If Part 1 was about the science of talent systems and Part 2 about the craft of prediction, Part 3 is about the storm rolling in.
Artificial intelligence is transforming both sides of the hiring equation simultaneously. And it's creating a mess.
"We have an escalating arms war going on between AI systems for candidates and AI systems supporting companies, trying to create the perfect resume for every application that will score best in an ATS."
It's an arms race born of bad data. Resumes are unstandardized. Job descriptions are vague. Criteria are inconsistent. AI doesn't fix this—it exploits it. Candidates optimize resumes to match keywords. Companies deploy AI to screen those optimized resumes. Round and round it goes, optimization without substance.
The Optimization Arms Race
At first, AI-generated resumes looked impressive. Sharper phrasing, keyword-rich summaries, clean formatting, perfect tailoring to each posting. Early adopters got a small edge.
Then everyone else caught on. Now the signal is eroding.
"A resume is a very poor unstandardized set of data against a job description as they have been used previously. You're trying to draw from this content and stuff it into that. The more alignment you can create, the better, but the truth is, it's all pretty flimsy."
AI is polishing a weak signal, not increasing the signal itself. The better the tools get at keyword alignment, the less differentiation they produce. Everyone starts looking strong at the surface. Very few stand out where it matters.
The same pattern plays out on the employer side. Teams point screening models at weak inputs and hope for clarity. The model becomes confident about data that doesn't deserve confidence. Precision looks high in testing and falls apart in reality. The optimization is real, but the target is wrong.
What to do instead : Improve input quality before optimizing. Standardize job definitions. Enforce the five prioritized competencies per role. Enrich candidate data beyond keywords—work samples, code links, portfolio artifacts, quantified outcomes. When you raise input quality, optimization starts meaning something useful.
The Convergence Problem The next stage of the arms race is worse: artificial perfection at scale.
"As this escalates, you're going to end up with potentially 100 candidates for a role all scoring 96, 97, 98, 99, or 100. It's very difficult to differentiate at that point."
On paper, everyone looks exceptional. In interviews, AI-polished answers start sounding alike. Language models learn the same playbook of acceptable stories and safe phrasing. The tools designed to highlight differences end up flattening them.
Here's what that means: top-of-funnel screening becomes commoditized. Competitive advantage no longer comes from who screens fastest or writes the prettiest job posting. Advantage shifts to whoever can differentiate deeper in the funnel with structured evaluation that AI can't replicate or fake.
Back to the fishing metaphor: "In talent acquisition, we're putting out our nets in the right waters for a set time. We bring in what we catch, then carefully evaluate against our criteria. The funnel is how we pressure-test that catch until we know which one is the real outlier."
The net will collect more fish as AI increases volume. Differentiation moves to the table where you sort the catch.
Practical guardrails : Require artifacts for high scores. Use the 0-3-5 rubric and demand proof behind every 5. Design prompts that penalize memorized answers and reward genuine capability:
A product candidate chooses between two flawed roadmaps while defending trade-offs A sales candidate handles a live discovery curveball with a time box A designer iterates a broken flow under accessibility constraints These tasks can't be gamed with polish. They reveal how someone actually thinks.
The Intelligence Spectrum: Stop Thinking "Human vs. AI"
Most debates about AI in hiring get stuck in a false binary—either machines take over or humans hold the line. Adam thinks that framing misses the point entirely.
Instead, think of a spectrum with four modes of intelligence:
Pure Human Intelligence : Rare, slower, but essential where trust and meaning live. Offer negotiations. Executive references. Final judgment calls where connection matters.
Amplified Human Intelligence : Humans working with AI as thinking partners. Drafts, critiques, research synthesis, brainstorming, probing assumptions. You stay in control but operate with superpowers.
Augmented Human Intelligence : AI leads with humans in the loop. The machine proposes a shortlist or summary. You validate, correct, add judgment.
Pure AI and Automation : Scheduling, reminders, status updates, document assembly—tasks that don't require human judgment.
"The real art is deciding where each workflow belongs. Amplified is what we do most of the time these days—bouncing something off a model to amplify your efforts. Augmented is where AI takes the lead and you provide a sanity check. Full AI is fine for scheduling. The question is which parts of your workflow live where on that spectrum."
