I ran a session for a recruiter community this week on MCP connectors, and the best question of the whole hour came in before we even started: what are the must-have AI workflows for a small firm? 

Honestly, it's the question I get from almost every firm I talk to, right? Everyone knows AI is a big deal. Almost nobody knows where to point it first.

So here's the answer. I run operations at CoRecruit, which somehow also means I run our coaching, our go-to-market and our marketing, and one simple setup powers all of it: connect AI to the conversations your team is already having, then ask it good questions.

What is an MCP connector?

An MCP connector is a link that lets an AI assistant like Claude read information from the tools your firm already uses, such as call notes, your ATS, your CRM, email and your shared drive. Instead of pasting notes into a chat window, you ask a question and the assistant goes and gets the context itself.

What you can connect:

• Call and meeting data: transcripts, participants and notes from your phone, video and in-person conversations.
• Your ATS or CRM: the source of truth your team keeps updated.
• Email and Google Drive: the context that happens outside of calls.
• Research tools: things like keyword research or your own website, depending on the task.

Setting one up is mostly a copy and paste job. You grab the server URL from your tool's knowledge base, add it as a custom connector in your AI assistant, log in, and you're connected. That's the whole thing. 

The interesting part is what you do next.

Why does AI give recruiters generic answers?

AI gives generic answers when it only knows what you typed into the prompt. A better model won't help with that, because the missing ingredient is your context: the calls, notes and conversations that make your firm different. Give the assistant that context and the output starts sounding like you instead of like AI slop.

Where your firm's context sits:

• Phone calls: 91% of the recruiter conversations running through CoRecruit.
• Video calls: 5%.
• In-person meetings: 4%.
• Email and notes: everything else your team writes down.

Models get better every few weeks, and I'm happy about that, but a smarter model doesn't automatically make AI more useful to your business. Context does. And if your note taker only joins video calls, it's missing the vast majority of what your team says all day. 

That's the first thing I'd fix before building anything fancy.

Is it safe to connect AI to your recruitment data?

It can be, if you start with read-only access. Read-only means the AI can pull information into the conversation but cannot edit anything inside your ATS, CRM or call platform. Start there, stay in control of what it can reach, and only add write permissions once you trust the workflow.

How to stay in control:

• Start read-only: the assistant can look but can't change anything.
• Set permissions per connector: let it always access a source, or make it ask you first every time.
• Add write access later: only after you've run read-only workflows for a while and know exactly what you want it to do.
• Check your obligations: review client contracts and your firm's data policies before connecting candidate data.

If you've never touched a connector before, please, please, please start with read-only. Write access is a different conversation, and a more nervous one. You can get a huge amount of value just from letting the AI read and answer questions.

What is the 10-80-10 principle for AI?

The 10-80-10 principle splits any AI task into three parts. You spend the first 10% giving context and describing the outcome you want, the AI does the middle 80% of the heavy lifting, and you spend the last 10% editing and coaching the result.

The three parts:

• First 10%, you set it up: the context, the data and the outcome you want.
• Middle 80%, the AI works: it pulls the context, searches, and drafts.
• Last 10%, you coach it: your opinion, your standard and your call on what goes out.

I borrowed this framework from much smarter people, and the last 10% is where all the value is. Think about hiring a new sourcer. Day one, you say, "find me ten candidates," and they come back with ten candidates. Some are great, some are hit and miss, and your job is to say this one is good because of this, that one is bad because of that. 

AI is exactly the same. Tell it the outcome, show it an example of good, then coach it like you'd coach a person.

Why should you filter AI by call type?

Because a business development call, an intake call and an interview all have different goals, and mixing them gives you muddy results. CoRecruit's MCP connector returns the meeting type along with the participants and context, so you can tell the assistant to pull only the calls that fit the job.

Match the call type to the job:

• Business development calls: coaching scorecards and rep baselines.
• Client intake and candidate calls: market maps and lead magnets.
• Client and candidate questions across calls: content topics.
• Interviews and reference checks: leave them out of sales analysis, or the data gets noisy fast.

Ask for a coaching analysis without specifying the call type and you'll end up with reference checks scored against a sales scorecard. Nobody wants that.

Which AI workflow should a recruitment firm start with?

Start with whichever one hits your biggest bottleneck, not all of them at once. For most firms the three strongest options are call coaching, market maps and content creation, because each one turns conversations you already have into something your team can use. Pick one, run it, then expand.

Three workflows worth starting with:

• Call coaching scorecards: score a week of business development calls against what good looks like, find the weakest part of your process, and coach there.
• Market maps: turn 90 days of intake and candidate calls into an anonymized one-pager covering roles, salary bands and client experience.
• Content from your own calls: mine client and candidate questions for topics, check demand, draft, then edit.

My background is building and leading sales teams, so I'm biased, but coaching is the one I'd start with. When I was a sales manager I'd spend a couple of days every week listening to calls and scoring them by hand. Now Claude pulls the week's business development calls, builds the scorecard and scores the team, and a report lands in our Slack channel every afternoon so reps get same-day feedback. Does that mean I stop listening to calls? Absolutely not. It tells me which calls deserve my time.

The content one saved me when I took over marketing at CoRecruit with zero marketing experience. Your team's conversations are full of questions and objections your future clients are also asking. That's content nobody else can buy.

How do you turn an AI workflow into a skill?

A skill is a saved set of instructions that tells your AI assistant how to do a task the same way every time. Once you get an output you like, ask the assistant to draft a skill from the work you just did, read it through, and save it. Next time, one request gets the same result.

How to build one:

• Run the task and coach it: keep giving feedback until the output matches your standard.
• Ask for a draft skill: "based on what we just made, can you draft a skill?"
• Read it carefully: check that it follows what you consider best practice.
• Save it: from then on, the assistant knows the order of operations and what to pull.

Remember the scene in The Matrix where Neo gets plugged in and says, "I know kung fu"? A skill is that for your AI. And you don't even need to know how to write one. I have no clue how, so I ask the AI to do it, then audit the result. A good follow-up is to ask, "based on the work we're doing, are there any other skills worth making?"

How do you start without overcomplicating it?

Pick your single biggest bottleneck, connect one data source, run one audit, and compare the result against your team's baseline. Then ask the AI to draft a skill from what you just ran. Starting all three workflows at once is wildly intimidating, and it's how most people stall.

Your first week:

• Connect one source: read-only access to your call data is plenty.
• Run one audit: coaching is my pick, but use your own bottleneck.
• Save the baseline: you'll want something to compare against.
• Draft a skill: so the next run takes minutes instead of an afternoon.

One more tip: talk to your AI instead of typing. Use the voice button on your desktop app or phone and just brain-dump what you want. You'll get better results faster than polishing a prompt, and it's a lot less painful.

This all compounds, too. Better calls raise your close rate at each stage, which sharpens your pitch, which gives you better stories to put out in the market, which brings in better-informed buyers, which makes the next call better. Ask the right questions of your own conversations and the flywheel keeps turning.

Josh Kirkham is VP of Operations at CoRecruit, where he leads operations, go-to-market, marketing, and coaching, and has built and led sales teams throughout his career.

‍