Skip to main content

AI Assistant Overview

Get help and get things done in Anima using natural language.

Beta: change-making is what's still evolving

The Analyst side — analysis, evidence, and troubleshooting — is mature and in everyday use. The Operator side — making governed changes on your behalf — is where coverage is still expanding and controls are still hardening. Treat the change-making examples as indicative, not a final feature list.


What is the AI Assistant?

The AI Assistant is a natural-language helper inside Anima Ad Manager. It works over the data and configuration for your currently selected domain, and it can:

  • Analyze performance — summarize trends, detect anomalies, and attribute changes across supply and demand
  • Answer questions about your setup and your numbers
  • Query your data — run reporting-style breakdowns and calculations on request
  • Diagnose issues when something isn't working
  • Propose and apply configuration changes — in Operator mode, with your approval (see below)

Two roles: Analyst and Operator

The assistant runs in one of two roles, which you choose in the chat (subject to your account permissions):

RoleWhat it doesCan it change anything?
AnalystGet answers and insightsReads your configuration and performance data, runs analysis, answers questions, and helps troubleshootNo — read-only
OperatorPerform tasks and editsRuns the same performance analysis and can propose and apply governed configuration changesYes — with your approval
  • Analyst is read-only by design. It can explain what a change would involve, but it cannot make changes itself.
  • Operator is the role that can write — it proposes and applies configuration changes, always with your approval.
  • Access to each role is permission-gated. Analyst is the default; Operator requires an additional permission on your account. If you don't see a role you expect, ask your administrator to grant Analyst or Operator assistant access.

How changes work (Operator mode)

When you ask the Operator to make a change, it follows a propose → approve → execute workflow:

  1. Query the current state and identify the impact
  2. Present the change as before/after values, along with the business impact and a rollback plan
  3. Wait for your explicit approval (e.g. "yes", "proceed")
  4. Execute the change — updates carry a reason documenting what changed
  5. Confirm what was applied
About the approval step

The approval step is part of how the Operator is designed and instructed to behave — it proposes changes and asks before acting, and every update it makes requires a reason describing the change. This is a workflow safeguard, not a server-side lock, so review each proposed change before you approve it. Stronger, system-enforced approval controls are on our roadmap.

Changes the Operator makes to wrapper configuration follow the same staged-release model as changes you make by hand: edits stage until you deploy a release. To undo a deployed change, you copy a previous release forward and deploy it — see Rolling back (there is no one-click rollback).


How it works — and why you can trust it

The assistant is more than a chat box over your data. A few things that matter for accuracy and safety:

  • It acts as you. Every read and write goes through Anima's API under your signed-in identity, so the platform enforces your permissions and keeps your data isolated from other accounts server-side, at the data layer — not as a UI convenience. It can't reach data you couldn't reach yourself.
  • Evidence-grounded answers. Performance questions are answered from pre-computed, confidence-scored statistical conclusions — trends with significance, anomaly clusters measured against a baseline, and dimensional attribution — rather than by eyeballing a table.
  • Real computation in a sandbox. For deeper analysis it runs actual Python (numpy / pandas / scipy) over your datasets in a server-side sandbox and works from the computed results, instead of pulling raw row-level data into the conversation.
  • Everything is on the record. Changes the assistant makes are written to the same audit trail as UI and API changes and tagged as agent-made, and each turn and tool call is captured in an execution trace.
  • Bounded and self-correcting. A single request runs within a fixed step budget and a ~300-second cap, with loop and stagnation protection — it produces its best answer and stops rather than spinning.

How to use it

  1. Click AI Assistant in the sidebar (available in both Ad Manager and Account modes, below the divider). It appears only if your account has assistant access.
  2. Choose Analyst or Operator (if you have Operator permission).
  3. Type your question or request.
  4. In Operator mode, review any proposed change and approve it before it's applied.

What you can ask

Analyst — insights and questions (read-only)

You sayWhat happens
"Give me performance insights for this week"Trend, anomaly, and opportunity analysis
"Which slot has the highest CPM?"Ranking by CPM
"Compare this week to last week"Period-over-period comparison
"Why did revenue drop yesterday?"Diagnostic analysis, attributing the change to segments
"Show me underperforming slots"Identifies weak performers
"How is my experiment doing?"Reports current experiment status
"How are my GAM line items delivering?"Reads your GAM delivery and change history, and can flag at-risk campaigns

Analyst — troubleshooting and learning (read-only)

You sayWhat happens
"Why aren't ads showing?"Guided diagnosis
"Help me fix low fill rate"Step-by-step triage
"Check my wrapper installation"Verification walkthrough
"What is header bidding?"Concept explanation

Operator — making changes (with approval)

You sayWhat the Operator proposes
"Set the floor to $1.00 for my article slots"Updates the floor on the named ad slots
"Disable Partner X"Turns off that demand partner
"Create an experiment for new floors"Sets up a wrapper experiment
"Draft a release with these config changes"Stages the changes into a release for you to deploy

Each of these is proposed first — the Operator shows you the before/after and waits for your approval before applying it.

