Why we built hello.david

The most useful thing our software did last year was refuse to finish a sentence.

A datasheet draft for an industrial client. The model had written three clean, cited paragraphs: capacity, enclosure rating, operating range, every number traceable to a spec document. Then it started a fourth: “It is the most reliable system on the market.” And the gate held it. Unverifiable superlative. No source. Not shipping.

That refusal is the whole product. This post is about why.

The real problem isn’t writing. It’s trust.

I build positioning for industrial companies. The marketing teams I meet are three, maybe five people. The catalogs they carry run to thousands of SKUs. The product truth (the real numbers, the norm references, the conditions under which a spec actually holds) lives with engineering, in decades of PDFs, in revision-controlled documents marketing has never read.

Now hand that team an AI writing tool. It produces more copy in an afternoon than they shipped last quarter. Volume problem solved.

Except volume was never the problem. Every draft that comes back confident and unverifiable makes the real bottleneck worse. Someone still has to check every claim against documents they don’t have, in language engineering owns. The team didn’t get a teammate. They got a fluent intern who makes things up, plus a review burden they can’t staff.

In consumer marketing, a wrong claim is embarrassing. In industrial B2B, marketing claims end up in datasheets, in offers, in tender documents. A Leistungsverzeichnis is legally binding. Slop isn’t a taste issue here. It’s exposure.

The reframe: a teammate is defined by what it won’t do

We spent a year building marketing agent systems for industrial clients before David existed. Governed knowledge bases with confidence tags. Brand contracts with banned-phrase lists. Voice-checks that run before a single file is written. An approval gate where a human, not an agent, moves work from draft to shipped.

The lesson from that year was uncomfortable for the AI industry’s favorite story: generation was never the hard part. Any decent model writes fluent copy. The hard part was everything around it: knowing what’s true, staying in one voice, and proving both to the person who signs off.

Which flips the question you should ask of any AI tool. Not “what can it do?” They can all do everything, allegedly. Ask what it won’t do. Won’t invent a number. Won’t leave your brand voice. Won’t ship without your sign-off. That’s not a limitation. That’s the job description of a teammate you can actually delegate to.

Three things follow from taking that seriously.

The knowledge has to be yours. Generic models know everything and nothing. They’ve read the whole internet and none of your spec sheets. David works from a governed knowledge base built from your own documents, and every claim it makes carries a citation back to the file and page it came from. If it can’t cite it, it can’t say it.

Governance has to be a lock, not a setting. Every tool now has a “brand voice” slider. A slider you can ignore is decoration. David runs three locks that aren’t optional: a brand-lock (your voice, your claims policy, your banned phrases), a source-lock (every claim cited to your documents), and an approval gate (a human ships it, or nobody does). The held sentence at the top of this post is those locks doing their job.

You should be able to read the rules. If software is going to decide what your company can say, you’re entitled to see how. So David is open source. Apache-2.0, your own model keys, no seats, no lock-in. Open prompts, open governance logic, an approval trail you can audit. Governance you can read.

Why “David”

The big AI marketing platforms are impressive machines. They’re also closed, priced per seat, and optimized for exactly one thing: volume. More posts, more variants, more, more.

David is named for the fight it picks. Small, open, governed, and aimed at the one dimension the Goliaths ignore: whether you can stand behind what ships.

Five questions for your own team

If you run marketing in an industrial company, you don’t need my software to test the thesis. Ask these on Monday.

Can you trace every claim on your homepage to a document engineering would sign? Who approved the last datasheet that went out, and is that recorded anywhere? If your best product manager read your website, what would they correct first? How many versions of your flagship product’s spec are in circulation right now? And when your team uses AI today: who checks its numbers, and against what?

If those questions are uncomfortable, the problem isn’t your team. It’s that nobody built marketing software for companies where being right is part of the product.

We’re building it now, in the open. The teammate who read every manual, built to show its work.