQL on Databricks. See our recent open-Jev walkthrough here . Explore the ai_decide SQL documentation or get started with the REST API . Subscribe to our blog and get the latest posts delivered to your inbox.
由 claude-code/sonnet 自動生成