Building dolv forced us to answer a question

I used to wave away: should the AI suggest the next action, or just do it?

We use the phrase 'grounded AI' in dolv and I want to explain what we mean by it, because it gets used loosely.

What it means in dolv:

When we first started testing dolv with early users, we did not have an explicit approval workflow.

The assumption was that teams would configure the level of automation they wanted. Some would run fully autonomous. Some would check in. Trust levels would vary.

One question we get from builders: what is the actual stack underneath dolv?

The short version: the frontend runs on Val Town. The marketing site runs on Astro. The integrations cover Google Workspace and Microsoft 365 natively with single OAuth consent screens.

Every early version of dolv was read-only. Pull the data, surface the insight, recommend the action.

The feedback was consistent: useful but incomplete.

When we built dolv's funnel health scoring, the obvious first version was equal weights. One-third each.

Equal weights broke immediately. A healthy TOFU was inflating the composite score even when MOFU and BOFU were in poor shape. The number felt fine. The pipeline was not.

Early on, we thought GTM teams needed better AI. Better outputs, better recommendations, smarter suggestions.

Then a head of growth said something that stuck:

Early in building dolv, we ran an internal test without the approval workflow. AI drafted, AI sent. No human in the loop.

It did not go catastrophically wrong. But one email went to a segment it was not intended for. The content was fine. The targeting was not.

Sharing this because I would have wanted someone to tell us earlier.

Mistake 1: We assumed teams wanted maximum autonomy.

Not the "we estimate roughly X hours" version. The real measurement.

I started tracking this for our own team a few months ago. The categories that showed up most:

- Re-keying the same information between the CRM and the content tool