beetl

Should I turn my heating on? I asked Germany's gas storage data

2026-11-10T00:00:00+00:00

It's November in Berlin and 1°C when I get up, and I still haven't turned the heating on. Partly to save money. Partly because I'd read that Germany's gas storage is low this year, and some stubborn part of me figured every radiator counts.

What we learned putting decision models inside a production AI agent

2026-10-07T00:00:00+00:00

Most of the AI in Beetl today is a large language model in a loop. Our assistant reads the question, calls tools such as list_data_sets and execute_query, and writes an answer. Automations do the same on a schedule: they check a condition in the data and write a report that ends in a finding, either "clear" or "needs attention".

Many of the decisions inside that loop are small: is this dataset relevant? Does this report actually show what it claims? Did anything change since last run? Each one costs a full language-model call today, or doesn't happen at all.

Decision models are a new kind of model built for exactly those questions. We spent two days wiring one into Beetl and measuring it with our own evaluation suite. Then OpenAI released its own, and we ran the same tests on it. This post is what we found.

How we wrote 20 integrations in one day

2026-09-14T00:00:00+00:00

We wanted Beetl users to connect their business applications and work with the data together, such as comparing sales in HubSpot with invoices from their accounting system.

First, we need to get that data into Beetl. This is why we built Beetl Connect, a toolkit that keeps the code for fetching data small and lets our platform handle running it.

Connect lets us build ready-to-use integrations for our catalogue and quickly create custom ones for individual customers. That could mean connecting an internal application or handling a customer's specific data needs. Both use the same toolkit and run on the same infrastructure.

Using Connect and a coding agent, we wrote and initially tested twenty integrations in one day. HubSpot gives us a good example of how we did it.

Fivetran charges you to delete data

2026-08-26T00:00:00+00:00

Since 1 January 2026 deleted rows count toward paid monthly active rows. That change, and what actually triggers a full table re-sync, are both documented by Fivetran.

AI Automations for the messy half of data quality

2026-08-18T00:00:00+00:00

I had been thinking about this feature for a while because we had already built most of the pieces it needed: pipelines, a query engine, scheduling, run history, and an AI assistant that could inspect data. I wanted to combine those pieces into a check that runs every day and tells me whether the data is still correct.

AI Automations are scheduled or manually run data checks in Beetl. You describe a business condition in plain language, and Beetl investigates the current data with read-only SQL before reporting Healthy, Needs attention, or Failed.

Of course, you can already do this deterministically. Write a query, assert that it returns zero rows, and send an alert when it does not. For a well-defined invariant, that is often exactly what you should do.

A deterministic check only covers the conditions encoded in its query. As exceptions and new source values appear, the query has to be updated or its results become increasingly noisy.

I wanted something more "agentic". By that I mean a check that can choose investigative steps beyond the path I explicitly defined. I wanted to describe the outcome I cared about, then let the system inspect the schema, sample the data, write queries, and follow up on whatever it found. The queries still had to be read-only, and I needed to see the evidence behind the result.

You already have BI. What you're missing is trustworthy joined data.

2026-08-11T00:00:00+00:00

If you have a BI tool and still cannot trust your reporting, the tool is fine. The work that is missing sits upstream of the chart.

The mapping your ERP vendor never documented

2026-08-04T00:00:00+00:00

In an ERP migration the export is easy. The hard part is the undocumented business logic the old system encoded over years, and recovering it before go-live.

Why data-driven decisions are hard to make routine

2026-08-01T00:00:00+00:00

Most companies already have the data needed to improve an important decision. Those records sit in several operational systems and need to become a result that arrives on time, can be checked, and continues to work after the first analysis. Rebuilding that path for every question is especially difficult for a company without a dedicated data team.

I have come to think that this repeated integration work is the main obstacle to making data-driven decisions routine. It is also why we are building Beetl: a company should be able to begin with one recurring decision and reuse the connections, mappings, deployment, and monitoring for the next one. Later AI systems rely on the same connections and definitions.

The Excel-and-Power-BI handoff that holds post-merger reporting together

2026-05-27T00:00:00+00:00

For months or years after a manufacturing acquisition, the Excel workbook is the consolidation engine. The Power BI card shows the resolved number but never the formula or the comment beside it. That is what fails at year-end audit.

Can you see per-order margin while the shift is still running?

2026-05-23T00:00:00+00:00

The actual profit margin of a manufacturing order in most cases is decided based on what happens during the shift, but almost no plant can see it while the shift is running.

Anatomy of a 5-4 Champions League Thriller: A Football Data Case Study

2026-05-15T00:00:00+00:00

Champions League knockout games aren't usually 5-4 thrillers. April's PSG vs Bayern semi-final first leg was. Bayern had 57% possession, an xG of 3.06 to PSG's 1.90, and six big chances to PSG's two. PSG won anyway. I dug into the data to see what might be going on.

DACH manufacturing now trails every major economy on digitalization

2026-04-30T00:00:00+00:00

Manufacturing companies in the DACH Mittelstand are sitting on a great deal of valuable data, and almost none of them are using it well. I've spent the last few months talking to plant managers, controllers, and owners across the region, most recently at Hannover Messe last week, and the pattern is consistent. The data exists. They know it exists. They can't really use it.