The first hire I planned, and never made

There's a gap between "I use AI for research" and "I have a research agent," and it's wider than it sounds.

Almost everyone reading this is in the first camp. A question comes up, you open a chat window, you get a decent answer, you move on. Useful, and it never shows up as recovered time on a calendar.

Why not? Because it only happens on the weeks you think of it, and the weeks you most need it are the weeks you're too busy to remember.

3 properties close that gap, and none of them are about which model you picked.

A named output artifact. The assignment produces a specific document that lands in a specific place: a one-page brief, a change log, a ranked question list. If you can't name the file, what you have is a topic.

A recurring slot. It runs Monday at 8 whether or not you thought of it. An agent that runs when you remember is a habit with better branding.

A verification step. Somebody checks the sources before the output gets used. This is the property people skip, and skipping it is what quietly turns the whole thing from an asset into a liability.

That third one deserves more than a bullet, because the trust problem here is genuinely different from the one in your own archive.

An agent reading your notes can retrieve the wrong thing. An agent reading the open web can produce a source that never existed, and models are measurably worse at citations than at almost anything else they do.

Published benchmarks put citation accuracy at the bottom of the pile. Reported hallucinated-reference rates run from the low teens well into the majority, depending on the model and the subject matter.

That's an argument for building the check in from day 1. It's 3 moves and it costs about 15 minutes a week per assignment.

Open every cited URL, all of them. Confirm the date on anything time-sensitive, because a correct fact with a 3-year-old date is a wrong answer when the claim is about now. And actually read the source behind any number you're about to put in your writing, your pricing, or your pitch.

Here's my hard-and-fast rule on this: if a claim is load-bearing, I've read the page it came from. Everything else can ride on the summary.

3 assignments, in the order I'd stand them up.

Content prep. Before writing anything substantial, assemble what's already been said, what the current numbers are, and where the consensus is soft. The artifact is a one-page brief whose last section is 3 questions the existing coverage leaves open, and that section is usually where the piece I actually write comes from.Competitive intel. Watch a fixed list of 5 or 6 companies and report only what changed. The artifact is a change log rather than a report, because a summary of a competitor reads the same every week and the delta is the part you need.

Customer and lead research. Before a call, a 5-line brief on who this is and what they probably need. Monthly, a ranked list of the questions that keep coming up across replies, forms, and calls. That second one is the sleeper: it's the best content-calendar input available and it rarely gets built, because building it by hand is miserable.

Now the part that decides whether any of this works. The prompt.

Open-ended prompts return essays. Constrained prompts return artifacts, and the constraint that does the most work is this one:

That single instruction converts a confident guess into a visible gap, which is exactly the behavior you want from something you're going to lean on. Pair it with a recency window so the agent stops handing you a 2019 blog post as current.

The full prompt spec runs longer than an email should. I put the whole thing in this week's post, along with the 4-hour arithmetic worked line by line and the 5 ways this goes wrong.

I worked the time-savings claim honestly, which meant charging it a tax most write-ups leave out. The gathering these 3 assignments replace ran about 4 hours a week for me. Verification runs 45 to 60 minutes on top, so the real recovery lands nearer 3.

The full breakdown has the complete prompts, the arithmetic, the 3-move setup, the 15-minute verification routine, and the failure mode that turns a useful agent into a liability: The Research Agent That Replaced My First Hire

A written week on the site, and this piece is the third in a chain worth reading in order if you're building the stack.

The 5-Tool AI Stack for a One-Person Growth Team named 5 roles worth a slot and gave the Researcher exactly one line. This week's post is that line getting its full breakdown, widened past lead enrichment into the 3 assignments above.The companion is Your Second Brain Needs a Search Agent, which covers the inward-facing half: an agent that answers from your own notes and calls. The 2 halves have different trust problems, and running both means knowing which check belongs where.

Social Media Examiner's 2026 AI Marketing Industry Report (published August 2026, 681 marketers surveyed in July). The headline number is saturated: 73% use AI every day, up from 37% in 2024. The one a few pages later is the one worth your time, because only 11% have an autonomous agent running in a regular workflow and another 24% are experimenting. That's this whole issue in 2 statistics: near-universal daily use, and about 1 in 9 who've turned it into a standing job.

GhostCite: A Large-Scale Analysis of Citation Validity (arXiv). This is the paper behind the numbers up top. Benchmarked across 13 models, citation hallucination rates ran from 14.23% to 94.93% depending on the model and the subject, and a sweep of 2.2 million real citations across 56,381 published papers found the invalid ones up 80.9% in 2025. It's a corner of the internet marketing people never visit, and it puts a number on the exact failure your verification step is there to catch.

80% of B2B tech buyers now use AI agents (August 17, on IDC's research). 8 in 10 B2B technology buyers now run agents as part of the buying process, and 71% prefer digital channels even on complex purchases. Your buyers are already pointing an agent at you. The asymmetry only cuts one way if you're not running one back.

AI Shopping Agents and Agentic Commerce 2026 (August 17). Agent usage stacks at the front of the journey: roughly 62% at product comparison, 23% at checkout, 19% post-purchase. That's a map of where the trust runs out, and it's the same boundary you're managing inside your own research agent.

Pick your next piece of writing and run one prompt against it before you start. Just one, and only the content-prep assignment.Ask for 3 questions the existing coverage doesn't answer, and require a source URL and a date on every claim, with the instruction to drop any claim it can't source.

Then open every URL it gives you. That last step is the whole exercise, and how many of the links check out will tell you more about which model to trust than any benchmark will.

William DeCourcy, Professor Leads

#ProfessorLeads #LeadGeneration #B2BMarketing #B2CMarketing #PerformanceMarketing #MarketingMetrics