From CV screening to redundancy decisions, AI is increasingly shaping who gets hired, managed and let go. But having a manager review the technology’s output isn’t necessarily the same as exercising human judgment. In the latest episode of “Skadbytes,” Deborah Kirk and Jonathan Stephenson from Skadden’s IP and technology team in London are joined by labor and employment attorneys Damian Babic and Jacob Alston to examine AI in the workplace through a twin lens: employment law and data protection. They discuss the U.K.’s new automated decision-making rules, the risk of indirect discrimination at scale and why employers must be able to show not just what the technology did, but what they did with it.
Episode Summary
When employers use AI to hire, manage and dismiss staff, complying with data protection law doesn't guarantee a fair employment decision. Host Deborah Kirk and counsel Jonathan Stephenson from the IP and Technology team welcome counsel Damian Babic and associate Jacob Alston from Skadden's London labor and employment practice. Following AI through recruitment, performance management and dismissal, the lawyers examine each stage under both data protection and employment law. Jonathan explains how the Data (Use and Access) Act 2025 changed U.K. automated decision-making rules, while Damian discusses how AI can apply indirectly discriminatory criteria at scale. They also explore why adding a human isn't the same as adding human judgment and why employers may need two audit trails.
Voiceover (00:00):
Welcome to “SkadBytes,” a podcast from Skadden exploring the latest developments shaping today’s rapidly evolving tech landscape. Join host Deborah Kirk and colleagues as they deliver concise insights on the pressing regulatory issues that matter to tech businesses, investors, and industry leaders worldwide.
Deborah Kirk (00:23):
Hi everyone, and welcome back to “SkadBytes.” I’m Debs Kirk, and I’m joined today by Jonathan Stephenson, who is a counsel in our IP and Tech team here at Skadden in London. And for the first time on “SkadBytes,” we are joined by two of our colleagues from the London Labor and Employment team, Damian Babic, counsel, and Jacob Alston, an associate in the team. Damian, Jacob, welcome and thank you for joining us.
Damian Babic (00:46):
Thanks, Deb. Great to be here.
Jacob Alston (00:47):
Thanks, Debs.
Deborah Kirk (00:49):
Now, this is an episode we’ve wanted to do for a while because AI in the workplace is a really good example of an area where our teams increasingly find themselves looking at the same issue, but from different angles or at least different legal angles. So there is a twin lens when we advise on employee data where each use case needs to be evaluated both from a data protection perspective, so that’s our world, and through the employment lens, which is these guys’ world. Now, AI is quickly making that same joined-up approach essential in many areas, including these. Employers are increasingly using AI to screen CVs, support performance management, monitor employees, inform pay and promotion decisions, and even assist with disciplinary processes and redundancy selection. I’d say the appeal is obvious, but the same tool can simultaneously raise employment, discrimination, privacy, automated decision-making, and AI regulatory issues.
Jonathan Stephenson (01:55):
Yeah, exactly that. And I think we find as these tools become more autonomous, the through line is a risk of less human oversight unless employers consciously build it back in. There’s also an important theme that I think runs through today’s discussion, and that’s a green light under one legal regime is not necessarily a green light under another. Data protection law may permit particular processing or following recent UK reforms, give organizations greater flexibility around automated decision making, but that doesn’t really answer whether the employment decision that comes out the other end is fair, non-discriminatory, and defensible.
Damian Babic (02:33):
Yeah, that’s right, Jonathan. And the reverse is true too. An employer may have a perfectly understandable employment objective, but that does not in itself address whether the data being used to achieve it has been collected and processed appropriately or whether the level of automation is acceptable.
Deborah Kirk (02:50):
Right. So rather than work through the legislation in the abstract, what we’re going to do here is follow AI through three stages of the employment life cycle. So who gets hired, how employees are managed, and how disciplinary dismissal and redundancy decisions are made. And what is interesting is that the legal considerations change as you move through that life cycle. So in recruitment, the concern may be who gets through the door and how to manage complaints from applicants who are rejected. In monitoring and performance, it becomes how people are measured and rewarded in a non-discriminatory way. And at the dismissal stage, the question is whether the employer has preserved a genuinely fair decision-making process and can demonstrate a fair reason for dismissal. But before we get into these scenarios, Jonathan, it is probably worth briefly explaining what has changed on automated decision making in the UK?
Jonathan Stephenson (03:49):
Yeah, the Data (Use and Access) Act 2025 or DUAA changed the UK framework from the 5th of February of this year. Previously, there was a fairly restrictive gateway around decisions based solely on automated processing, and that produced legal or similarly significant effects. The DUAA gives organizations considerably more flexibility. Outside the stricter regime for certain decisions involving special category data, organizations can use the ordinary range of lawful bases, including legitimate interests. But where a significant decision is based solely on automated processing, and by that we mean there is no meaningful human involvement, safeguards still apply. The individual needs to be informed, given an opportunity to make representations, and to be able to obtain meaningful human intervention and to be able to really contest the decision. So the gateway has widened, but that doesn’t necessarily mean that employers can simply hand consequential workplace decisions over to a machine.
