Let’s talk about why you should stop buying AI tools you cannot audit.
The Real Problem With AI in Talent Acquisition Is Not Cost. It Is Control.
I recently joined Hung Lee’s Brainfood Live discussion on a question that sounds provocative at first: what if AI actually costs more than humans? The full episode is worth watching because it surfaces a topic many TA leaders are not yet thinking about deeply enough: the real cost of AI in recruitment. (Crowdcast)
But here is my view.
The biggest issue is not token cost.
The biggest issue is that TA teams are buying AI-enabled software they do not understand, cannot audit, and may not be able to defend when challenged.
That should make every TA leader uncomfortable.
Because when AI enters recruitment, we are no longer talking about a productivity tool that rewrites emails or summarizes meeting notes. We are talking about technology that may influence who gets seen, who gets advanced, who gets rejected, and who gets hired.
And if your answer to “how did the tool reach that conclusion?” is anything close to “the vendor said the algorithm is solid,” you do not have an AI strategy.
You have a liability.
First: Token Cost Is Not the Real Problem
There is a lot of noise about AI becoming too expensive. More tokens. More compute. More model usage. More agentic workflows. More cost.
I understand the concern, but for Talent Acquisition this is mostly the wrong conversation.
Yes, frontier models are expensive. The most capable models will always cost more because they require more compute, more infrastructure, and more energy. But that does not mean every AI-enabled recruitment workflow needs to run on a frontier model.
In reality, frontier models are mainly needed to build, design, and improve complex software and advanced solutions. They are not needed for 90 percent of what automation, AI workflows, and agents actually do in TA.
Most operational use cases can run on smaller, cheaper, more focused models. Some can even run locally. And this will only become more relevant as model efficiency improves and smaller models become better at specific tasks.
The hardware side is also not standing still. Moore’s Law is often oversimplified, but the core trend remains clear: computing power has improved dramatically over decades, and transistor density has followed a long-term exponential trajectory. It is not a natural law, and the strict cost-per-transistor version is more complicated today, but the direction is still toward more compute capability over time. (Our World in Data)
Add model compression, smaller models, edge deployment, quantization, and more efficient inference, and the conclusion is simple: token cost will matter, but it is not the strategic problem TA leaders should obsess over.
The real question is not: “How many tokens will this use?”
The real question is: “What exactly does this tool do, what does it cost to run, and can I prove how every output was created?”
That is where most TA teams are dangerously underprepared.
The Agent Panic Is Also Mostly the Wrong Problem for TA
Another concern raised in the broader AI debate is that people may build autonomous agents without really understanding how much token usage they create. An agent could keep running, keep calling tools, keep reasoning, keep spending money, and nobody notices until the bill arrives.
That risk exists.
But in Talent Acquisition, this is not the main concern.
Why? Because serious TA teams will not let individual recruiters casually build autonomous agents and let them loose on onboarding, scheduling, sourcing, screening, or candidate evaluation.
At least they should not.
Recruitment processes involve personal data, sensitive decision flows, and candidate rights. In Europe, AI systems used for recruitment or selection, including systems that analyse and filter applications or evaluate candidates, are explicitly listed as high-risk under the EU AI Act. (AI Act Service Desk) GDPR also gives people protection against solely automated decisions that produce legal or similarly significant effects. (GDPR)
So no, this is not a playground.
TA teams are not simply “users” of AI. They are buyers, implementers, and accountable operators of technology that touches hiring decisions.
That changes the standard completely.
What Actually Matters When Buying AI in TA
For TA, two things matter more than the hype:
- Running cost
- Auditability
Everything else is secondary.
1. Running Cost: Stop Buying Features. Build the Business Case.
The only business case that matters is brutally simple:
How much will this solution cost to implement, operate, maintain, govern, and improve and what measurable improvement will it create?
Not the demo cost.
Not the licence cost.
Not the “AI add-on” cost.
The total cost.
Implementation. Integration. Legal review. Process redesign. Training. Vendor management. Change management. Internal support. Ongoing usage. Data governance. Reporting. Auditing.
Most AI business cases in TA are far too shallow. They focus on saving recruiter time, as if the only goal of AI is doing more admin with fewer people.
That is lazy thinking.
Yes, AI can reduce administration. Yes, it can automate repetitive work. Yes, it may allow a team to do more with less.
But the bigger opportunity is not only efficiency.
The bigger opportunity is process quality.
Imagine removing enough admin work so recruiters can finally run proper job intakes instead of taking vague hiring manager instructions. Imagine recruiters having time to challenge whether the business really needs “another marketeer” or whether the team actually needs a different capability. Imagine every hiring manager being trained and supported in structured interviewing. Imagine every candidate being properly prepared for the purpose of each interview. Imagine every role having a real decision meeting where the hiring team evaluates evidence instead of trading opinions.