Not every phrasing maps to a shipped action

The Operator can create and update the resource types Anima manages — ad slots, layouts, demand partners and seats, wrapper configs, experiments, and releases, plus GAM ad-unit/size/key mappings. Requests outside that set (for example, orchestrating an end-to-end deploy in one step, or toggling a setting that doesn't exist yet) may be partially handled or explained rather than executed. When in doubt, the Operator will tell you what it can and can't do.


Four investigations the Analyst runs well

The Analyst isn't just a search box — it follows the same structured investigations as the troubleshooting runbooks, doing the scoping, baselining, and attribution for you. These four are its strongest, most-exercised flows. For each, here's the prompt, what it does, and what you get back.

"Why did revenue drop?"

Say: "Why did revenue drop last week?" (name a slot, bidder, or the whole domain if you can)

It does: scopes the entity and time range, establishes a baseline, checks for anomalies, then attributes the change across traffic, supply, and demand to find which segments moved — and checks whether a recent config change lines up with the timing.

You get: what changed, when, and the top contributing factors ranked by impact (with the time range quantified), plus a recommended next step. If it's inconclusive, it hands you the top hypotheses with a validating check for each. Full manual version: Low revenue.

"Which bidder is hurting me?"

Say: "Which bidder is underperforming this week?" or "Is Rubicon healthy?"

It does: pulls per-bidder performance, ranks the seats and flags outliers, checks whether a concerning metric is new or long-standing, and reviews the bidder's configuration.

You get: each bidder rated healthy / degraded / critical, its key metrics compared against baseline and peers, and specific recommendations for the underperformers. Full manual version: Bidder health.

"Give me a health overview"

Say: "Give me an executive overview of my domain" or "How's my performance this week?"

It does: summarizes domain-level KPIs and groups findings Critical → Warning → Healthy, investigating any shift that warrants a closer look.

You get: the headline numbers with severity context and two or three actionable next steps — a good Monday-morning read.

"Audit this slot"

Say: "Audit the article_sidebar_top slot" or "Why is my mobile header slot's fill rate low?"

It does: pulls the slot's configuration (sizes, load, refresh, layout) and performance, breaks down any underperformance by device, browser, or bidder, and checks the slot's place in its layout.

You get: the slot assessed on revenue, eCPM, viewability, and fill rate, a configuration review, and specific recommendations — size, load, refresh, or floor changes. Full manual version: Slot audit.


Example conversation

Illustrative

The figures below are made up to show the shape of a conversation, not real account data.

You: "Set the floor price to $1.25 for all my article slots"

Assistant (Operator): "Here's the change I'd make:

SlotCurrent floorNew floor
article_header$1.50$1.25
article_incontent_1$1.50$1.25
article_incontent_2$1.50$1.25

Likely impact: lower floors may raise fill rate but can reduce average CPM. Rollback: set these floors back to $1.50. Proceed?"

You: "Yes"

Assistant: "Done — floors updated for 3 slots, each with the reason you approved. Deploy a release to apply the changes to your live site."


Tips for best results

Be specific

Less effectiveMore effective
"Update floors""Set the floor to $1.00 for sidebar slots"
"Check performance""Compare this week's CPM to last week"
"Fix ads""Why aren't ads showing on article pages?"

Review before you approve

In Operator mode, the assistant proposes each change and asks for your approval before applying it. Read the before/after and the rollback plan carefully — you're the final check.

Use natural language

You don't need special commands. Just type naturally:

  • "Show me yesterday's top performing slots"
  • "What would happen if I raised floors to $2?"
  • "Help me set up an A/B test"

What the assistant can't do

TaskWhyAlternative
Make changes in Analyst modeAnalyst is read-only by designSwitch to Operator (if permitted)
Write to your GAM accountThe assistant can read GAM — your entity catalog, change history, and delivery — and flag at-risk campaigns, but it never writes back to Google Ad ManagerUse the GAM UI to change GAM inventory
Create partner accountsExternal to AnimaContact the partner directly
Predict exact outcomesAd performance is uncertainRun an experiment
Act across other domainsScoped to your selected domainSwitch domains first

Privacy and data

  • The assistant only accesses your Anima account data, scoped to your selected domain.
  • Conversations are private to your session.
  • When the Operator updates configuration, it must supply a reason describing the change.

Feedback

If the assistant makes a mistake or could be more helpful:


Getting started

Try these starter prompts (Analyst mode, read-only):

  1. "Give me insights on my performance this week"
  2. "What are my top and bottom performing slots?"
  3. "Analyze my revenue trend and flag anything unusual"
  4. "Walk me through setting up an experiment"

Click AI Assistant in the sidebar to get started