A scheduled automation hands the Enterpret Agent a standing job and a cadence. Write what to analyze, set how often it runs, choose what it does with the result, and the agent does the work in the background and delivers on its own. The result is either a report sent where your team works, in Slack or email, or an action taken in a connected tool, like a filed ticket or an updated record.

This article covers the full lifecycle: creating an automation, starting from a ready-made template, reading what it delivers, managing it after it is live, and troubleshooting when something looks wrong.

Everything the Enterpret Agent does in a live Session, it can do here on a cadence, with no one watching.

Building an automation is six steps. The first four are setup, the test run shows the output before you commit, and saving activates the schedule.

Worked example. A product manager owns the mobile onboarding area and wants a weekly digest of what customers are saying: what changed, which themes are emerging or persistent, representative quotes, delivered to the team's Slack channel every Monday morning.

Open the automation builder and write a prompt describing the job, the same way you would brief the agent in a Session. State the scope, the window, what to compare against, and the shape of the output you want.

Example prompt:

Generate a weekly Voice of Customer digest for the mobile onboarding product area. Cover the past 7 days against the prior 7 days, and check the past 4 weeks for persistent themes. For each significant theme, give the volume count and week-over-week delta, classify it as emerging, trending, or persistent, and include two or three verbatim quotes with source links. Lead with what changed, not what exists. Prefer tables over prose.

A sharper prompt produces a sharper report. If the prompt is rough, use the refine action to sharpen it. The agent rewrites the draft into a fuller brief, tightening the scope, naming the comparison window, and structuring the output. Keep or adjust the result before moving on.

The agent runs against your Context Graph, so every count and quote in the result links back to the real feedback record. You do not need to tell it where the data lives.

Choose how often the automation runs and when. A schedule has a frequency, a day, a time, and a timezone.

For the worked example: weekly, Monday, 8:00 AM, set to the team's working timezone.

The scheduled time is when the run starts. The result arrives once the agent finishes the work, fanning out across your feedback, synthesizing themes, pulling and verifying quotes, assembling the output. Allow about 30 minutes between the scheduled start and delivery, more for a broad or deep job. Set the time earlier than you need the result in hand.

A scheduled automation either delivers a report or takes an action in a connected tool.

For the worked example: the team's Slack channel, so the Monday digest is waiting when the team logs on. The recurring question nobody wanted to keep asking now answers itself.

By default the automation works over your Context Graph and the Enterpret Agent's built-in analytical skills. You can extend it.

Connectors authenticate per user, under your own account. When the agent takes an action in a connected tool on a scheduled run, that action is attributable to you and scoped to your access.

Before you save, run a test to preview the output. The agent executes the job once, on demand, and shows what a real run produces.

Read the test output as the result's first reader. Check that the scope is right, the themes are the ones you care about, the quotes are verbatim and linked, and the format is the one you asked for. If something is off, return to the prompt, adjust, and test again. The test run is where you tune the prompt, before the automation goes live, because editing a live automation means waiting for the next scheduled run to see the change land.

Save the automation. From here it runs on the cadence you set and delivers to the destination you chose, with no further input.

Rather than writing a prompt from scratch, start from a ready-made pattern. A template loads its prompt into the automation builder, already shaped for the job: scope, comparison window, output structure. Adapt the scope to yours, set the schedule and what it does with the result, test, and save.

Ticket triage and CRM account flags are action-in-connector patterns: instead of delivering a report, they write a result back into Linear, Jira, or your CRM. They need the relevant connector attached and authorized.

A pattern that does not quite fit is still worth running. The catalog is a starting point, and the agent will run any objective you can brief it on, not only the patterns listed here.

An automation produces one of two things when its run completes: a report, or an action in a connected tool.

Most automations deliver a report. It opens with an executive snapshot you can read in under a minute, carries a deeper research section underneath, and links every count and quote back to the real feedback record behind it. It is the same kind of output the Enterpret Agent produces in a Session, delivered on its own when the run finishes rather than in your chat window.

The executive snapshot sits at the top and answers one question: what changed since last time. It carries a headline stating the most important movement in the window, the key changes with volume counts and the delta against the prior period, and representative quotes for the findings that matter. It leads with movement, not inventory, classifying themes as emerging, trending, or persistent so a theme that has been loud for months reads differently from one that spiked this week.

