By David,
Ask a general-purpose AI chatbot about Azure Virtual Desktop and you'll get a competent explanation of host pools, FSLogix and scaling plans. Ask it which of your host pools cost the most this month, or why logons were slow yesterday, and it has two options: say it can't see your environment, or guess. The second is the dangerous one, because a confident wrong answer about your own estate looks exactly like a right one.
That gap is where an AVD assistant either earns its place or becomes a liability. This post covers what admins actually ask, why the answer has to come from live telemetry, what to check about where your data goes, and where an assistant should stop. The first four sections apply to any tool. The last shows how we built Atlas Advisor, the assistant inside Atlas AVD.
The short version: The useful questions about AVD are about your estate: cost, capacity, health and what to fix next. An assistant can only answer them well if it queries live data and shows the rows behind the answer, so you can check it. Before you connect one, find out which AI provider sees your questions, what environment data goes with them, and who pays for the usage. And keep it read-only, or keep a person approving every change: an assistant that can log users off or delete hosts on its own is a risk most teams don't need to take.
1. What admins actually ask
Spend a week logging the questions your team asks about an AVD estate and they fall into four groups. Each one needs a different source of data, which is the first clue that a chatbot working from its training data can't answer them.
| Question | What the answer needs | Where that data lives |
|---|---|---|
| Cost: "Which host pools cost the most this month?" "Why is spend up?" "Are we over budget?" | Actual and forecast spend by resource, tags and budgets | Azure Cost Management |
| Capacity: "Is there capacity pressure?" "Do we have quota for ten more 8-core hosts in East US 2?" | Session counts against host limits, vCPU quota and usage per region | Host pool data, the Azure quota APIs |
| Health: "Show unhealthy hosts." "Why was login slow yesterday?" "Why are users disconnecting?" | Session host status, logon checkpoints, connection errors | The AVD service, AVD Insights in Log Analytics |
| Priorities: "What needs my attention?" "What should I fix next?" | All of the above, ranked by impact | Every source at once |
The last row is the one people really want answered, and it's the hardest, because it needs the other three combined and weighed. The Cost Intelligence and Environment Health Scoring pages describe the kind of analysis that has to sit underneath it.
2. Why the answer has to come from live telemetry
A language model is good at language. It isn't a database, and it doesn't know what happened in your subscription this morning. For an AVD assistant to be trustworthy, the model's job should be to understand the question, fetch the right data with a defined query, and explain what came back. The numbers should never come from the model's memory.
Three properties tell you whether an assistant works that way:
- It shows the rows behind the answer. "Your most expensive pool is Design, at $4,100 this month" should come with the table it was computed from, so you can check it and spot the pool it missed.
- It's specific about time. "Month-to-date actuals" and "forecast" are different numbers. A good answer says which it used.
- It says when it can't answer. If AVD Insights isn't connected, a question about logon stages has no data behind it. The right answer is "this needs Insights", not a plausible paragraph.
When you evaluate any assistant, ask it a question you already know the answer to, then one where the data doesn't exist. The second test tells you more.
3. Data handling: where your prompts go and who pays
Before you connect an assistant to production, get clear answers to these:
- Which AI provider processes the questions? The vendor's own model, a model the vendor resells, or a provider you choose?
- What goes with the question? Answering "why is pool X slow?" means sending some environment details (pool names, host names, user principal names, timings) to wherever the model runs. Find out exactly what.
- Can it run in your own tenant? If your policies require AI processing to stay in your Azure tenant, check whether you can point the assistant at an Azure OpenAI resource you own.
- Whose terms apply? If you bring your own provider key, that provider's data-use and retention terms govern your prompts. Read them for the API tier you're on, not the consumer app.
- Who bills the usage, and is it marked up? Bring-your-own-key means the provider bills you directly. A bundled assistant means the vendor pays and recovers it somewhere.
- Does anything work without AI at all? An assistant that can still answer common questions with no model connected is easier to approve, and it keeps working when a key expires.
4. What an assistant shouldn't do unattended
The step after answering is acting: "you have three idle hosts; shall I shut them down?" It's tempting, and in an AVD estate it carries real risk.
- Disruptive actions hit people. Logging off a "disconnected" session loses someone's unsaved work. Draining or restarting a host during business hours moves users mid-task.
