How Recipes Keeps a Recurring Analysis Consistent
Next month’s pipeline review should explain what changed in the business. Asking an AI assistant to work out the analysis again introduces another question: did the business change, or did the way we measured it?
A Recipe saves the sources, selection rules, and calculations you’ve agreed on. Run those steps against fresh data for the next review. When the question changes, review the proposed edits to the analysis.
Saved Analysis
An LLM helps write and revise the plan. Supported analytical transformations run through deterministic operations:
- Author: define the question, select the tools, and state which records count.
- Compile: translate the transformation into commands the engine can execute, checking columns and parameters.
- Verify: run the candidate against data and inspect the results and warnings.
- Pin: save the command for reuse, with the step’s wording and tool forming part of its identity.
- Replay: use the machine-checkable plan (the contract) or saved command on the next run.
Without a usable contract or command, that stored analytical step stops for repair. Verification covers the tested inputs; it cannot prove correctness for every future dataset.
Calculation Review
An assistant can write a calculation that runs successfully while counting the same opportunity more than once. This synthetic example shows why the analytical steps need review.
Suppose the question is how much pipeline has had sales activity. Two opportunities total $300,000: A is worth $100,000 and has three calls; B is worth $200,000 and has one. Joining opportunities to calls produces these rows:
| Opportunity | Activity | Amount |
|---|---|---|
| A | Call 1 | $100,000 |
| A | Call 2 | $100,000 |
| A | Call 3 | $100,000 |
| B | Call 4 | $200,000 |
Adding the amounts gives $500,000. An extra $200,000 of pipeline could change the discussion about coverage, even though no new opportunity exists.
We ran this synthetic example through QFlow’s merge executor. It warns that the join multiplied rows and that downstream sums will be inflated. The join can still complete because one-to-many relationships are often intentional.
The corrected plan reduces matching activities to one opportunity ID each, joins those IDs to the original opportunities, and sums the two amounts. We ran that sequence through the grouping and merge executors: $300,000.
Both joins executed successfully. Reviewing what each row represented made the difference. Saving the corrected plan carries that counting rule into the next review, using fresh data.
Recipe Review
When the business question changes, proposed edits can be compared with the saved steps before application. The output and warnings provide evidence for that review, and unfinished verification stays visible.
The underlying data and business definition still need scrutiny. A Recipe can faithfully repeat a calculation based on incomplete records or the wrong population.
The model can also write the report’s explanation. Correctness review checks claims against recorded results, but interpretation still needs judgment. Even the correct $300,000 says little about whether the pipeline supports the plan until you consider its age, stages, and expected close dates.