> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getthread.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Assistive AI Accuracy Dashboard
> Measure how often Magic AI's priority, category, and title suggestions are accepted, and investigate the overrides behind every exception by board and override source.
The Assistive AI Accuracy dashboard shows how well Magic AI's suggestions are landing on your service desk — and, where they aren't, exactly who or what changed them. Use it to tune AI, defend accuracy numbers in a review, and find the boards where technicians disagree with the AI most.
Open it from **Thread admin → General → Analytics**.
## What this dashboard answers
* How accurate are Magic AI's **priority, category, and title** suggestions this month, and is that improving?
* Which **boards** have the lowest accuracy?
* When a suggestion is changed, **who or what changed it** — a technician, an automation, or a change made in the PSA?
* Which **individual tickets** sit behind an override, and what did the value change from and to?
## Filters and time windows
A filter bar at the top of the dashboard controls both tabs:
| Filter | Options |
| ---------- | ------------------------------------------------------ |
| Date range | Last 3 months, last 6 months, last 12 months, all time |
| Board | A single board, or all boards |
| Source | Multi-select — pick any combination |
| Feature | Prioritization, Categorization, Title |
Changing a filter re-runs the underlying queries rather than re-slicing what's already on screen, so results can take a moment to come back. A **Clear filters** button appears as soon as anything differs from the defaults, and resets everything — including the date range.
## What's on the dashboard
### Accuracy tab
**KPI tiles** show the latest month's accuracy for each feature, each with a small up or down badge comparing it to the month before, plus a tile for total exceptions and a tile for total AI suggestions. Hover any tile for a tooltip explaining precisely what it measures.
**Trend chart** plots monthly accuracy as one line per feature, colored consistently throughout the dashboard — teal for Prioritization, blue for Categorization, orange for Title.
**Magic insights** is a collapsible AI-written narrative interpreting the trend for a leadership audience. It is built from a **fixed trailing three-month window**, not from your current filter selection, so it won't follow you if you widen the date range.
### Exceptions tab
**Override patterns table** groups exceptions by board and override source, and shows per-feature accuracy alongside the volume — so you can read accuracy and the exceptions driving it together in one row.
**Override source** classifies *who or what* changed the AI's suggestion:
| Override source | What it means |
| --------------- | ------------------------------------------------------------------------------------------------------- |
| Technician | A person in Thread changed the suggestion. Genuine disagreement — worth investigating. |
| PSA webhook | The change arrived from your PSA. Also genuine disagreement, but not attributable to a specific member. |
| Ambiguous | No traceable actor behind the change. |
| Automation | A rule or automation changed the value, not a person exercising judgment. |
| Stale snapshot | A side effect, such as a board move, rather than a real decision about the suggestion. |
**Clicking a row opens a drawer** listing the underlying tickets, each showing the AI-suggested value next to the final value, with search, sort, and export.
Both the table and the drawer export to CSV.
## Metric definitions
| Metric | Definition |
| --------------- | ---------------------------------------------------------------------------------------------- |
| Accuracy | The share of AI suggestions for that feature that were kept, rather than later changed. |
| Exception | One case where Magic AI suggested a value and it was subsequently changed. |
| AI suggestions | The total number of suggestions Magic AI made in the period, and the denominator for accuracy. |
| Override source | The business classification of who or what changed the suggestion — see the table above. |
## Things to watch
**The Exceptions table total will usually be smaller than the Exceptions KPI tile.** By default the table shows only **Technician**, **PSA webhook**, and **Ambiguous** overrides, because those are the ones that represent real disagreement. The KPI tile and the header scope note count **all five** override sources. This is intentional, not a discrepancy.
* **The earliest date shown is the earliest accuracy data recorded for your workspace** — not a display limit. If your history is shorter than 12 months, **All time** and **Last 12 months** legitimately show the same range.
* **The header scope note** — for example *"Aug 2025 – Jul 2026 · 7 boards · 412 overrides"* — is computed from the same filtered and grouped data as the table beneath it, so the two are meant to agree.
* **Changes made directly in the PSA cannot be credited to a member.** This is a PSA API limitation around impersonation, so those overrides land in the PSA-attributed sources rather than naming a technician. See [*Things to know & gotchas*](/analytics/things-to-know-and-gotchas).
* **Every figure is computed upstream** — the dashboard only formats what it receives, so a number here matches the same metric elsewhere in Magic Analytics.
* **Data refreshes daily.** Today's suggestions and overrides generally appear tomorrow.
## Next steps
Act on low accuracy by improving your AI configuration.
Ask follow-up questions on any tile in plain English.
## Related topics
- [Magic Analytics Topics You Can Explore](/analytics/topics-you-can-explore.md)
- [Magic Analytics FAQs](/analytics/faqs.md)
- [Magic Analytics: Things to Know and Gotchas](/analytics/things-to-know-and-gotchas.md)