Let AI Rank. Never Let AI Reject

I hold a Singapore passport and I live in Dubai. That gives me an unusual seat: two of the world’s most ambitious business hubs placing the same bet at the same time, in two very different styles. Every company operating in either market is working inside that bet. 

Something changed in the last two years, and most companies have not caught up with what it means. For thirty years we bought automation: software that does the steps you specify. Now we are buying agency: software that takes a goal and works out the steps itself. 

The difference is the difference between a lift and a taxi. In a lift you press floor twelve and it takes you to floor twelve. Press the wrong button and it takes you to the wrong floor very efficiently. In a taxi you say “take me to the airport” and the driver picks the route, avoids the traffic and makes a hundred small decisions you never see. 

We can now run software agents that take a brief (twelve data analysts in Dubai, starting in March) and then go and do the work: find the people, write to them, ask the first questions, book the interviews, and come back with a shortlist and the reasoning behind it. Hiring cycles that took nine weeks now take three. 

Recruiters stopped doing the work and started supervising it. Some loved it. Some found it much harder, and they were right to: when you do the work you know it is right, and when you supervise it you have to decide how much to check. 

That is the question now landing on every executive desk and in every boardroom. Not “what can we automate?” That debate is finished. The new question is: what are we willing to hand over, and how will we know it was done properly? 

Companies in Singapore and the UAE are not starting this conversation from zero, because their governments are not. The UAE appointed the world’s first Minister of State for Artificial Intelligence in 2017 and set out a national AI strategy to 2031. Dubai wrote its own assurance rules and seal. Abu Dhabi has put AED 13 billion behind becoming an AI-native government. Singapore has had a national AI strategy since 2019, built a testing regime for AI systems through its regulator, and is now merging its workforce and skills agencies into one. The instincts differ: the UAE states the ambition and funds it, while Singapore builds the institution and tests it. Both have pointed in the same direction for nine years. For any business operating there, expectations on how AI is used with people are arriving faster than in most markets, and the companies that get ahead of them will set the pace rather than react to it. 

What the machine should do, and what must not

When someone asks me for a list of tasks to hand over. I use three questions instead, and they work on any process in any business. 

1. Does it repeat? If something happens a thousand times a month in roughly the same shape, a machine will do it better than a tired person at four in the afternoon. Reading applications. Checking eligibility. Scheduling. Chasing documents. 
2. Can it be undone? If a mistake can be corrected tomorrow at no cost to the person affected, hand it over. If a mistake quietly ends someone’s application or career, keep a human on it. Can you explain it in one sentence to the person affected? Not to your board, but to the applicant. If you cannot, that step is not ready to be automated, however accurate the model is. 

What falls out of those three questions is a simple division of labour: the machine prepares the decision, the human makes it. In hiring that has one concrete rule, and it is the most useful thing I can offer any leader: let AI rank, never let AI reject. 

A ranking is an opinion, and opinions can be overruled. A rejection is a door closing. In a regulated industry or a well-known brand especially, a door that closes with no name attached becomes a complaint, then a viral post or a lawyer’s letter, then a question from a regulator. “The algorithm decided” is not an answer anyone can give. 

The risk here is not the dramatic one. Nobody is about to let software pick your next CFO. The quiet risk is that humans stop reading. You keep a person in the loop on paper, and that person approves ninety-eight percent of what the system proposes, because the system is usually right and the queue is long. A human in the loop who is not actually looking is not oversight. It is decoration, and it is worse than nothing, because it makes you feel safe. 

So when you design the process, do not just ask where the human sits. Ask what that person is expected to examine, how long it should take, and what happens when they disagree. If disagreeing is inconvenient, nobody will disagree. 

The CV is the bottleneck 

The CV is a semi flawed document and we have been pretending otherwise for five hundred years. It is a sales document about the past, written by the candidate, edited for the reader they imagine. It tells you where someone has been and almost nothing about what they can do next. 

It also protects the wrong people. The CV rewards those who attended schools we recognise, worked at companies we have heard of, and know how to write about themselves the way employers like. It penalises the mother returning after six years, the person who learned online, the graduate from a university you cannot place, the manager whose title says nothing about what she actually did. I ran an executive search firm for twenty years. I know precisely how that game is played, because we coached people to play it. 

Three layers replace it. 

1. Skills: what a person can do, rather than what their title says. Titles mislead: “associate” or “executive” means twelve different jobs across twelve business units. In our own platform we work with a taxonomy of more than 35,000 skills and match on meaning rather than keywords. The most valuable thing it ever showed me was how close two very different-sounding jobs actually are. A claims assessor and a compliance analyst share most of their working skills. In one organisation we watched one of those roles being cut while the other was being advertised, in the same month, because the titles did not match. 

For a company, that single insight is worth more than any hiring tool. It is the difference between paying severance and paying recruitment fees in the same quarter, or simply moving people. 

