Overview

The Dema Agent is an AI analyst built into Dema. Ask a question in plain language to explore your data, create a chart or table, build a dashboard, or write an analysis.

---

​

Your first conversation

Open Agents and enter a question: If you are not sure where to start, choose one of the suggested prompts. Conversations are saved, so you can return and continue where you left off. The agent keeps the earlier messages as context for follow-up requests such as “Now break that down by market.”

---

​

The message box

The message box also lets you attach files, select tools, dictate messages, and choose a model: The @ and / menus also work in instruction fields, such as routine instructions, agent instructions, Settings instructions, and skill bodies. Files are not available there because those inputs are not tied to a conversation.

---

​

Model selection

The model picker in the message box lets you choose the AI model for your conversation. You can select a specific model or use Auto.

​

Auto model selection

Auto picks a model for you at each turn depending on the task complexity: a quick lookup gets a light model; a turn that needs deeper reasoning gets a stronger one. Based on our internal benchmarking, Auto is the better default for:

• Short sessions - under ~7 turns. One-off questions, ad-hoc analysis, exploration. Around 20% cheaper than typical current usage with same response quality. Single-question sessions save the most, closer to 40%.
• Sessions where difficulty varies turn to turn - a few demanding turns among lighter ones.

For longer iterative work - building an app, refining a skill, developing an analysis over many turns - we recommend choosing one model and staying with it for the session. These conversations build up shared context, and changing model mid-thread means rebuilding it, which costs more than the per-turn saving is worth.

---

​

The workspace panel

The workspace panel on the right has two tabs:

• Workspace shows the agent’s plan and the charts, tables, or apps it creates.
• Files shows the agent’s files and the outputs from the current conversation. See Files and memory.

---

​

Competitor ad research

The agent can look up a competitor’s public ads in the Meta (Facebook and Instagram) Ad Library and the TikTok Ads Library. Search by brand name or keyword, and narrow the results by country and date. For Meta ads, you can also filter by active or inactive status. These are built-in tools, so you don’t need to connect an integration. For some ads, the agent can also return details such as reach or impressions by region.

---

​

Instructions

Instructions tell the built-in assistant how to respond. Use them for language, formatting, or company-specific guidelines that should apply to every conversation. Manage them at Agents → Settings → Instructions. Yours are private by default; share one and your teammates can enable it too. Keep instructions short and direct, for example: “Always report in EUR” or “Round percentages to one decimal.” Use a skill instead when the guidance applies only to a particular task.

---

​

Going further

---

​

Getting good answers

1. Ask a specific question. “Show revenue by channel for the last four weeks” is clearer than “How is revenue doing?”
2. Include the comparison or business context you need, such as “Compare it with the same period last year.”
3. Use follow-up messages to refine the result. The agent keeps the earlier messages as context.