Run the same prompt, the same sequence and the same sending infrastructure against two lists — one enriched properly, one bought as-is — and the results will differ by more than any model upgrade has ever delivered. We have measured this on client programmes repeatedly and it is not close.
This is unintuitive, because the model is the visible part. It is also the part vendors compete on. But an AI writing a personalised opener is only as good as the facts it is given, and most teams are giving it a job title, a company name and hope.
What "enriched" actually has to mean
Enrichment is not appending an email address. For outreach that survives contact with a real buyer, an enriched record answers four questions:
- Who is this company, currently? Not the description from three years ago. What they sell, to whom, at what size, on which stack.
- What changed recently? Hiring, funding, leadership, product launches, tooling. Change is what creates a reason to talk this month.
- Who is this person inside it? Scope and seniority, not just a title string. A "Head of Growth" at forty people and at four thousand are different jobs.
- How confident are we in each of those? A field with no confidence attached is a field you cannot safely put in an email.
The confidence point is the one people skip
The failure mode of AI outreach is not blandness. It is confident wrongness — an opener that references a product the company sunset last year, or congratulates someone on a role they left. One of those lands in an inbox and the account is gone, permanently, along with some of your reputation.
The fix is boring: carry a confidence score on every field, and let the generation step degrade gracefully. High confidence gets the specific opener. Low confidence gets a good generic one. Nothing gets a hallucinated one.
The failure mode of AI outreach is not blandness. It is confident wrongness.
Coverage is a distribution, not a number
Vendors quote coverage as a single percentage. That number is nearly useless because it is an average across a database you are not going to use all of. What matters is coverage within your ICP.
We routinely find providers advertising strong overall coverage that fall to a fraction of it on, say, sub-200-headcount companies in a specific region — exactly the segment the client cares about. Measure coverage on a sample of your own target accounts before you sign anything. It takes an afternoon and it has changed vendor decisions on every audit we have run.
The waterfall beats the single source
No single provider wins everywhere. The economics work out strongly in favour of a waterfall: try the cheapest source that is good at this field for this segment, fall back to the next, stop when a value clears your confidence threshold, and only pay for resolved rows.
This is more engineering than buying one subscription, which is why most teams do not do it. It is also usually a large cost reduction and a step change in quality at the same time, which is why the teams that do it rarely go back.
Enrich continuously, not once
The last mistake is treating enrichment as an event. Records decay — roughly a quarter of contact data goes stale within a year, faster in high-growth segments. If your engine enriches at import and never again, its accuracy is highest on the day you stop paying attention to it.
Re-resolve on a schedule, and re-resolve on trigger when a signal fires. Then the AI on top has something true to work with, which turns out to be the whole game.