AdvancementSeptember 21, 2026

You can’t ask a prospect researcher for five years of AI experience. So what do you hire for?

The prospect researcher’s job is changing, and hiring should change with it. In this short article, we share our view on which skills will become more valuable as AI takes on more of the routine research, and what that means for your next hire.

Shahryar Abbasi4 min read

If you are hiring a prospect researcher, or you are one deciding what to learn next, AI creates an awkward problem. We all know AI is changing the role, but this version of prospect research has only just begun and is already moving faster than anyone can build a long track record in it. There is no settled credential that proves someone is ready.

A better question is which familiar skills become more useful when AI joins the work. From what we have seen building AI for advancement, the following two seem likely to become even more valuable.

Skill one

Attention to relevant detail

This is slightly different from the attention to detail we are used to, which catches visible mistakes: a misspelled name, an old title, a number copied incorrectly.

An AI hallucination has a different shape. A false claim arrives with the same confidence and polish as a true one: a property is linked to the wrong Jordan Lee, an estimate appears as a fact, or wealth becomes evidence of affinity. The work looks clean and confident, which is exactly what makes the mistake dangerous.

Even though newer models do this less often (OpenAI reported that GPT-5 made about 80% fewer factual errors than o3 on specific factuality tests), that improvement creates a small paradox: as the work becomes more reliable, we become more willing to trust it, so the remaining errors may travel further before someone stops to question them.

AI gets facts wrong less often than it used to
Made-up facts in a summaryhow often a summary states somethingthe document does not sayFactual errors in an answerGPT-5 against the older o3 model3.0%1.8%0.2×20232026o3GPT-5
Vectara hallucination leaderboard, best model each year. OpenAI, GPT-5 launch, August 2025.
AI makes fewer factual mistakes than it did two years ago, on both measures.

This changes how the researcher approaches the fact-check. They ask, “What supports this claim?” and then, “What would make it false?”

The researcher can try to disprove the claim or ask a separate AI evaluator to challenge it. We spend a lot of time building feedback loops like this, where one agent gathers the evidence, another challenges the conclusion, and the researcher decides whether the evidence is strong enough.

Lower error rates also do not guarantee fewer errors for the team. If AI produces ten times as much work while its error rate falls fivefold, the review queue still receives twice as many errors. Better models change where we need to apply judgment rather than removing the need for it.

What actually reaches the desk

Before AI

Ten briefs cross the desk in a day, and one of them has something wrong in it. The researcher catches it, because there are only ten.

AI fills the desk

Now a hundred briefs a day. Ninety of them would never have been written before.

And each brief is better

Mistakes are five times rarer than they were. One brief in fifty carries an error.

Two mistakes still get through

A hundred briefs at one mistake in fifty is two mistakes a day. The work got more reliable and the review queue still doubled.

Ten briefs a day with one mistake is the starting point. 100 ÷ 50 = 2.
A day on one researcher’s desk
Ten briefs a dayOne hundred briefs a dayone mistake in every fifty briefsTwo mistakes reach the desk
Briefs you had beforeAdded by AIHas a mistake in it

Skill two

Systematized multitasking

AI can set many lines of research in motion at once, while the researcher can still review only one prospect at a time. The role starts to resemble air traffic control: keeping every open loop visible without trying to hold all of it in your head.

The simplest version is one queue where every prospect has a status, an open question, and a next action. It also helps to limit the number of sensitive cases under active review. You might keep three open while the rest wait, then lower that number if you notice details starting to blur.

Three prospects open at a time
Under reviewbeing read nowwaiting on a sourcequestion sent backWaiting
Being read nowOpenIn the queue
The researcher keeps three prospects under review and the rest wait in one queue.

People who think they are bad at multitasking may be unusually good at this version because they already know not to trust memory and are used to creating systems. They keep one queue, write down the next action, and finish one review before pulling in another. Nothing disappears between a browser tab, a spreadsheet, and the CRM because the system remembers it for them.

The models are already fast and will keep changing, so we will keep updating the feedback loops and workflows around them. Risely Advancement Intelligence can draft a prospect research framework in about 66 seconds, compared with roughly an hour before (read the Westmont case study). We will continue to share what we learn as we build, so follow along if you are hiring for this role or learning it yourself.

See it work on a real campus queue.

A live walkthrough shows Risely’s agents doing this work on one shared record, with your team approving every move.