Analytics: from spreadsheet to decision

1. What a pipeline is, in Flowcall terms

A pipeline is the work you repeat every week, written down once: get the data, shape it, check it, decide, deliver. Each of those is a task on the canvas; a link between two tasks means "run this after that". Tasks with no link between them run at the same time.

Two rules make everything else simple. First, every run gets its own folder; tasks pass data to each other by writing files there, and you can open any of them later from the Runs tab. Second, a task that fails after its retries stops everything downstream, so a bad file never becomes a bad decision.

Do this: open the sample pipeline, press Run now, and watch the Runs tab fill in.

Example used: Your first pipeline

2. Replace a manual report

Most recurring reports are: run a query, eyeball it, paste it into Excel, email it. Flowcall does the same four things without you: an SQL task runs the query (against a connection you set up once under Settings → Connections), a quality check does the eyeballing with rules that never get tired, and a Move-a-file task drops the result in Files or on a shared drive as Excel, date-stamped.

The habit to build: write the quality rule before you trust the query. rows >= 1 catches an empty extract; range: activation_rate 0 1 catches a broken calculation; anomaly: signups catches a feed that silently halved.

You never type file names between steps: when you link a quality check after the SQL task, it picks up the SQL task's output from a dropdown labelled with the step that produces it.

Do this: install the template, run it, and download the result as Excel from the Results tab.

Example used: Warehouse query to Excel report

3. Catch the 10,000-instead-of-1,000 mistake

The costliest data errors are the plausible ones: a user count ten times too large, a duplicate row that doubles a total, a blank that becomes zero. The quality-check task is a list of plain rules that the data must pass before anything downstream runs. unique: event_id finds duplicates. not_null: account_id finds blanks. range: users 1 500 finds the impossible value. anomaly: users compares today's total with the last runs and flags a three-sigma jump once it has five days of history.

When a rule fails, the run fails, the Results tab shows the offending column stats, and, if alerts are on, you get an email with the failing rule.

Do this: install the template and run it. It fails on purpose — find the duplicate and the blanks in the Results tab, then fix the rules so it passes.

Example used: Daily data quality gate

4. Route a decision to a person

Reports inform; pipelines in Flowcall can decide. An approval task pauses the run, puts the actual table in the inbox, and waits for someone to approve or reject with a comment. Give the step a dollar value so the decision log can total what was at stake.

The part nothing else does: an outcome check. Tell the step to run another pipeline N days later and judge the decision with a rule ("it worked if rows == 0 in churned_accounts.csv"). Over time the decision log shows a hit rate for the decisions your team approves, and each one has a downloadable receipt.

Do this: install the template, run it, approve the renewal offers from the inbox, then open the decision log.

Example used: Renewal review

5. Let the assistant do the typing

Turn on the assistant under Settings → Assistant: paste a free Hugging Face token, or use your own Anthropic key or a model on your own machine. Then describe what you want in the Assistant tab, in plain words: "every weekday at 7, email finance revenue by plan from sales.csv". It looks at your files and connections, drafts the whole pipeline, tests it in a scratch folder (no emails sent, nothing written), fixes what failed, and shows you one card with the steps and the test result. Nothing is saved or scheduled until you press Create.

For a small change to a pipeline you have open ("add a quality rule here"), it proposes just those steps, and you apply them with a button.

When a run fails, the Runs tab has "Ask the assistant why this failed": it reads the log, explains the cause and proposes the fix.

The example has a bug on purpose, so there is something real to ask about.

Do this: install the example, run it, press "Ask the assistant why this failed" and apply the fix. Then ask it to build a small pipeline from one of your own files.

Example used: Ask the assistant to fix it