The pitch every AI sourcing tool makes
Open any AI sourcing product built in the last eighteen months and the pitch reads the same. Paste a role, get a list of candidates. Higher match scores, faster shortlists, less LinkedIn. Gartner's 2025 hype cycle placed generative AI in recruiting near the peak of inflated expectations, and G2's AI sourcing category has more than doubled in listings year on year.
The demos are impressive. The retention data is not.
Specialist recruiters, the people who run one niche deeply, tend to churn off these tools within a quarter. That is not a UX problem. It is a category error.
What a specialist actually does
A specialist recruiter is not running a keyword search. They are running a judgement engine that they have trained on their own market for years. They know that a "Head of Data" at a Series B fintech is a different animal to the same title at a listed insurer. They know which competitor teams are quietly disbanding. They know the two people you never call at the same time.
That context does not live in a job spec. It lives in their inbox, their notebook, and the last twenty conversations they had.
Where the current AI wave breaks
Three honest failures show up again and again.
1. Generic taxonomies. Most AI sourcing tools normalise titles and skills against a global taxonomy. Useful at volume, useless at depth. When your entire desk lives inside "clinical operations" or "commodities trading", a global taxonomy flattens exactly the distinctions you were hired for.
2. No memory. Every search starts from zero. The tool does not know you already spoke to this person in March, that they declined for a specific reason, or that their manager just moved firms. Signals compound; these tools do not.
3. Output is a list. The specialist's actual job is not producing a list. It is producing the next move, in their voice, with the right context. A ranked list is the raw material, not the work.
What the intelligence layer needs to do
The fix is not a better ranker. The fix is a partner that:
- Reads the same market you read, in your niche vocabulary, not a global one.
- Remembers what you learned last week and folds it into what it shows you today.
- Drafts the next move, sourced and attributed, in a voice that sounds like you.
- Sits alongside the CRM you already run, so nothing gets migrated and nothing gets lost.
That is a different product category to sourcing. It is closer to how a senior recruiter would describe a great junior researcher: someone who does the reading, remembers the context, and hands you a shortlist you can act on without re-doing the work.
The honest read
AI sourcing tools are not useless. For high-volume, low-specialisation desks with predictable role shapes, they meaningfully compress time to first shortlist. G2 reviews back this up in RPO and volume agency contexts.
For the specialist market, they are the wrong shape. Not because the models are weak, but because the job is not "find more candidates". The job is "compound my judgement over years". That is a memory problem, a language problem, and a voice problem before it is a ranking problem.
Holi is built for that problem. The intelligence layer between what you learn and who you place.