Deborah Kirk (04:54):
Absolutely. And that distinction between having a human somewhere in the process and having meaningful human judgment is going to come up repeatedly. So Damian, before we get into the scenarios, there is one other concept we probably need to explain: direct and indirect discrimination, so help us here.
Damian Babic (05:12):
Yeah, sure, Debs. Jacob, do you want to take that one?
Jacob Alston (05:14):
Of course. So direct discrimination is where someone is treated less favorably because of a protected characteristic such as race, sex, or disability. Indirect discrimination is different. It applies where a rule, criterion or practice which applies to everyone in the same way and looks neutral on its face may still put people sharing a protected characteristic at a particular disadvantage. Subject to limited exceptions, direct discrimination generally cannot be justified. Indirect discrimination can be justified where the employer can show that the practice is a proportionate means of achieving a legitimate aim.
Damian Babic (05:49):
Yeah, that’s right. And indirect discrimination is particularly relevant to AI because AI can apply the same criterion consistently and at enormous scale. Now, consistency sounds attractive, right? But if the criterion itself disproportionately disadvantages a particular group with or without a certain characteristic, automation can replicate that disadvantage across hundreds or thousands of decisions really quickly. And that’s the real issue here. The consistency that makes these systems attractive can also increase the risk that criteria that is indirectly discriminatory is applied more widely.
Deborah Kirk (06:24):
It’s so interesting to hear that in the way that it applies to your world. So let’s make this a little more concrete and start at the beginning of the employment relationship: recruitment. So Jonathan, why don’t you tell us what might the technology actually be doing at this point?
Jonathan Stephenson (06:39):
Yeah, good question. So the way I think about it is, imagine a large employer using an AI recruitment platform. The system reviews CVs, it runs automated skills assessments, ranks candidates for role fit, and it might use a chatbot for first round screening. There are obvious potential benefits here: speed, consistency, and the possibility of reducing some forms of human bias. And the ICO itself, the UK supervisory authority for data protection, has recognized these potential benefits. There is another interesting dynamic here too though. One survey found that almost half of the UK job seekers were using generative AI in their applications. So in some recruitment processes, we are effectively seeing AI-assisted screening of AI-assisted applications. But the question for the employer is what its system is actually doing underneath that.
Damian Babic (07:31):
Yeah, that’s right, Jonathan. And if the model has been trained on, say, historic recruitment data, it may learn patterns from the people the organization hired successfully in the past. So, for example, if the historic workforce was not particularly diverse, the model may reproduce those demographic patterns, applying historical biases, even though nobody instructed it to actually do that. So you can therefore have a system applying apparently neutral criteria consistently, but while producing results which are indirectly discriminatory at scale before any human has even seen the candidate’s name or their application. And systems can miss context as well that a human decision-maker might understand. So a gap in someone’s employment history might be treated as a negative signal, for example, even though it relates to disability, family leave or another circumstance that needs really to be understood in the round in light of an employer’s obligations around reasonable adjustments, for example, rather than simply scored negatively.
Jacob Alston (08:28):
And when a job applicant is rejected, they might well ask why they were unsuccessful. And if they’ve been treated differently to someone without their protected characteristic, then there’s a discrimination claim potentially, and the employer will need to explain why its treatment was not discriminatory. And at that stage, if the employer’s saying, well, the algorithm gave them a lower score, that on its own is not much of an explanation.
Deborah Kirk (08:49):
Understandably. And the ICO also has been considering the use of automated decision-making in recruitment in its Recruitment Rewired work. It engaged with more than 30 employers and found that some organizations didn’t even fully recognize when they were actually using solely automated significant decision making. In other words, they thought the technology was supporting a human decision when in reality the system was effectively making the decision. The ICO also identified problems around privacy of information, fairness and bias assessments and DPIA, so data protection impact assessments, which are essentially the pre-launch assessment of how use of personal data could affect individuals and how those risks are going to be managed. The regulatory message here is not do not use these tools. The potential benefits are recognized. The issue is really ensuring that the employers understand what the system is actually doing and put the appropriate safeguards around it, particularly when we’re dealing with automated decision making.
Jonathan Stephenson (09:56):
Yeah. And looking through a different lens, we’re now seeing the responsibility question play out in litigation too. In an ongoing US class action concerning an AI recruitment platform, applicants allege that automated screening disproportionately excluded candidates based on characteristics including race, age, and disability. And the litigation has now reached the class certification stage. One of the particularly interesting questions is where responsibility sits when an employer relies on a third-party system, whether it’s with their employer, technology provider, or potentially both of them.
