The Analysis Agent
Every question you ask in Analytics Chat, in Report Chat or to Aliveo in Slack is answered by an AI analysis agent. It works the way a good analyst does: it checks what data and context it has, writes code to analyse your data, looks at the result, fixes anything that's off, and explains what it found in plain language.
How a Question Gets Answered
- Gather context. The agent starts with what it needs to know about your business: the accounts and date range in scope, each account's KPIs, funnels and column descriptions, your campaign tags, and your saved context. See How Aliveo Understands Your Data and How the Context System Works. If your request is too ambiguous to answer correctly, it asks you before going further.
- Plan and write code. It decides what to look at and writes Python to load and analyse your data. Each step shows a short progress label, so you can follow along.
- Run it in a secure sandbox. The code runs in a locked-down Python environment. It is checked before it runs, unsafe operations are blocked, and it can only read the files it is meant to.
- Check its own output. The agent reads the result of every step, and looks at every chart it draws, before deciding what to do next. If a step fails or a number looks wrong, it fixes the code and runs it again.
- Answer. Every figure in an analytical answer comes from code that ran for this question, never from the model's memory or an earlier answer. The answer leads with the finding and puts the supporting tables and charts inline.
Fine-Tuned LLMs with RAG-Based Prompts
The agent runs on fine-tuned LLMs with RAG-based (retrieval-augmented) prompts. Instead of loading everything into every request, Aliveo retrieves what is relevant to your question:
- the passages of any knowledge documents you @-mention that best match your question,
- the campaigns whose names best match your question, with their tags, and
- a set of tested analysis functions suited to the dataset and the question, for example for root-cause analysis or forecasting.
This keeps each prompt focused on your question and your business.
Skills
Skills give the agent specialised instructions for a particular kind of task. Aliveo picks the right skill from your question. You can also choose one yourself by typing / in the message box:
- actioning: turns a change you ask for ("pause campaign X", "set tCPA to $40") into a table you can review and Apply.
- budget_allocation: recommends how to split budget, and at campaign level which budgets or bid targets to change, as tables to apply on every connected ad platform. See Budget & Bid Allocation.
- info: answers from context without running analysis.
- brainstorming: strategy and market exploration.
- web_search: pulls in live external information.
- creative_features: labels the videos in a Google Drive folder by creative feature.
Some skills appear only for certain platforms or accounts; actioning, for example, appears only for accounts Aliveo manages. Others are never in the / list and are picked automatically: on LinkedIn accounts, for example, the agent looks up LinkedIn targeting options when you ask about targeting. See Starting a Chat.
Automation Agents
Scheduled Automation Agents use the same building blocks but run step by step. Each step has its own plan, the code that carries it out, and a validation check on the result. On every scheduled run, the steps execute in order with fresh dates, failed code is repaired, and each step's output feeds the next. The finished report is then delivered to you.