A narrow problem worth automating
Most AI-in-marketing writing is either too abstract to implement or a product demo. This is neither — it's the setup I've now run with three clients, including what broke.
The problem: you have 300–800 target accounts and enough signal to rank them, but the signal is unstructured. Job postings, funding announcements, product launches, executive hires, technology changes, support tickets, call transcripts. A person could read all of it and produce a good ranking. Nobody has three weeks.
This is a genuinely good fit, because it's a classification and extraction task over text — the thing language models are actually reliable at — rather than a judgement task dressed up as one.
The architecture
Four layers, and the separation matters more than any individual component.
1. Structured signals — no model involved.
Firmographics, technology, product usage, CRM history. This is deterministic scoring in SQL. Don't route it through a model; you'll add cost, latency, and non-determinism to a calculation that already works.
2. Unstructured extraction — model does the reading.
For each account, feed the text you have — recent job postings, news, transcripts, support history — and extract to a fixed schema:
- Is there evidence of the trigger event we defined in our ICP work? (yes/no/unclear)
- Which of these five buying-committee roles have been hired recently?
- What problems are mentioned in their public communications?
- Any evidence of a competing solution in place?
The critical constraint: extraction only. You're asking what the text says, not what the model thinks about the account. That distinction is the whole design.
3. Deterministic scoring — no model involved.
Combine layers one and two with weights you set. The model contributed facts; the weighting of those facts is a business decision that must be inspectable, adjustable, and stable across runs.
This is where most implementations go wrong: they ask the model to produce the final score. Then nobody can explain why account 47 ranks above account 12, and the same lead scoring credibility collapse I've written about before happens on a faster timeline.
4. Human review of the top tier.
The top 30 get read by a person before anyone acts. Non-negotiable.
What actually improved
Coverage. Previously the team researched maybe 50 of 400 accounts properly. Now all 400 get the extraction pass and the person's time goes to the top 30. Same hours, eight times the coverage.
Consistency. Three people researching accounts apply three different standards. The extraction layer is uniform, which makes the ranking comparable in a way it never was.
Recency. Manual research is a snapshot that ages. Re-running weekly is cheap, so accounts that develop a trigger get surfaced within days rather than at the next quarterly review.
What broke
Hallucinated evidence. Asked whether an account showed evidence of a trigger, the model occasionally produced a confident yes with a fabricated justification.
The fix: require a verbatim quote from the source text for every positive finding. No quote, no finding. This eliminated nearly all of it, and it's the single most important implementation detail in this post.
Recency blindness. A funding announcement from 2022 was weighted like one from last month. The model isn't reliably reasoning about dates — pass dates as structured fields and handle time decay in layer three, not in the prompt.
Enthusiasm. Ask a model to assess fit and most accounts come back as good fits. It's optimising for a helpful answer. Extraction questions with binary or enumerated answers don't have this problem; open-ended assessment questions always do.
Silent drift. After a model version update, the same accounts scored differently. Now I keep a fixed set of 20 accounts with known-correct extractions and re-run them whenever anything changes. Ten minutes, catches drift before it reaches the sales team.
What I'd tell someone building this
Keep the model in the reading seat. It's genuinely excellent at "does this document say X, and where" and unreliable at "is this a good account."
Every failure above came from asking the second kind of question. Every success came from asking the first and doing the judgement in code you can read.
That's the whole lesson, and it generalises well beyond account prioritisation.
Hilal Tasdan
B2B SaaS Growth Marketing Consultant & Fractional CMO. Partner in Growth.