My new favorite - a novel (for us) use of the now-old AI by Workato connector, built by one of my teammates. Whenever a recipe fails, we generate an integration log in a standard callable function. This function routes the issue to datadog, which further routes it downstream for handling. We've had that since forever, and over time experienced a lot of friction with business teams not being able to distinguish data issues that they need to own from engineering issues that we would own. So increasingly every job failure came back to my team, which at our scale (1-2% failure rate on ~300K jobs monthly) wasn't sustainable. The solution: two new Ai by Workato steps in our logging recipe.
Step 19 parses specific information out of the recipe job payload, including error details (necessary because data format varies from job to job)
Step 20 classifies the parsed error message into one of several buckets, which then guide the downstream routing of the issue.
Originally we had planned to do this using a Genie, and that approach also would work (and would be needed for any more sophisticated post-analysis actions) -- but this turned out to be much simpler and avoids other complications like getting LLM keys, token usage, etc.
Most recently, identified and built some fairly easy optimizations on high-consuming recipes that cut 1M credits/month from our baseline consumption. (But don't tell my AE 😉). We had legacy integrations with inefficient triggers, or that repetitively made the same updates even if they weren't needed.
Replacing naive realtime triggers with scheduled (often query-based) triggers often reduces credit consumption by 100x, 1000x or more especially when triggering on systems/objects that are frequently updated, like the Salesforce Account object
Another nice trick is to use SQL Transformations to analyze records in bulk, and spit out only updated records that need to be processed; sort of a custom change data capture (for when Data Pipeline isn't the right solution)
Yet another version of this is using lookup tables or data tables to cache key-value pairs consisting of a key and an md5() hash only of the fields you care about on records with frequent updates posted to other systems. When you get an update trigger, hash the updated values and compare with the cache; if no change, stop processing.
The ultimate point of this is not to be a skinflint, but to get baseline costs down to give us more runway for growth, especially on AI features. Getting more budget for integration costs is always an uphill battle, so showing the ROI of efficient integrations is always a good way to win hearts and minds when you really need that budget 😃
This automates failure triage that used to take an engineer 15-30+ minutes to dig into per issue. Now it's instant and anybody can follow up to resolve.
Aaron W. Yes, we have something similar but in a different domain. It allows members of specific slack channels to receive alerts about customer orders. If an order fails, a trigger recipe automatically makes a headless Genie call to diagnose the failure. Users can then interact with the Genie subsequently to review details.
All out of the box connectors, except for a custom connection via HTTP to access our order system
Genie with minimal direction and 3 skills + KB:
Trigger monitoring platform events from Salesforce
A skill that pulls datadog logs for the order
A skill that pulls order status from the order system
And a knowledge base consisting of a single spreadsheet listing known failure modes and owners per type of failure
Had more skills planned but they weren't needed to accurately diagnose and recommend resolution. We are planning to add more skills later for things like automatically creating tickets for bugs, connecting to incident.io, etc.
That's why "not a drop to drink" lol...
Oh, and although I don't really play video games, I love thinking about designing one...
