How XeroFlow uses AI chat, anomaly detection, semantic search, and automation recipes to keep your agency running smarter.
11 min read
AI Chat
Natural language queries grounded in your agency data.
Anomaly Detection
8 analyzers surface spend spikes, deadline risks, and bottlenecks.
Semantic Search
Meaning-based search powered by Cloudflare Vectorize.
XeroFlow includes a conversational AI assistant powered by Groq that understands your agency's data. Ask questions in natural language -- "What did we spend on Meta Ads for Acme Corp last month?" or "Show me overdue tasks on the Q1 Campaign board" -- and the AI fetches the relevant context from your tasks, clients, projects, and financial records.
Use @mentions to pin specific entities into the conversation context. Type @Acme Corp to reference a client, @Website Redesign for a project, or @Homepage Mockup for a specific task. Mentioned entities are fetched and placed at the top of the AI context window, so responses are grounded in real data rather than generic advice. Conversations can be pinned for quick access later.
@Mentions Ground Every Response
When you @mention a client, project, or task, the AI fetches that entity's real data and pins it to the top of the context window. Responses are always grounded in facts, not guesswork.
The proactive AI agent runs eight independent analyzers in the background, each focused on a specific risk area. These analyzers scan your data on a schedule and surface findings before you need to ask. The eight analyzers cover spend anomalies (unusual daily or weekly spend spikes), budget thresholds (approaching or exceeding limits), deadline risks (tasks likely to miss their due dates), and workload imbalances (team members with disproportionate assignments).
Additional analyzers monitor invoice aging (overdue client payments), campaign performance degradation (declining metrics over time), stale tasks (items with no updates for an extended period), and approval bottlenecks (client approvals stuck waiting). Each finding includes a severity level, supporting evidence, and a recommended action.
Traditional keyword search misses results when people use different terminology. XeroFlow's semantic search uses Cloudflare Vectorize to understand the meaning behind your queries, not just the exact words. When you search for "logo project for the coffee brand," it finds tasks related to branding deliverables for your cafe client, even if no task contains those exact words.
Behind the scenes, tasks, clients, briefs, and knowledge base entries are embedded as vectors using a change-detection system. Only modified items are re-embedded, keeping the index fresh without reprocessing your entire dataset. The search combines semantic similarity with recency scoring, entity importance weighting, and diversity penalties to return the most relevant and varied results.
Every query to the AI system first passes through an intent classifier that determines what kind of question is being asked. The classifier distinguishes between financial queries, process questions, task searches, time tracking inquiries, and general knowledge requests. Each intent type triggers a different retrieval strategy optimised for that data domain.
Classification runs on the edge using Cloudflare Workers AI for sub-50ms latency, with a Groq fallback for ambiguous cases. The system also supports LoRA adapter routing, meaning you can fine-tune classification accuracy over time using your own agency's conversation data without retraining the base model.
Automations in XeroFlow follow a trigger → action model. A trigger is an event that happens on a board -- a status changes to "Done," a due date arrives, a person is assigned, or a new task is created. An action is what happens in response -- send a notification, post a message in a chat channel, update another column, move the task to a different group, or send an email via Resend.
You build recipes visually by selecting a trigger type, configuring its conditions, and chaining one or more actions. Recipes are scoped to a specific board and can be toggled on or off individually. Common patterns include notifying an account manager when a task is marked for review, emailing a client when a proof is uploaded, and archiving tasks automatically when their status reaches "Completed."
The AI agent does not wait to be asked. Based on the anomaly detectors and your team's activity patterns, it generates actionable recommendations that appear in the Activity Hub. These suggestions range from "Consider redistributing 4 tasks from Sarah to Tom this week" to "Acme Corp's Meta spend is trending 18% above budget -- review campaign targeting."
Recommendations are scored by relevance and urgency, then delivered through the "For You" tab in the Activity Hub. The system learns from your interactions -- if you consistently dismiss a certain type of recommendation, it reduces the frequency. If you act on recommendations, it surfaces similar insights more prominently.