2. Evidence: proof rather than claims. Anyone can write “strong analytical skills”. Ask instead for a small piece of the real work, marked against a rubric agreed in advance. This is where AI genuinely earns its place: it can give every applicant a test rather than only the forty who survived the CV screen. The ranking after evidence consistently looks different from the ranking after CVs. Different people rise.
3. Potential: what someone could do next. This is the most valuable layer and the most dangerous. Valuable because the role you are hiring for will change within eighteen months. Dangerous because “potential” has historically been the word we use when we promote people who remind us of ourselves. Measure it with evidence, such as learning rate and performance in adjacent tasks, never with a feeling.

None of this scales without a shared vocabulary. Singapore’s quiet achievement was not the funding it put behind lifelong learning. It was agreeing nationally on one way to describe skills, so that a training course, a job advertisement and a workforce dataset mean the same thing by the same word. The same lesson applies inside any large company: one skills language across HR, learning and the business. Nobody cuts a ribbon for a taxonomy. It is still one of the highest-return infrastructure decisions available to any large employer this decade. 

It is also the reframe I would offer anyone working on nationalisation quotas, Emiratisation included, or on diversity hiring targets. A national graduate is rarely unqualified. They are frequently un-adjacent: described in the wrong vocabulary and invisible to your recruiters’ search. Fix the vocabulary and the pool changes shape without a single standard being lowered. 

Fairer than what? 

The fear in every boardroom is that AI will bake bias into hiring at scale. It is a reasonable fear and it has a famous example: a large technology company built a CV screening tool, trained it on ten years of its own hiring, discovered it had learned to mark women down, and scrapped it. 

Look closely at what happened there. The machine did not invent that bias. It inherited it, accelerated it, and made it visible. 

That last word is the one that matters. When a human panel is biased, you cannot audit it. You cannot ask a hiring manager to produce the weights he applied on a Tuesday afternoon after a bad commute. Human bias is untested, unlogged and unfixable at scale. Machine bias is measurable. It is uncomfortable, but measurable, and anything measurable can be managed. 

So the question is never “is AI fair?” It is: fairer than what? Fairer than your current process, measured how, over what period? 

Here are four rules I would write into any vendor contract. 

1. Define fairness in writing before you buy anything. Fair to whom, across which groups, at which stage? Equal selection rates and equal accuracy are different objectives and are sometimes incompatible. If a vendor says “our AI is unbiased” and nobody has defined the term, you have agreed to nothing. 
2. Measure outcomes by stage, monthly. Screened, assessed, interviewed, offered: check pass rates across the groups you care about. Annual review is not enough, because bias drifts as the data drifts. 
3. Require a one-sentence reason per candidate. If your team cannot tell an applicant, in plain language, why they were ranked where they were, that system should not be making decisions about people. 
4. Have it audited by someone who is not selling it to you. That is the entire point of independence. 

We put our own platform through a national testing framework early. It was slow, expensive and uncomfortable, and it made the product better. We came out knowing things about our own models that we had not gone looking for. My advice to companies is to stop treating assurance as a compliance cost and start treating it as buying power. When large enterprise buyers declare they will only purchase AI that passes an independent test, the market builds different products. The companies that insist on this now will set the standard their competitors later have to meet. 

One last point on fairness, and it is personal. I also founded a small venture working on employment for people on the autism spectrum. What we learned there generalises. Change the format of the assessment: drop the unstructured interview, publish the criteria in advance, allow written responses, and test the actual work. When you do, the so-called talent shortage changes shape. A great deal of what we call a shortage of ability is a shortage of suitable assessment. 

Efficiency is a one-off; intelligence compounds 

Most companies buy AI for hiring to save time, and it works. Time to hire falls sharply, and cost per hire falls by roughly half. In one deployment with a large manufacturer, around 7,000 applicants were pre-screened and shortlisting time dropped by about 40 percent. 

Now let me argue against my own sales pitch. Efficiency is a one-off. You take it, you bank it, and it is gone. You cannot cut the same sixty percent twice. If speed is the only reason you did this, year three will disappoint you. 

The value that compounds is the structured data you start generating about human capability, at the scale of a business unit, a company or a sector. That is workforce intelligence, and it answers five questions no HR report has ever answered well: 

• What skills do we actually have, as opposed to what our job titles claim? 

• Which skills are we one course away from? 

• Who is the internal candidate we keep overlooking? 

• What really predicts success and retention in our roles? We hold a patent in this area, and we keep finding that the university and the brand-name employer predict far less than people assume, while demonstrated skill predicts far more. 

• Which roles should we stop hiring for entirely, because the work has changed? 

Four deployments make this concrete, and each has a direct lesson for a corporate talent strategy. 