Try this exercise : List your end-to-end hiring workflow from intake to offer. Label each step with one of the four modes. Ask why. If you place a step under pure AI, write the risk if the model is wrong. If you place it under pure human, write the cost of keeping it there and the minimum augmentation that could reduce cost without harming trust. Revisit quarterly.
Where Humans Remain Essential: The Connection Factor Even as AI grows more powerful, certain parts of the funnel remain irreducibly human. Adam calls this the "connection factor."
Algorithms can optimize resumes, rank assessments, predict performance odds. They cannot build trust with a candidate who has options. They cannot sense the spark in a conversation that makes someone believe in a mission. They cannot persuade an outlier to choose your environment over a competitor's.
This isn't romanticism—it's competitive reality. Top performers weigh three questions AI can't answer: Do I trust the people? Do I believe in the problem? Will this place help me grow?
Human conversations, referrals, and leadership presence answer those questions. No model can stand in for that final stretch.
The strategic implication : Protect time and energy for high-value interactions. Standardize and automate the early funnel. Concentrate human attention on high-leverage moments—live work sessions, team meet-and-greets, executive touchpoints, decision huddles. This is where belief, motivation, and commitment get forged.
The arms race will accelerate. Resumes will get more polished. Screening will get more automated. But the winners won't be those with the best AI—they'll be the ones who built systems that can reliably identify and connect with outliers when everyone looks perfect on paper.
That's the future Adam's been building toward: systems that treat hiring as a learning loop, prediction as a science, and human judgment as the irreplaceable core of what makes great teams.
Case note: Why Behavox chose SquarePeg
When Hopewell led the redesign at Behavox, the team evaluated several AI screening tools. The goal was not to add another keyword filter. The goal was to raise the quality of inputs, make early signals comparable, and keep the entire pipeline explainable to executives, Legal, and candidates.
Hopewell outlined five requirements.
Enrichment beyond keywords. The tool had to normalize titles, map skills to a shared taxonomy, extract outcomes, and flag timeline issues. SquarePeg’s enrichment turned free text into structured signals that mapped cleanly to the five competency job briefs.
Explainable rankings. Any shortlist had to include rationale that a hiring manager could read in under a minute. SquarePeg attached evidence notes to its rankings so reviewers saw the why, not just the who.
Fraud and duplicate detection. With rising volume, the team needed early checks for duplicate profiles, credential inconsistencies, and obvious spam. SquarePeg’s fraud flags helped clean the pool before interviewers invested time.
Speed to live and integration. Behavox wanted an ATS first integration, single sign on, and audit trails with minimal lift from IT. SquarePeg connected quickly and pushed enriched profiles back into existing workflows.
Compliance and auditability. Legal and Security asked for SOC 2 controls, privacy by design, and a clear view of how recommendations were produced. SquarePeg’s glass box approach and independent assurance met those requirements.
The outcome was a better early funnel. Recruiters reviewed top candidates in under an hour for key roles. Hiring managers trusted the shortlists because they could see the rationale and the artifacts that supported it. Post hire, the 30, 60, 90 checkpoints fed back into the model weights and interview prompts. The system learned because the data was clean and the decisions were visible.
Hopewell’s summary of the choice was practical. The tool had to make resumes comparable, not just prettier. It had to show its work. It had to plug into the five competency model without forcing a new language on the team. SquarePeg checked those boxes and kept the glass box visible for stakeholders who needed to understand how calls were made.
Operating Model: Design for AI Without Losing the Plot The companies that thrive won't just be the ones deploying the most AI tools. They'll be the ones who redesigned their systems around AI—automating what should be automated, amplifying human strengths where they matter, and doubling down on structured evaluation in the later funnel.
Five Design Principles Start with better data. Standardize job definitions to five competencies. Use enriched, structured candidate data instead of keyword search alone. Create shared taxonomies for skills and levels. AI trained on garbage produces garbage at scale.
Instrument for outcomes. Track ROTI, quality of hire at 90/180/365 days, offer acceptance, and time-to-signal. Use these to audit whether your models are actually predicting what matters.
Keep the glass box visible. Document how models make recommendations. Store rationale for shortlist decisions. Make it easy to explain choices to candidates and stakeholders. "The algorithm said so" isn't an answer.
