Why we built an AI for dividends, not stocks

Ask a general-purpose finance AI what a stock's dividend will be next quarter and it will answer confidently. In our testing, that confident answer was wrong roughly 30% of the time. For an income investor, a wrong dividend number isn't a rounding error — it's the whole point. So we built something narrower, and far more careful.

The 30% problem

Big language models are trained on the whole internet, which means they're trained on stale, contradictory, and occasionally fabricated dividend data. They confuse the declared rate with the trailing rate. They miss special dividends, spin-offs, and changes in payment frequency. And when they don't know, they guess — fluently. That fluency is the dangerous part: a wrong number that sounds right.

The dataset

To measure it honestly, we built an evaluation set of real dividend questions with known, verified answers — forward rates, ex-dates, yields, frequency, and payout history across hundreds of names, including the messy edge cases that trip everyone up. Crucially, the answers come from our own reconciled data lake, timestamped so we know exactly what was true and when.

The eval

We ran general-purpose models against that set and scored them on accuracy, not vibes. The pattern was consistent: strong on the household names, shaky on everything else, and rarely willing to say "I don't know." The errors clustered exactly where income investors get hurt — frequency changes, recent cuts, and funds that return capital.

A model that's 70% right about dividends isn't 70% useful. It's a trap, because you can't tell which 70%.

What we did instead

Our AI doesn't recall dividend facts from training — it looks them up. Every dividend number it gives you is pulled live from a reconciled database, not generated from memory. The model's job is to reason, plan, and explain; the data's job is to be correct. When the data is uncertain, the answer says so instead of inventing a figure.

• Retrieval, not recall — numbers come from the data lake, every time.
• Reconciled sources — disagreements between providers are flagged, not averaged into a wrong middle.
• Honest uncertainty — "I don't have a confident figure" beats a confident wrong one.

Why it matters

Income investing rewards being right about boring details: when the check arrives, how big it is, and whether it's safe. A general AI optimizes for sounding helpful across everything. We optimized for being correct about one thing that actually pays your bills. That's the whole reason HeyDividend exists.