The deep dive below the snapshot is the research section, where the agent shows its work on the findings that warrant it: hypotheses about what is driving a change, the causal chains it traced through the feedback, and a confidence level on each conclusion. A busy reader can stop after the snapshot, a curious one keeps going.

Citations make a report verifiable rather than a summary you have to trust. Every aggregate number is a citation, and so is every quote. Click the count or quote behind a claim you want to check, and Enterpret opens the underlying feedback records in your dashboard, filtered to exactly what drove that number, so you can confirm the finding, read the full context, or pull more examples. When no validated quote exists for a finding, the report flags it as missing rather than filling the gap.

A report can also be produced as a PDF, cited and ready to share with someone who was not in the channel. The PDF carries the same snapshot-then-deep-dive structure and the same citation links, so a finding in the file is as verifiable as the same finding read in chat.

When the automation's job is to act rather than report, the result lands in the connected tool instead. It files or tags a ticket in your tracker, updates a record in your CRM, or posts into the tool where that work already lives. The action runs under the connector's access, the same as in a Session, and is attributable to the person who created the automation.

Every automation you create lives in one place: the Automations list. From there you find an automation by its name and owner, edit it, pause or resume it, delete it, and review its run history.

The list groups automations by state:

Each row shows the automation's name and its owner, the person who created it. In a workspace where several people run automations, the list holds everyone's; filter to your own to see just the ones you created.

Edit an automation from the list to rewrite the prompt, change the cadence, switch what it does with the result, or adjust the connectors and skills it uses. Edits take effect on the next scheduled run and apply going forward, not retroactively, which is why the test stage is the place to tune the prompt.

Pause an automation to stop it running without losing its configuration, for a digest that is noisy during a launch week or a monitor for a roster being reworked. Delete it to remove it permanently. A deleted automation stops running and leaves the list.

Review run history to see, per automation, when it ran, whether the run completed, and the result it produced. Run history is the first place to look when a delivery looks wrong: a run that started but did not deliver shows up here, before you re-check the prompt or the destination.

Editing is reserved for the person who created the automation. An admin can pause, resume, or delete any automation in the workspace, but cannot edit one they did not create. Each automation shows its owner.

Most automation problems trace back to one of four things: a run that errored, a result that never reached its destination, a connector that lost access, or a prompt that asked for more than one run could carry.

Two checks resolve a large share of issues first:

A run failed or errored. A run can start and then stop before it produces a result; the failure is recorded in the automation's run history. A single errored run on an otherwise healthy automation is usually a transient model or service error, so run a test to confirm the automation is healthy now. Repeated failures usually mean the job is trying to do too much in one pass, or a connector it depends on lost access. Narrow the prompt and confirm connector access, then test again. If runs keep failing after that, contact support.

A report did not arrive. The automation ran, but no report showed up where you expected. Confirm the delivery destination on the automation is the channel or person you expect, since a report sent to the wrong place looks like one that never arrived. For delivery to a Slack channel, confirm the Enterpret Slack app is present in that channel, because delivery to a channel the app has not been added to will not land. Add the app, then test.

A connector lost access or authorization. An automation that reads from or writes to a connected tool depends on that connector staying authorized. OAuth sessions can lapse, an admin in the connected tool can revoke access, or your own access in the tool can narrow. Open the Connectors panel, check the status of the connector the automation uses, reconnect or re-authorize it by signing in again with your own account, and confirm your account still has the access the automation needs, for example membership in the target Slack channel or visibility into the Linear project. Then test.

A prompt was too large or the output was truncated. A run can complete but return less than expected: a report that stops partway, a table that cuts off, an analysis covering a narrower slice than the prompt asked for. This usually means the job asked for more than a single run could carry. Narrow the scope to one product area, one account roster, or a single window rather than the whole corpus at once, and bound the output by asking for representative quotes rather than all of them. If wide coverage is needed, split it into several automations, each scoped to a slice. Tune at the test stage until a single run returns the complete report.

How to escalate. Most cases above resolve yourself. Contact support when the problem outlasts those fixes: runs that keep failing after a narrowed prompt and confirmed connector access, a report that still truncates despite a focused prompt, or anything you cannot diagnose from the run history. Email [email protected] with the automation name, what you expected versus what happened, the run time in question, and the delivery destination or connector involved. The more precisely you can point at the failing run, the faster the team can trace it.

Have a question about scheduled automations, or one you would love to run? Reach out at [email protected] and we will jump in.