That is where AI can create real value.
Not by replacing the recruitment process, but by freeing up capacity to make the process better.
Within the Recruiting Excellence Framework, the goal is not simply faster hiring. The goal is the ability to consistently deliver the right talent exactly on time when the business needs it. Technology only matters if it improves that ability across cost, timing, and quality.
So before you buy an AI tool, ask:
Will this make us better at delivering the right person exactly on time? Or will it just automate work we never properly designed in the first place?
If the answer is unclear, do not buy it yet.
2. Auditability: The Biggest Problem in AI-Enabled TA
This is the real issue.
Many AI-enabled recruitment tools are still black boxes.
They produce a score, a recommendation, a ranking, a summary, or a “match percentage,” but the user cannot see exactly how that output was created.
That is unacceptable.
If a tool influences a recruitment decision, you need to be able to audit it at individual case level. Not generally. Not theoretically. Not through a vendor statement saying “our model is explainable.”
For every candidate, you need to know:
1. The exact input
What was the prompt? What data was included? CV? LinkedIn profile? Motivation letter? Job description? Intake notes? Assessment results? Interview notes? All of it needs to be visible.
2. The exact reasoning path
What did the tool look at? Which criteria did it apply? Which documents did it use? Which data points mattered? Why did it weigh one element more than another?
3. The exact output
What score, recommendation, ranking, or decision-support output did it produce?
And you need to store this in a way that can be retrieved later.
Because if you rank 10,000 candidates per month and one year from now a candidate challenges your process, “we had a human in the loop” may not be enough to protect you.
Especially if the recruiter bulk rejected everyone below a score of 80.
Let’s be honest. Ranking candidates is already decision influence. It may not be the final rejection decision, but it shapes attention, priority, and access. If the ranking determines who gets reviewed and who does not, it materially affects the process.
That means the standard must be higher.
The ATS Matching Score Problem
Let’s take the most obvious example: the AI ranking tool inside your ATS.
Many ATS providers now offer matching scores, fit scores, candidate ranking, or recommendation tools. Some are genuinely AI. Some are not. That discussion is for another day.
The more important point is this:
If your ATS provider cannot explain exactly why one candidate received 78 out of 100 and another received 81 out of 100, that is a massive red flag.
Not “we use a robust dataset.”
Not “our matching algorithm compares candidate data to job requirements.”
Not “the model has been tested.”
Not “the recruiter still makes the final decision.”
None of that is enough.
You need the full case-level log.
The exact input.
The exact reasoning.
The exact output.
So when someone challenges the process, you do not panic. You open the tool, pull the case record, export the audit trail, send it to legal, and continue your day.
That is the level of control TA leaders should demand.
Anything less is not mature AI adoption.
It is blind trust dressed up as innovation.
The Practical Solution: Buy AI Like a Business-Critical System
TA leaders need to stop buying AI tools like nice-to-have productivity software.
Recruitment AI should be evaluated like business-critical infrastructure.
Before selecting any tool, demand answers to these questions:
What exact process does this tool improve?
If the process is broken, fix the process first.
What is the total running cost?
Include implementation, maintenance, governance, usage, support, training, and compliance.
What data does the tool process?
Know what goes in, where it goes, where it is stored, and who can access it.
Can every output be audited at individual case level?
If not, the tool should not influence candidate evaluation.
Can we export the full input, reasoning, and output log?
If not, you are dependent on vendor promises when things go wrong.
Does this improve Recruiting Velocity, Hiring Budget, or Quality of Hire?
If it does not improve cost, timing, or quality, why are you buying it?
This is the shift TA needs to make.
From excitement to control.
From demos to evidence.
From “AI-enabled” to auditable.
From tool adoption to measurable improvement.
Final Thought: If You Cannot Audit It, Do Not Buy It
AI will transform Talent Acquisition. That is not the debate.
The debate is whether TA leaders will lead that transformation properly or outsource their judgment to vendors selling black-box tools with polished interfaces.
Token cost is not the issue.
Control is the issue.
Auditability is the issue.
Running cost is the issue.
And the uncomfortable truth is this: many TA teams are not ready to buy AI responsibly because they do not yet know what standards to demand.
That needs to change fast.
At the Recruiting Excellence Foundation, we help TA teams select the right AI and automation tools, evaluate the real business case, and implement solutions that are GDPR-proof, auditable, and aligned with the recruitment process you actually need.
We also build bespoke AI tools for TA teams that want full control over their workflows, data, and audit trails.
Because in recruitment, the future does not belong to the team with the most AI features.
It belongs to the team that can prove every decision was made properly.
If you are selecting AI tools for Talent Acquisition and want to do it properly, reach out.