You Spend 9.5 Hours a Week on Resumes. Build the Loop That Gets Them Back.

Hiring is not hard because people are bad at it. Hiring is hard because you keep doing the same five steps by hand, every single time a role opens.

Why your hiring pile never gets smaller

Picture a fourteen-person HVAC and plumbing company that needs two field technicians before the summer service season starts. The owner posts the role, forty-one resumes show up in a week, and now the actual work of running the business waits while someone reads all forty-one, one at a time, twice.

That is not a staffing problem. That is a workflow with no memory. Every time a role opens, you start from a blank inbox and redo the same five steps you did last time: write the post, read the pile, rank it in your head, chase the good ones, forget why you passed on the rest.

Two engineers on LinkedIn's own hiring team put a number on this. Talking to small business hiring managers, they found the average owner spends 9.5 hours a week just reviewing candidates and deciding who to reach out to. For a small team, that is close to a full day gone before a single interview happens. Their reframe is the useful part: hiring is not a one-shot task where you post a job and the right person appears. It is a loop. You post, you see who applies, you realize some of your "must-haves" were really "nice-to-haves," you adjust, you source more, you re-evaluate. Plan, act, observe, adapt, again.

LinkedIn built their version of that loop with an engineering team, a framework called LangGraph, and a full observability stack. You do not need any of that. You need the same four stages, built with tools you can turn on this afternoon.

The loop, in four stages

Strip the enterprise architecture away and the loop underneath is simple enough to sketch on a napkin. Here is the version worth building.

The detail that makes this a loop and not just a checklist is the arrow curving back to step one. Every hire, good or bad, teaches you something about what "must-have" actually means for this role. That feedback is supposed to change the next posting. Most small businesses throw that information away and start from zero every time. Do not.

Build it this week: the five-step version

You do not need LangGraph, a planner model, or an engineering team of two. You need a form, a prompt, a spreadsheet, and one rule about when a human has to sign off. Here is the buildable version, tool by tool.

Step 1: Build the intake form. Use Google Forms, Typeform, or a simple Notion database. Ask for exactly four things: role title, the three non-negotiable requirements, the pay range, and the deadline. Keep it to one screen. This is the "context" the rest of the loop runs on, so it has to exist before anything else does.

Step 2: Turn intake into a job post with one prompt. Paste the four answers into Claude or ChatGPT with a prompt like: "Using this intake, write a job post in plain language, under 250 words, that leads with the three must-haves and states the pay range openly." Read it once, fix anything that sounds generic, post it. This step used to take an hour. It now takes the time it takes you to read one paragraph.

Step 3: Screen applicants against your own bar, not a generic one. As resumes or WhatsApp/email applications come in, drop each one into the same AI tool with your three must-haves from step 1 and a fixed instruction: "Score this candidate 1 to 5 against each must-have, explain the score in one line each, then give an overall recommendation of interview, maybe, or pass." Paste every result into one spreadsheet column so you get a ranked list instead of a pile. This is the exact move LinkedIn's engineers described as "evaluate the applicants for you based on the qualifications that you define," just run by hand instead of by a framework.

Step 4: Put a human at the only decision that matters. The agent ranks. It does not reject. Read the "pass" column yourself before anyone gets an automated no, and skim the "maybe" column especially closely; that is where AI screening tools get accused of throwing away good candidates who did not phrase their experience the way the model expected. This single rule is the difference between an assistant and a liability.

Step 5: Schedule, then close the loop. Book interviews with Calendly or Cal.com so nobody plays email tag over a time slot. After you hire, or after you decide to reopen the role, spend five minutes updating your step 1 template: which must-have turned out to matter, which one did not, what pay range actually attracted people. That update is what makes the next round faster than this one.

The one lesson worth stealing more than the tech

The most useful thing in LinkedIn's talk was not the framework choice. It was a lesson about what happens when you let the loop run without limits. Their engineers found that letting the model make free-form decisions at every step produced inconsistent, hard-to-trust output, so they built fixed checkpoints: a template the AI has to fill in the same way every time, and specific points where the loop has to stop and hand control back to a person.

You should copy the checkpoints, not the automation. Your screening agent should always output the same three things in the same order: score, one-line reason, recommendation. That consistency is what lets you scan forty results in five minutes instead of rereading forty different-shaped answers. And your "pass" and "hire" decisions should never happen without a human reading them first, not because the AI is unreliable, but because a hiring decision is the one place in this loop where being wrong is expensive and hard to undo.

What this actually costs to build

Total new tooling cost for a first version of this loop: close to zero. The form and the scheduler have free tiers that cover a small business easily, and the screening step runs on whatever AI subscription you likely already have open in another tab. The investment is an afternoon of setup, not a budget line.