A national jobs portal in Singapore matches citizens to work on a skills taxonomy rather than on job titles. As a result, a citizen sees roles they would never have typed into a search box, because the title is unfamiliar even though the work is largely the same as what they did last year. No new jobs were created, but every citizen’s visible labour market widened overnight. Now picture that logic applied to your internal job board. Your employees would see roles they never knew they were qualified for. It is one of the cheapest retention interventions available. 

Universities use the same machinery for admissions and for internship matching, and the second case introduces a distinction worth carrying into any workforce strategy. We match on capability (can this person do the work) and on compatibility (will they thrive with this team, this manager, this working culture). Most systems only ask the first. Placements fail on the second. A capable hire in the wrong environment leaves in six weeks and tells forty friends, 

A national scholarship body uses assessment technology for one of the most consequential selections a country makes. Nobody lets software choose a national scholar. What the technology does is make the human judgement better: the same structured evidence on every candidate, scoring that holds steady across different panels on different days, and a complete record of how each conclusion was reached. The same approach fits your own highest-stakes selections, from graduate programmes to leadership succession. AI ranks and documents; people decide. 

And one of Singapore’s largest employers hires frontline staff with no resume at all. Candidates are assessed on what the job actually requires, while agents handle the integration, the interviews and the scheduling, so recruiters spend their time on people rather than on the queue. Consider who that opens the door to: the five-year gap, the badly formatted single page, the returning parent, the older worker. None of them were ever going to win a CV competition. All of them can demonstrate they can do the job. 

One warning, and it is the dull one that decides everything. None of this works without a shared data spine. If every business unit buys its own tool with its own vocabulary, you end up with twelve dashboards that cannot talk to each other, and in four years someone commissions an integration project that fails. One skills language across the enterprise is an infrastructure decision, not an HR decision, and it belongs at the same level as the cloud decision. 

And one metric change costs nothing: stop reporting time to hire as your headline number and report quality of hire instead. That means people who are still here, performing, and promoted at twelve and twenty-four months. Time to hire measures your process. Quality of hire measures your judgement. 

What cannot be delegated 

Six decisions stay with the leadership, whatever the technology does. 

1. Purpose. The same platform can be tuned for speed, cost, access or long-term capability. Those pull in different directions, and if you do not state the purpose, the vendor’s default becomes your policy. I have watched that happen without anyone in the room deciding it. 
2. Where the human sits. Not as a principle but as a process document naming the step, the person and what they must examine. Principles do not survive a busy quarter. 
3. What data you refuse to use. Some signals are predictive and still unacceptable. Postcode predicts outcomes. So does the school someone attended at fifteen. That a variable improves the model is not a reason to use it. This is a values decision, not a data science one. 
4. Who is accountable, by name. Every model and every agent in production has a human owner. At some point a person will be turned down, and they will write to you, post about it, or call a lawyer. 
5. The route of appeal. How does someone contest a decision made with AI, who reviews it, and how fast? Build it on day one. Candidate and employee trust in these systems will be set by how the first hundred complaints are handled, not by the accuracy figures in the vendor pitch. 
6. Who you invest in. Every hour the machine returns can go to cost savings or to building people. Only one of those compounds. 

There is a seventh, unofficial one: use the tools yourself. Not a demo, not a briefing, but an hour a week, hands on. I meet senior executives making enormous decisions about systems they have never personally touched. You would never approve an acquisition without walking the site. 

And tell your workforce the truth. If roles are going to change, say so early and say what you will do for people. In my own company, the moment we became honest about which tasks the agents would take, resistance dropped. The fear was never the technology. The fear was being surprised. 

My bet for 2030 

These are my predictions, offered as predictions rather than forecasts. 

1. The application becomes a demonstration. For high-volume and entry-level roles the CV disappears, replaced by twenty minutes of the actual work. 
2. Individuals carry a verified, portable record of their skills. The live question is who sets that standard, because if employers and governments do not, vendors will, and your workforce data ends up in someone else’s format. 
3. Both sides bring agents, which forces employers to publish criteria a machine can read. Internal mobility overtakes external recruitment as the primary use of the technology, particularly in large enterprises. 
4. Independent audit becomes a licence to operate, expected by boards, regulators and enterprise buyers the way financial audit is expected now. 

Underneath all of the above sits the hypothesis I would defend hardest. By 2030, candidates will be abundant and the technology will be a commodity, because everyone will have the same models. What will be scarce is employers that people trust to judge them fairly. That is not a technology advantage. It is a governance advantage, and it is available right now to any company willing to build it. 

So my ask of any leader reading this is deliberately small. Pick one process where people are queuing. Put an agent on it. Decide the purpose before you start, keep a person on the rejection, check the results by group every month, and put your own name on the outcome. Then do the next one. 

The technology will keep improving without your help. Trust will not. 

Nina Alag Suri, CEO & Founder X0PA AI