Insights

Insights reads the record of how your organization used AI over the past few weeks and returns a handful of findings. A finding is a conclusion with an action attached: one frontier model is a third of your spend and no routing policy covers it, five Connectors were authorized and never called, four tools keep firing together and should be a skill. You are not handed a chart to interpret.

The analysis that produces them is called Dreaming, and it works in two stages, both overnight while nobody is using the product. Every night it reduces the day’s telemetry to a per-day summary. Once a week it reads back across those summaries and works out what is worth telling you. Insights is at wf-develop.merge.dev/insights/findings, organization-wide, visible to admins.

Nothing is ever applied on your behalf. Every finding is a recommendation that waits for a person to accept it.

Findings, not metrics

Insights is the findings and nothing else, each one a conclusion with a fix attached, with a summary row across the top carrying the severity of each for anyone scanning.

The metrics underneath them, with no recommendation on top, live on the Dashboard. Reach for those when you want to check a number yourself or slice spend a way no finding slices it. Findings answer “what should we do.” The Dashboard answers “what is the number.”

What Insights reads

Insights uses telemetry you already produce. There is nothing to instrument, no new retention surface, and the sources compose. A missing signal never blocks a night’s run: Dreaming degrades to what it has, and each finding records which signals were available when it was written, so an organization running with tool call payloads redacted can see what it is not being told.

Dreaming: how findings are generated

Each night, Dreaming reduces the day’s Tool calls and Gateway request logs to a per-day summary for your organization. Once a week, each analysis reads the span of those days it needs and writes a finding only if its pattern is present, which is why the page shows three findings some weeks and five others. Raw telemetry is read once, on the night it lands, so a ninety day window costs no more to analyze than a seven day one and survives your log retention window expiring underneath it.

Everything is scrubbed on the way in, using the rules you already run: the same redaction path as AI Guardrails, with your organization’s Standard Entity Rules and Custom Regex Rules. Personal information and identifiers are removed, and what remains is aggregated to the group level. The one exception is LLM spend, which stays attributable to an employee key so cost findings can name where the money goes. Prompt text, tool call arguments, and tool results are never carried past that first reduction, and no finding quotes them.

Insights reads only telemetry you already retain under your existing log settings. It creates no new retention surface, and turning payload redaction on costs you the analyses that need call content while leaving sequence, frequency, and failure patterns intact.

Each analysis reads the window that suits the pattern it looks for. A co-occurrence pattern is read over days, because it has to be current. An idle grant is read over weeks, because a shorter window catches people on vacation. The window a finding used is stated on its card.

The first week is empty. Insights holds its findings until your organization has produced roughly a week of usage, because below that a pattern and a coincidence look the same. Individual analyses cross their own threshold at different points, so the page fills in over the first few weeks rather than arriving complete.

Finding types

Severity is shown as a badge on the card and a colored chip in the summary row. High and Medium are exposure, something costing money or sitting open. Opportunity is upside, and reads blue rather than a warmer color so a page carrying upside alongside exposure does not read as all problems.

Five analyses run today. New analysis types ship on their own cadence and start appearing the week they go live, each with a fix of its own, so expect this list to grow.

Overpowered direct calls, flexible tier candidates, and cache opportunities read Gateway request logs. Connected but unused and suggested skills read Agent Handler tool calls.

Every fix has a limit worth knowing before you commit to it. A routing policy applies from the moment it is saved and does not retroactively cover calls already made. Revoking a grant leaves the Connector in your catalog, and its owner is prompted to reconnect on next use. Session boundaries behind a suggested skill are inferred from a 30 minute idle gap, so check a set against a couple of real sessions before publishing it organization-wide. The flexible tier trades latency for price, so move one job at a time rather than in bulk.

Reading a finding

Collapsed, a card shows its category, its severity badge, the headline stating the conclusion, and the action button. The action is always a fix verb, never “view” or “review.”

Expanding adds the evidence in the order you would check it:

• The evidence sentence, opening with the window used and rounding its figures.
• Stat tiles, each with the comparison it needs to be read honestly. “Share of spend: 30%” sits above “$8,100 of$27,000.”
• Charts and tables showing the specific rows the conclusion rests on: the grants being revoked, the co-occurring tool sets, the jobs moving tier.
• A pointer to wherever the fuller picture lives, as a sentence rather than a second button, since one of the two is the fix and the other is further reading.

Cards open independently, so you can compare two side by side.

Acting on one

Where Insights carries the fix out, the action button opens a confirm step that names the concrete change rather than asking whether you are sure: the four tools going into the skill. Under those rows sits the caveat, before you commit rather than after. If it changes your mind, cancel and the finding stays open.

Resolved. Confirming resolves the finding. The card collapses to a green state whose headline states what changed, with Undo to reopen it.

Dismissed. You never have to accept a finding. Dismiss when it is real but not something you intend to act on: the idle Connector is a pilot that has not started, the batch job is meant to run that way. The card collapses to one line, with Undo, and the finding does not come back.

The period selector at the top of the page offers Last 7 days, Last 30 days, and Last 90 days, and starts on 30. It changes how far back the page reads, not how far back an analysis looked: each one has its own fixed window, stated on the finding, for the reasons above.

Next

Debug an individual call behind any of these numbers with Tool calls.