For twenty years the applicant tracking system was a filing cabinet with a workflow bolted on. Resumes went in, stages moved, emails went out. Then AI arrived as a sidebar: a button that summarizes a resume or drafts a rejection note. Useful, but the recruiter still did every click.
The next generation is different in kind, not degree. The AI does not sit beside the pipeline. It works inside it. It reads the live data, takes actions you approve, and logs what it did. That shift raises a practical question for anyone who hires engineers: what should the software do on its own, and where does a person have to stay in charge?
Three generations of hiring software
The first generation stored data. The second added AI that writes. Most teams live here today: they paste a resume into a chat window, get a summary, and paste the summary back into the ATS. The loop works, but the AI never sees the real record, and nothing it produces is connected to anything.
The third generation lets the AI act on the record itself. At 100Hires we shipped this as a hosted MCP server. It exposes the whole pipeline (candidates, applications, jobs, messages, interviews) as 130 typed tools to Claude, ChatGPT, Cursor, or any client that speaks the protocol. A recruiter types “move every candidate below score 40 to rejected and cite the threshold” and the assistant plans the calls, runs them, and reports back. Other vendors are building the same surface, and the mechanism is the same everywhere, so here is how it works.
What MCP is, in plain English
Model Context Protocol is an open standard for connecting an AI assistant to outside software. Anthropic introduced it in late 2024; Claude, ChatGPT, VS Code and Cursor all speak it today. The vendor publishes a list of tools with names and schemas: list applications, move stage, schedule email, create note. The AI client reads that list at startup and calls the tools when a prompt needs them. The user signs in once through OAuth and approves a scope that inherits the role they already hold in the ATS.
That changes two things. First, the assistant works with live data, not a pasted fragment, so “who applied for the DevOps role this week and has Kubernetes on their resume” returns the actual shortlist. Second, permissions follow the person.
If your account cannot delete a job in the interface, the assistant cannot delete it through your session either, and clients such as Claude let you switch individual tools off, so a sourcing assistant never carries the “reject application” tool. Your password stays out of it: you sign in on the vendor’s own page, and the assistant holds a short-lived token that expires on its own.
Five jobs an agent already does well
We run our own hiring on this setup, and the split settled quickly: the agent took the administrative load, the people kept the judgment. Five jobs it handles well.
- Triage. A posting for a mid-level engineer can draw 200 to 400 applications. The agent scores each one against criteria the hiring manager wrote (required stack, years, location, work authorization) and sorts the inbox.
- Follow-ups from real notes. “Draft a follow-up to everyone stuck in phone screen for more than five days” produces messages that reference what the candidate actually said, because the agent read the notes.
- Stuck-candidate alerts. Candidates who sit in a stage past a threshold get flagged to the owner, or moved, by rule.
- Batch stage moves. Fifty applicants from “new” to “screen” in one call; a rejection batch carries the reason on every card.
- Reporting. “Pull every interview from the last 30 days and summarize what each candidate said about salary” chains two tools and writes the result into the record.
None of these decide who gets hired. They decide who the human looks at first, and they remove the copy-paste.
Where the human stays in the loop
The adoption numbers say the door is open, with a condition attached. In a January 2024 Paychex survey of 309 HR professionals and 1,003 recent hires, 45% of HR professionals already used AI-driven onboarding and another 25% planned to.
Employees onboarded with AI support were 30% less likely to quit in their first year. On the candidate side, a Tidio survey of 1,068 people found that 76% would not mind being onboarded by AI, but 65% still want to meet the people they will work with. Only 31% would accept a hiring decision made by AI alone; with a human involved in the decision, acceptance rises to 75%.
That last number is the design constraint. In practice, four things keep the person in charge.
The score is an input, not a verdict. A good AI score shows its work: a total, a per-criterion breakdown, and a two-line summary a hiring manager can disagree with. A black-box number invites two failure modes, blind trust and blind distrust. The workflow our own team uses most is a candidate assessment: the assistant pulls the profile, the applications and the interview evaluation forms, checks them against the criteria, and writes a recommendation into the notes. The interviewer reads it and argues with it. That is the right relationship.
Every action is logged and reversible. When the agent moves, rejects, or messages, the candidate timeline shows which account asked, what ran, and when. A batch you regret can be walked back: select the same candidates, qualify them again or change their stage in one action. Candidate data leaves the system only when a tool is called for it; there is no background sync feeding the model.
Permissions follow the person, not the bot. The assistant works inside your role. A recruiter’s assistant cannot do what the recruiter cannot, and a sourcing assistant should run on the read tools alone.
Interviews stay human. For engineering roles this is not a philosophical point. A live technical conversation is still the highest-signal step in the process, and the agent’s job is to make sure the right people reach it faster.
Six questions to ask any “next-gen” vendor
- Does the AI see live records, or does it work from what I paste in?
- Does it inherit my role, and can I switch individual tools off?
- Is every action it takes written to the candidate timeline?
- Can I undo a batch action?
- Does the score explain itself per criterion?
- Is the integration built on an open standard like MCP, or on a proprietary bot I cannot move to another client?
A vendor who answers yes to all six has built an agent you can supervise. A vendor who answers yes to none has built a chatbot with a database behind it.
The short version
Next-generation recruiting software is not a smarter search box. It is an assistant with a scoped set of keys, a full audit trail, and a habit of asking before it acts on anything that matters. Give it the triage, the follow-ups, and the reporting. Keep the interviews, the offers, and the final call. That split is where the productivity shows up, and it is the one candidates say they will accept.
Written by: Alex Kravets is the CEO of 100Hires, an applicant tracking system for startups and small teams.
