# Insights and the AI analyst The findings surface, last-mile emissions, and the conversational analyst you can ask about your own operation Section: app. Audience: public. Updated: 2026-10-10. --- Two pages in the **"Analytics"** section turn your data into something to act on. ## Insights **"Insights"** has three views: **"Overview"**, **"Findings"** and **"Emissions"**. A finding is something the platform has noticed and thinks is worth your attention. The overview summarises; the findings view is the list. Findings carry their evidence. **"View evidence"** opens what the finding was based on, so you can judge it rather than take it on trust. Use that before acting on anything surprising. Insights are produced on a schedule by the platform rather than on demand, so a finding reflects the last run, not this second. ### Emissions **"Emissions"** reports last-mile emissions: a monthly total for the final delivery leg, rolled up nightly, with the emission factor and data-quality rating each period was built from. A figure is produced once a parcel is delivered and its lane resolves, so until the first delivery through a connected carrier the view reads **"No emissions calculated yet"**. It took over from the CO₂ view the retired Dispatch page had, and old links to that view open here. When the figures were built from substituted inputs, such as a catalogue lane rather than one your carrier supplied, the view says they are estimates, and the export is offered as **"Export estimate (CSV)"** instead of **"Export CSRD Scope 3 Cat 9"**. Do not file an estimate as a compliance figure. ## The AI analyst **"AI Analyst"** answers questions about your own operation in plain language: your orders, shipments, carrier performance, alerts, returns and findings. It looks the answer up in your Horizon data before replying, and says what it looked at. It is signed-in only. Signed out, the input says so plainly rather than failing when you press send. ### Starting a conversation An empty chat opens on a short briefing: tiles for **"Open alerts"**, **"In transit"**, **"Low stock"** and **"Out of stock"**, with today's counts. Clicking a tile asks the analyst about it. Below them, starter questions such as **"Review open alerts"**, **"Carrier performance"** and **"Stalled shipments"** do the same. Or type your own question and press Enter; Shift and Enter adds a new line. While it works, the analyst shows what it is doing, such as **"Searching your orders"** or **"Analyzing carrier performance"**, and for how long. If an answer takes a while, **"Stop waiting"** frees the page: the answer keeps processing and appears in the conversation's history. If a message fails, **"Retry"** sends it again. ### Reading an answer Each answer can carry: - charts of the figures it used, such as **"Carrier on-time rate"**, **"Median transit time by carrier"** and **"Open alerts by severity"**, when the answer has numbers worth drawing; - a confidence rating: **"High"**, **"Medium"** or **"Low"**; - **"Sources"**: the orders, shipments, carrier metrics, alerts and findings it relied on. Click one to see its details. A finding's details offer **"View in Insights workbench"**; - **"Suggested next steps"**, on the latest answer: follow-up questions with the reason for each. Clicking one asks it. **"Copy"** copies an answer. **"Good answer"** and **"Poor answer"** rate it. ### Conversations **"History"** lists your past conversations, grouped by **"Today"**, **"Yesterday"**, **"This week"**, **"This month"** and **"Earlier"**, each with its message count. A conversation started from a finding is named after it. Click one to reopen it. **"Search conversations…"** filters the list by title. Each row can be renamed or deleted; deleting asks first. **"New conversation"** starts afresh. On a narrow screen the list is a drawer that **"History"** opens. ### What it can and cannot do - It reads; it does not act. It cannot buy a label, change an order, send an email or edit a setting. A suggested next step is a question to ask, not an action it takes. - It sees only your account's data, with personal details such as customer names left out. Inside the Shopify admin, it sees only the store you are signed in from. - For forecasts, it reads the latest results of the nightly playbooks, such as demand forecasts, delivery risk, customer segments and carrier recommendations. It does not run them again, so a playbook that has not run yet has nothing to offer. - It is built to answer only from the data it looks up, never to invent figures, and it rates its own confidence. Still check anything expensive, as below. ## Investigating a finding The two pages are joined. A finding in **"Insights"** can be handed to the analyst to look into, which opens the analyst with that finding already in hand. The link that does this is tidied out of the address afterwards, so going back does not re-trigger the investigation. ## What to trust The analyst is a reading of your data, not an oracle. Treat a number it gives you the way you would treat a colleague's summary: useful, worth checking against the page that owns the data before you act on something expensive. Your usage of the analyst is metered. See [Usage and limits](/app/usage-and-limits).