Custom Pipeline Use Cases

Overview

A Custom Pipeline is worth building when a process is too specific for a standard product feature, but too recurring to keep doing by hand. This page catalogs the workflow families teams actually automate, the shape each one takes, and what to include in your prompt so the AI Agent builds the right thing on the first attempt.

Every one of these starts the same way: you describe the workflow in chat, the Agent builds and configures it, you review a first run, and then you schedule it.

Choosing the right tool first

Not everything needs a pipeline. Three capabilities overlap, and picking the wrong one is the most common reason a project takes longer than it should:

The three compose. The usual path to a production pipeline is: prototype the platform request in chat with Discovery API, confirm it returns what you expect, then ask the Agent to turn that into a pipeline. The pipeline uses the same connections and the same request shape, so a request that worked in chat works in the pipeline.

1. Audience sync and suppression

Keeping ad-platform audiences in step with first-party data in your warehouse or CRM.

• Retargeting sync — push a segment (high-LTV customers, cart abandoners, trial users) into Meta, Google Ads, TikTok or LinkedIn as a custom audience, and keep it refreshed.
• Suppression lists — the same mechanics inverted: push existing customers and recent converters as an exclusion so you stop paying to re-acquire them. Worth refreshing more often than your targeting audiences — a lagging suppression list costs money every day it is stale.
• Lookalike seeding — a synced audience becomes the seed for a lookalike.
• Audience inventory — a read-only report of which audiences exist, how big they are, and which ones stopped refreshing.

Example prompt: “Build a weekly pipeline that reads high-LTV customers from Improvado Storage and pushes audience updates to Meta Ads and Google Ads. Emails are in email, phones in phone_e164. Name the audiences with the segment and the run date.”

2. Campaign operations

Writing back to ad platforms — the workflows that used to require either a spreadsheet and a person, or an engineer.

• Campaign launch from a brief — read a brief from Google Sheets, validate naming and UTMs against your convention, then create the campaign chain (budget → campaign → ad set → creative → ad) paused, and report which rows succeeded with links.
• Governance renames at scale — your naming-convention rules flag offenders, the pipeline renames them in the platform.
• Budget and status control — compare spend against plan and pause the overspenders, or cut a budget, with a before/after notification.

Example prompt: “Create a campaign launch validation pipeline that reads a Google Sheet with campaign briefs, checks naming conventions and landing-page tracking, then returns a report with pass/fail results. Don't create anything yet — validation only.”

3. Data-quality watchdogs

Scheduled checks over your marketing tables: freshness, missing rows, zero-spend anomalies, broken naming. The pipeline publishes a summary of the failed checks as an artifact you can download or have mailed out.

If what you want is a threshold alert on a metric rather than a table-level check, look at Discovery API rules first — they give you the same monitoring through Marketing Data Governance, build the underlying pipeline for you, and land in your digest email.

Example prompt: “Create a daily data-quality pipeline that checks yesterday's Meta Ads and Google Ads spend tables for missing rows, stale data and zero-spend anomalies. Send a summary artifact listing the failed checks.”

4. Reverse ETL into business tools

Pushing marketing metrics back out of Improvado into the systems where the rest of the company works — attribution or lifecycle metrics onto HubSpot or Salesforce records, campaign status into a Notion or Google Sheets tracker, alerts into Slack or Google Chat.

These are the pipelines most likely to replace an existing no-code automation. Two things to say in the prompt: which field is the match key, and what should happen when the value already matches (skip it — writing unchanged records burns quota and pollutes record history).

5. Long-running backfills

Reprocessing months of history in resumable batches. This is the case where a pipeline earns its keep over a script: it checkpoints as it goes, so a restart mid-way resumes from the last completed batch instead of starting over.

Run the first pass on a small slice — one day, one account — and review the output before going full-volume.

Example prompt: “Backfill 12 months of Meta Ads spend by day into our warehouse, in daily batches, resumable if it fails partway. Start with just September 2025 so I can check it.”

6. Recurring reports and files

Generating a file on a schedule — a client-ready export, a reconciliation sheet, a QA dump — and publishing it as a run artifact you can download from Run History.

What to include in your prompt

The Agent builds a better pipeline the first time when your description covers these five things:

1. 1Input — which data source, destination or table, and which accounts.
2. 2Business rules — the filters, thresholds and decisions, in your own words.
3. 3Output — a table, a file, a platform change, a notification, or a combination.
4. 4Cadence — on demand, or a schedule (and in which timezone you mean it).
5. 5Blast radius — whether the first run should be a dry run, and whether anything it creates should start paused.

You don't need to specify endpoints, API versions or code. If the Agent needs one of those, it resolves it from Improvado's own curated API guides for that source.

Related articles