- The model can be steered by data. Host names, ticket text and event log messages all end up in the model's context. A crafted string in any of them is an attempt to give it instructions. A read-only assistant limits that to a misleading answer; an assistant that can act turns it into a change.
- Mistakes need an owner. When a change goes wrong, someone has to be able to say who decided to make it and why.
A sensible line: an assistant can read, rank and recommend freely. Anything that changes your environment either stays with a person, or runs only from a small, closed list of reversible actions with explicit approval and an audit trail. Destructive actions (deleting hosts, removing users, changing scaling during working hours) shouldn't be on the list at all.
5. How Atlas Advisor does it
Atlas Advisor is the assistant built into Atlas AVD, included on every plan, Community included. It runs inside your Atlas deployment, in your own Azure subscription, and answers from the same live data the rest of Atlas uses.
It answers from your data, and shows it
Ask "Show expensive host pools" and Advisor runs a defined query against your live cost and host pool data, writes a one-line answer, and puts the rows underneath, with a filter and an export. The screenshot below is from our demo environment, so the names and numbers are sample data.
The answer states that the figure is month-to-date actual spend, says how many rows it used, and lists them. If a number looks wrong, you can see why without opening the Azure portal.
Two ways to run: built-in, or with your own AI provider
- Built-in, with no AI key. Advisor's built-in engine answers a fixed set of common questions from your live data: "How many users are logged in?", "Show unhealthy hosts", "Why was login slow yesterday?", "Is there capacity pressure?", "Am I over budget?", "Any cost spikes?", "Show unattached disks". Nothing leaves your deployment. Each question is answered on its own, without memory of the previous one.
- With your own AI provider. Connect Claude, ChatGPT, Grok or Azure OpenAI and you can ask anything in plain language, follow up, and have Advisor work through multi-step investigations. Azure OpenAI can be a resource in your own Azure tenant, and it can authenticate with your Atlas deployment's managed identity, so no key needs to be stored.
Recommendations, ranked by impact
Advisor also reviews the environment continuously and lists what to fix next, by priority and category (security, cost, capacity) with the estimated monthly saving where there is one.
Read-only, by design
Advisor answers, ranks and drafts the steps. It can't run a script, change a setting, restart a host or sign anyone out. A person on your team makes every change, through the normal Atlas screens, under their own permissions and in the audit trail. That's the line from section 4, applied.
Where the data goes
- Without an AI key, Advisor's built-in answers never leave your deployment.
- With a key, your questions and the environment details needed to answer them go to the provider whose key you added, and nowhere else. Choose Azure OpenAI in your own subscription to keep them in your tenant.
- Atlas never receives your Advisor conversations. Chat history is stored in your own Atlas deployment.
- The provider bills your AI usage directly. Atlas doesn't mark it up.
Questions to try
With the built-in engine or a provider connected, these make a good first ten minutes:
- "What needs my attention?"
- "Show expensive host pools"
- "Any cost spikes?"
- "Why was login slow yesterday?" (needs AVD Insights; see why AVD logons are slow for what the stages mean)
- "Is there capacity pressure?" (and the quota side in requesting an Azure vCPU quota increase)
- "Show unattached disks"
- "Show open recommendations"
The Atlas Intelligence page covers the analysis engines Advisor draws on, and the AVD sizing guide is a good companion for the capacity answers.
FAQ
Can Atlas Advisor make changes to my environment?
No. Advisor is read-only. It answers, recommends and drafts steps, and a person on your team makes every change.
Which AI providers does Atlas Advisor support?
Claude, ChatGPT, Grok and Azure OpenAI, using your own key. Azure OpenAI can run in your own Azure tenant and authenticate with your Atlas deployment's managed identity.
Do I need an AI key to use Atlas Advisor?
No. Without one, Advisor's built-in engine answers a fixed set of common questions from your live data, and nothing leaves your deployment. A key adds free-form questions, follow-ups and multi-step investigations.
Does Atlas see my Advisor conversations?
No. Conversations stay in your Atlas deployment and go only to the AI provider you connect, if you connect one.
Who pays for the AI usage?
You do, directly to your provider, at their rates. Atlas doesn't mark up AI usage, and Advisor is included on every Atlas AVD plan.
Can a general-purpose chatbot answer questions about my AVD environment?
It can explain AVD concepts well. It can't see your subscription, so any answer about your own costs, hosts or sessions is a guess unless the tool queries your live data.
Atlas Advisor details reflect Atlas AVD as of November 2026.