HR Screening Stalls When AI Slide Decks Skip the Intake Form
HR Screening Stalls When AI Slide Decks Skip the Intake Form
You are screening eighty candidates this week. You have the interview grid open, the scorecard template, and the same compensation spreadsheet that has survived three HRIS migrations. The last thing you need is another tool that generates a beautiful summary of a candidate that does not match what the hiring manager actually asked for.
OpenAI just bought a presentation startup called NextSlide. The team is being folded into ChatGPT. The immediate read is that slide generation gets smarter inside an AI assistant. That is likely true. It is also beside the point for your actual work.
Your work is not slides. Your work is the intake form, the disposition code, the interview feedback loop, and the sixty-second decision about whether a candidate moves forward. That is where the friction lives.
The Category Here Is Not Presentations. It Is Workflow Compression.
NextSlide is one instance of a broader category: AI tools that promise to compress a task you already do manually. The presentation angle is the demo. The real product is the compression of judgment into a formatted artifact.
You already have two tools in that category. You have the ATS’s built-in interview summary generator, which produces a paragraph from your notes. You have the email template library that auto-fills candidate follow-ups. Both are limited. Both are used daily. Both fail in predictable ways.
The ATS summary generator works if you type structured notes. If you have a rambling phone screen or a candidate who spoke in circles, the summary looks clean and is subtly wrong. The template library works if the scenario matches. It does not handle the edge case where the candidate declined because of a commute issue.
NextSlide or something like it will eventually be bolted onto ChatGPT or your ATS. The question is not whether it can make a slide. The question is what happens when the slide has to match the hiring manager’s actual rubric.
That part is real. The rest is friction.
Who This Is For, And Who Should Ignore It
If you are an HR business partner who spends more than two hours a week converting interview notes into a formatted update for a director, this category matters. If you are screening eighty candidates a week, you need to know where the automation breaks before you trust it.
If you are a recruiter who works in a single ATS and never produces a custom summary for a hiring panel, ignore this until your vendor integrates it. The standalone version will not survive your workflow.
I will be direct: the tool is not built for you. It is built for the person who wants to present to a board. You are not that person. But the same underlying model—generating a structured artifact from unstructured input—is exactly what your weekly summary work is. That is why you need to look at this category with a skeptical eye.
A Concrete Workflow Example: The Panel Debrief Summary
Here is a scenario. Thursday morning. You have four interviews for a senior operations role. The hiring manager wants a one-page summary per candidate by 2 PM. You have notes from three panelists, one phone screen transcript, and a gut feeling that the second candidate is stronger than the notes suggest.
Without an AI tool, you do this: read the notes, write a half-page summary, flag the discrepancy between the phone screen and the onsite interview, and send it. That takes about eighteen minutes per candidate. With eighty candidates a week, that is a non-trivial chunk of your day.
With an AI slide or summary tool, the pitch is that you paste the notes and get a formatted output. The demo looks good. The output is clean. It has bullet points. It uses the right headings.
Then you check the summary against the notes. The candidate’s main weakness—unclear about the second interview’s metrics—does not appear. The tool smoothed it out. You have to add it manually. You also have to check that the tool did not invent a strength that was not in the notes. This happened in a test I ran with a similar AI summary tool: it said a candidate had “demonstrated leadership in cross-functional initiatives” when the note only said “has worked with three teams.”
The time saved on formatting is about four minutes. The time spent verifying the content is about six. The net cost is two minutes per candidate.
That is the integration friction. It does not remove the judgment call. It shifts the work from writing to checking. And checking is more fatiguing because it requires sustained attention.
You still have to check.
What Works Better Than Expected
I have to admit one thing that surprised me. The AI output for structured data—like a list of interview dates, or the candidate’s work history pulled from a resume—is reliable. If you ask for a timeline of a candidate’s job changes, the model gets it right. That is useful.
For the part of your job that is data transcription, this category works. If you have a candidate who worked at three companies and you need a clean chronological list, the tool is faster than you. It does not miss the month. It does not transpose the year.
I expected this to save time. What actually happens is closer to shifting the work. The data part is faster. The interpretation part is still yours. And the risk is that you start trusting the clean output more than you should because the data part was accurate.
That is the trap. The first three outputs are correct. The fourth one has a subtle error in the candidate’s stated reason for leaving a job. The model inferred “pursued new challenges” from a note that said “left after restructuring.” Those are not the same thing. The hiring manager will not notice. You will, if you check. But you are tired.
Where It Breaks Inside Your Real Stack
The breaking point is the intake form. Every candidate you screen passes through a structured intake: the recruiter’s notes, the disposition codes, the interview feedback fields. The AI tool does not see that structure. It sees text you give it.
If you paste the intake form text into the tool, it will produce a summary that ignores the field labels. It will not know that “Potential concerns” is a required field. It will not know that the hiring manager specifically wants the candidate’s compensation expectations flagged. You have to tell it that. That is instruction overhead.
In my experience, instruction overhead is worse than manual work. You have to write the prompt, check the output, and then edit the missing pieces. The ATS’s built-in summary generator at least knows the field names. It is dumber but it is contextual.
NextSlide or any standalone tool will not integrate with your ATS’s data model without a connector. That connector will break when your ATS updates. That is a maintenance cost you do not need.
On paper this should work. In practice the friction shows up somewhere else. It shows up in the fifteen minutes you spend wiring the tool to your process, the six minutes of verification per candidate, and the nagging doubt about whether the clean output is accurate.
Comparison With What You Already Use
Your ATS’s summary generator is worse at formatting. It produces paragraphs that look like a robot wrote them. But it does not fabricate content as easily because it is limited to the fields you filled in. That is an advantage for screening work. The AI tool is better at prose and worse at fidelity.
The email template library is limited, but it does not hallucinate. It does not invent a candidate’s motivation. It just has fixed text.
If you want to pilot something, do not pilot a standalone presentation tool. Pilot a prompt that uses your ATS’s existing text export and a generic GPT model. That gives you the same benefit without the integration risk. If you already have ChatGPT access, you can do this today with a well-structured prompt.
But be prepared for the verification cost. It is real. You are trading typing for reading. Reading is harder when you are tired.
Verdict: Pilot With Boundaries, Do Not Adopt Blindly
My recommendation is a conditional pilot. Use it for candidates where the notes are highly structured and the hiring manager’s rubric is simple. Do not use it for edge cases like a candidate who changed careers or had a gap.
The pilot conditions are:
- You only use it for the data transcription portion, not the interpretation.
- You always verify against the original notes for any claim about candidate motivation or performance.
- You time-box the verification. If it takes more than five minutes per candidate, drop the tool.
If the tool fails the five-minute verification test, avoid it. Your job is not to produce polished documents. It is to make a judgment call under time pressure. The tool does not help with that. It adds a layer that you have to look through.
That is the honest assessment. The demo looks good. The slide output looks professional. The work of screening eighty candidates does not change. It just gets a new wrapper that you have to unwrap every time.
I will leave you with this. The best tool for your job is still the one that knows the field labels and does not invent a strength you did not note. Until the AI category gets that context, it is a formatting tool, not a screening tool.
And you already know how to format.
That part is real.
```
Comments
Post a Comment