Build, test, and run a pipeline

You build a pipeline by describing it. Start a chat in Rippit, or ask your own AI assistant through the Rippit MCP server, and say what you want to keep happening. You never have to say “pipeline”: “every Monday,” “each new ticket,” “keep monitoring,” and “alert me when” are enough.

​

Describe what you want

Think of it as briefing someone on a standing job rather than asking a question. A chat answer is for right now; a pipeline is the version that still happens every Friday when nobody remembers to ask. The one habit worth forming: describe the whole job, not just the interesting part. Most people write the analysis (“summarize billing complaints”) and stop, because that is the part they care about. Rippit builds every step from your description, so anything you leave out is something it has to ask you about. A good description names four things:

• Which data it reads. The source, and which rows matter.
• What judgment you want. Not a filter over metadata, but the reading you would ask a person for: what the theme is, why it happened, whether it is getting worse.
• When it runs. “On every new conversation,” “every Monday at 9am Eastern,” or both. For a schedule, say how far back each run should look.
• Where the result goes. A table you will query, a Slack channel, an email list, a ticket in another system.

A few things that help:

• Say whether you want one result or one per group. “A digest for the team” and “a brief for each account owner” are different pipelines.
• Say what should happen when nothing matches. Otherwise you get silence on a quiet week and cannot tell it apart from a broken pipeline.
• Expect questions. Rippit asks about what the data can’t tell it, such as the recipients, the send time, and the timezone.
• Prove it in chat first. If you are unsure the analysis works at all, ask it as a normal question, then say “do this every week.”

See Pipeline examples for prompts to start from.

​

What a well-built pipeline looks like

Rippit makes these choices when it builds a pipeline, and they are worth knowing when you review one. The principle is to put each piece of work where it is exact and cheap. Which step to use maps each kind of work to its step.

• Counting and filtering happen in a rollup, never in a model or an agent. A rollup reads every row in its window and its totals are exact.
• A judgment every record needs is an AI column on the source. It runs once per record, and the rollup then filters and groups by it.
• Agents write and synthesize over a group. Each row costs one agent session, and whatever an agent re-reads with its tools it pays for on every run, so it is handed the columns it needs.
• The rollup’s rows are the unit you act on. One row per recipient, account, or topic means one message each.

A few more habits make a pipeline dependable:

• Judge once, reuse everywhere. AI columns on a source are available to every other pipeline and to chat, so a second pipeline rarely needs its own.
• Bound the first run. Going live processes every existing row that passes the row filter. Filter to a date range when you only need recent history.
• Match the lookback to the schedule. A weekly rollup looks back seven days, so each row covers its own period and nothing is reported twice.
• Keep alerts quiet. Use a threshold so a group writes a row only when it qualifies, and decide what should happen when nothing qualifies.
• Name people and accounts by name. If a row’s key is an ID, look up a name, or the output shows the ID.
• Send only to recipients you trust. A recipient should come from a verified column or a value you approved, not from text a model produced.

When not to use a pipeline: for an answer you want once, ask in chat. A pipeline earns its setup when the same work should happen again.

​

Review the draft

Rippit explains the chain of steps in plain words, then builds it as a draft. In Rippit chat, the pipeline appears beside the conversation as a graph: each card is a table or a step, and you can open any card to see how it is set up. A draft does nothing on its own. No rows are processed and nothing is sent until you put it live. If a step has a problem, such as a column that doesn’t exist, the draft says which step and why, and Rippit fixes it before going further. To change the draft, say what you want: “look back 30 days instead of 7,” “add a column for sentiment,” “send it to the support channel too.”

​

Test it

A test runs every step once, on a sample of your data, and writes the results to a test worksheet, a separate table you can read without touching the live tables. Testing is the step worth your patience, because a schedule turns a small mistake into a recurring one.

• A pipeline that only enriches samples 200 rows by default. Ask for more or fewer.
• When a pipeline has a rollup, the test runs the rollup once, then runs the steps after it on a sample of the rollup’s rows: up to 50 by default when one of them is an agent run.
• When a pipeline sends, Rippit asks before testing it and samples a single row unless you ask for more, so one message goes out.
• A test runs right away, whatever the schedule and delivery times say.
• You can test on a narrower slice, such as last week’s billing conversations, without changing the pipeline.

Read the test worksheet the way you would read the real output: are the labels right, is the summary useful, did the message say what you wanted? When something is off, say what, and Rippit changes the step and tests again.

​

Check the cost

After a test, ask what the pipeline will cost. Rippit estimates two numbers from the test run: the one-time cost to process the rows you already have, and the ongoing cost per day as new rows arrive. The estimate counts the model calls that fill AI columns. It does not count agent runs or search provider costs, so a pipeline that uses them costs more than the estimate shows. To gauge those, run a test on a few rows and check the credits it used. If the estimate is higher than you want, the usual fixes are a tighter row filter, a cheaper model for simple columns, or moving agent runs after a rollup so they run on a few rows instead of every record.

​

Go live

Tell Rippit to put the pipeline live. It summarizes what will run, when, and where it will send, and starts the pipeline only after you confirm. A pipeline with unresolved problems cannot go live. Once it is live:

• Enrichment starts. AI columns, formulas, and lookups fill in for existing rows that match the row filter, then for each new row as it arrives.
• Schedules take effect. Rollups run at their next scheduled time. Rippit starts a scheduled run ahead of the delivery time so the results are ready, then holds each message until the time you asked for.
• Output lands in real tables. You can query them, chart them, ask about them in chat, and use them as the source for another pipeline.

​

Change a live pipeline

Ask for the change in the pipeline’s chat, or from your assistant. Changes to a live pipeline are drafts: the running version keeps running, unchanged, until you say to make them live. That holds for removing a step too. Most things can change at any time, including instructions, filters, schedules, recipients, and adding new steps. Two things to know:

• A column added later starts from now. It fills rows that arrive or change after it goes live. Rows processed earlier stay empty for that column.
• A live rollup keeps its shape. Its columns and grouping are fixed, so a rollup that needs a new column is added as a new step.

​

Pause and resume

Pausing stops new runs right away; a run already in progress finishes. Pause the whole pipeline, or only the steps that send, to keep enrichment running while messages stop. A paused step stays paused through later changes until you resume it. A pipeline can’t be deleted yet. To retire one, pause it.

​

Keep an eye on it

Open Pipelines in the left sidebar. Each pipeline shows a status: A pipeline with only some steps paused, or with draft changes waiting, still shows Active. Open it to see each step’s state. Open a pipeline to see its graph. Click a table to see its rows. AI column and rollup cards switch between Output, the rows they produced, and Config, their instructions, model, and schedule. An action card shows when it sends, the rows it sends from, and the error from its last run if that run failed. The page is for inspecting, not editing. To change, test, pause, or fix a pipeline, click Ask Rippit. It opens the most recent chat that changed the pipeline, or a new chat when there is none, such as a pipeline built from your own assistant, with the pipeline open beside the conversation.

​

Build from your AI assistant

Everything on this page works from any MCP client. Ask your assistant to “use Rippit to” build the pipeline, and it drafts, tests, estimates, and manages it through the pipeline tools. The same safeguards apply: your assistant asks before testing a pipeline that sends, and puts one live only when you ask. A pipeline built from your assistant shows up on the Pipelines page in Rippit like any other.