Enforce evidence for exceptional scores. Any 5 on the 0-3-5 scale must include an artifact or cited observation. This rule keeps hype in check and forces people to show their work.
Protect the human edge. Reserve time for connection and persuasion. Train managers on the conversation skills that close outliers. The best system in the world is worthless if great candidates choose your competitor.
Governance Without Bureaucracy Create a simple risk register for AI in talent. One page, not fifty:
Name the model uses you allow Record the data sources List potential harms (proxy bias, over-reliance on weak signals) Define mitigations (fairness checks, red team reviews of prompts, candidate appeals process) Keep this document alive, not filed away.
On security and privacy : Treat candidate data as sensitive. Limit prompts that include personal information. Prefer on-platform analysis over ad hoc uploads. If you operate in multiple regions, default to the strictest standard.
Data Standards: Raising the Floor The arms race feeds on weak inputs. The most powerful countermeasure? Raise the floor on data quality.
Job data : Five competencies per role. Clear definitions for 3s and 5s. Assessment method named in advance.
Candidate data : Extract structured signals from resumes and profiles. Map titles, tools, and outcomes to a shared taxonomy. Deduplicate profiles. Flag fraud and timeline inconsistencies.
Interview data : Store question banks and scoring rubrics. Collect evidence notes and attach artifacts. Track inter-rater patterns to detect drift.
Outcome data : Feed 30-60-90 day checkpoints back into the system using the same 0-3-5 language.
When you keep these layers clean, AI becomes a multiplier instead of a noise machine. Models can rank signals rather than words. Humans can see why the model made a call. Post-hire results can tune the prompts and weights.
Measurement That Keeps Everyone Honest In Part 2, ROTI replaced vanity metrics as the north star. In Part 3, add model accountability on top:
Precision and recall by stage : Measure what the model proposes versus who actually advances. Aim for clear thresholds by role and level.
Drift detection : Watch for changes in pass rates by source, region, or demographic group. Investigate the why before adjusting prompts.
Decision documentation : Save a short rationale for final decisions, human or machine. When decisions are challenged, you can show your work.
Cycle time to confidence : Track how quickly the system produces a shortlist you'd be comfortable defending to leadership. Speed only matters if confidence stays high.
A Real Example Consider a company hiring enterprise account executives. The team maps the role to five competencies: discovery, multi-thread deal strategy, forecast accuracy, domain storytelling, and negotiation hygiene.
Solve for X screens for evidence of these five. Solve for Y uses a live discovery role-play with two curveballs. Solve for Z probes how candidates navigate slow legal cycles and competitive pressure.
The funnel runs 12-6-3-1 on a two-week clock. AI helps summarize discovery transcripts and flags common failure patterns. Humans review the summaries and score with the 0-3-5 rubric.
At 30, 60, and 90 days, managers score new hires on pipeline health, stage hygiene, and forecast deltas against reality. The team closes the loop by tuning prompts and reweighting questions that predicted 60-day performance.
Over two quarters, win rates rise while cycle time falls. The system learned because the people running it insisted on evidence and shared language.
Why This Actually Matters "We lean into the science of talent acquisition much more in the later stages."
The message isn't to abandon AI. It's to place it correctly. Use AI to expand the net, clean the inputs, accelerate analysis. Use humans to judge outliers, earn trust, and close.
The future of hiring isn't a battle between humans and AI. It's a design problem. Put the right intelligence in the right step with the right guardrails. Build the connective tissue that makes the pieces work together. Then keep learning every quarter.
Closing Thought The arms race will continue. AI will keep making resumes shinier, interviews smoother, and candidates harder to tell apart.
The organizations that win won't be the ones gaming the top of the funnel. They'll be the ones who embraced structured prediction, human connection, and systems that learn over time.
In the end, the question isn't whether AI will replace recruiters. The question is whether recruiters will build systems that harness AI without losing what only humans can do: see the outlier, earn their trust, bring them into the fold.
That's the system Adam Hopewell built. That's the future of talent acquisition for anyone willing to do the hard work of thinking systematically instead of tactically.
The frameworks are here. The science is proven. The only question left is: will you build it?
.svg)
.svg)
