Usage Dashboard Reference

Every companion tab in the Admin Panel has a Usage sub-tab. First-time viewers have three reactions: “That’s a lot of charts” → “What do these numbers mean?” → “Should I worry?”

This page answers each one.

---

Usage Tab at a Glance

[Screenshot: Ada Dashboard → Usage, full view]

From top to bottom you typically see four regions:

1. Monthly total card — big numbers, glance-friendly
2. 7-day trend line — last week at a glance
3. Token bar chart — time-sliced detail
4. Distribution chart — toggle “by model” or “by purpose”

---

Region 1: Monthly Total Card

Cumulative since the first of this month:

---

Region 2: 7-Day Trend Line

[Screenshot: 7-day trend line example]

Y axis: daily spend (USD) X axis: date

What to look for:

• Stable trend — roughly consistent daily usage is normal
• Unusual peaks — a 10× day suggests an AI loop
• Weekend vs weekday — reflects how the business runs
• Holidays — support workloads should drop during long breaks

---

Region 3: Token Bar Chart

Same data, bar chart format. Each bar is a day (or an hour, depending on your slice).

Toggle: look for a Daily / Hourly button above the chart.

Input vs Output use different colors:

• Lighter = input tokens (your prompt + conversation history)
• Darker = output tokens (AI’s reply)

Output typically accounts for 70–80% — replies are longer than prompts.

---

Region 4: Distribution Chart

This is the most useful chart — shows where the money actually goes.

Two toggles:

By Model (Provider)

Percentage share per model:

GPT-4o-mini   ████████████████████████  60%
GPT-4o        ████████                  25%
Claude Haiku  ████                      15%

Use to judge: are expensive models eating the budget?

By Purpose (Usage Type)

Percentage share per feature:

Reply generation  ████████████████  55%
Tool calls         ██████████        30%
Memory system      █████              15%

Use to judge: which feature is expensive? If memory burns 50%, conversation history is too long and needs compression.

[Screenshot: distribution-by-purpose example]

---

Real Debugging: Reading the Charts

Case A: Yesterday’s Spend Doubled

Steps:

1. Trend line → identify the spike day
2. Switch to Hourly bar chart → find the spiking hour
3. Go to Activity tab and look at that hour’s conversations
4. Common causes:
   • One customer kept asking long questions
   • AI got stuck in a loop (talking to itself)
   • Expensive model accidentally used for bulk work

Case B: Projected Monthly Cost Is Close to Over-Budget

Steps:

1. Divide “month total” by elapsed days → daily average
2. Multiply by total days in the month → projected monthly
3. If over, switch to “by model” to find the heaviest spender
4. Reduce that model’s usage

Case C: Conversations Stable, Cost Spiked

Usually output tokens or context length exploded:

1. Look at input vs output ratio — output at 90%+ means replies are too long
2. Cap maxTokens (see Cost Optimization)
3. Check whether context compression is enabled

---

Exporting Data

The Usage tab typically has an Export CSV button (top-right). Export 30 / 90 / 365 days.

Columns include:

• Timestamp
• Model
• Purpose
• Input tokens
• Output tokens
• Cost (USD)
• Triggering platform and user

Hand it to Vi: “Tell me which time slot cost the most this month, and which model had the best ROI.”

---

Related Docs

• How Tokens Are Priced
• Budget Alerts
• Cost Optimization