Damian Babic (10:32):
That actually really matters from an employment perspective because in the UK, the employer will still ultimately need to explain why any particular candidate was treated as they were.
Jonathan Stephenson (10:41):
Exactly. So this then becomes a procurement and a governance issue too. An employer can’t necessarily assume that buying the technology means outsourcing the risk. If you can’t sufficiently understand how the tool reaches its recommendations, test it appropriately or obtain enough information from the provider to defend a challenged decision, that can become a problem very quickly. And where candidate data is processed outside the UK and the EU to bring about privacy, there is another layer. The employer needs to understand where that data is going and whether the necessary international transfer arrangements are in place.
Deborah Kirk (11:17):
That’s right. And there is of course, also an EU AI Act angle to all of this. Recruitment AI is one of the employment uses treated as high risk under the AI Act regime. The full high risk requirements for these standalone systems have now been pushed to December 2027 in terms of their enforcement dates, but that additional runway doesn’t create a regulatory vacuum. Employment law and data protection law clearly apply today. So an employer should not read the delayed AI Act date as permission to defer governance around a recruitment tool that it is already using. There’s just another thing on their list to think about in December of next year.
Jacob Alston (11:57):
I think, Deb, just coming back to the point on human involvement and the question is not simply whether a person appears somewhere in the process, the person needs to understand the recommendation, be able to interrogate it, consider information that the model may not have captured, and have the genuine authority to override an AI tool decision and make their own final decision.
Jonathan Stephenson (12:17):
Yeah, exactly that. And adding a human to the workflow is not necessarily the same thing as adding human judgment.
Deborah Kirk (12:22):
Right. And once a candidate is through the door, the same questions don’t then disappear. They just simply change form. So imagine an employer using continuous performance tools to score every, say, customer interaction or analyze communications, track sales, and feed those metrics into performance reviews or pay or promotion even decisions. There is of course an obvious upside here too. Used well, that kind of data can give managers a much broader and potentially more consistent picture of performance than periodic subjective reviews. And at a first glance, that can look like objective management information. So Jacob, tell us where does the employment risks start to appear?
Jacob Alston (13:05):
So I think, Debs, one obvious issue is about what the metrics fail to capture. So take your example of activity-based productivity scoring. An employee on maternity or other family leave or with periods of sickness absence due to a long-term health condition may generate less activity or a part-time worker may produce fewer outputs in absolute terms. A disabled employee may have reasonable adjustments in place, meaning changes that the employer has made to remove disadvantages arising from a disability. For example, that could be a change to working hours. So if the system simply treats lower activity as lower performance, the same supposedly neutral metric can systematically disadvantage those employees. And that potentially is an indirect discrimination risk and for disabled employees, possibly also a breach of the duty to make reasonable adjustments.
Damian Babic (13:50):
Yeah, that’s right, Jacob. And just looking at monitoring beyond the discrimination angle, every employment contract, whether or not it’s written down, includes an implied duty of trust and confidence. Now broadly, that means that an employer must not behave in a way that is likely to destroy or seriously damage the employment relationship. Now, pervasive monitoring that doesn’t account for individual circumstances can engage that duty. Now in a sufficiently serious case, a breach of trust and confidence can form the basis of a constructive dismissal claim where the employee is forced to resign in response to a fundamental breach of their employment contract by their employer. So whether a monitoring tool is lawful to deploy from a data protection standpoint is only really part of the analysis.
Jonathan Stephenson (14:33):
And that’s a really good example of the twin lens we discussed earlier. An employer might have a lawful basis under data protection law for processing performance data, for example. But if an automated pay or promotion model trained on historic data systematically disadvantages say part-time workers or a particular ethnic group, having a lawful basis to process the data doesn’t insulate the resulting employment decision from discrimination risk.
Deborah Kirk (14:59):
That’s right. And monitoring is becoming a live policy issue in its own right. So the government is looking specifically at workplace monitoring technologies, including from a transparency point of view, thinking about the worker voice and the implications for the wider employment relationship, which reinforces the point that Damian just made. The question is not simply are we legally permitted to collect this data? It is also how the technology is introduced, what employees sort of understand about it and what it does to that employment relationship.
Jacob Alston (15:33):
And it’s also why consultation and communication matter. Where an employer is introducing technology that materially changes how employees are monitored or assessed, involving employees and where relevant their representatives can be important both legally and practically for good employee relations. If employees understand what is being measured, why it is being measured, and crucially how they can challenge an outcome that appears wrong, this is likely to reduce the risk of friction as these technologies become more widely used.
Jonathan Stephenson (16:01):
And from the governance side, the same human judgment point appears again. A manager who receives an AI-generated performance score needs to be able to interrogate it, consider the employee’s individual circumstances and depart from it where appropriate. A human clicking approve at the end of an automated process doesn’t necessarily give you that meaningful human involvement.
Deborah Kirk (16:23):
Which takes us to the final stage of the life cycle, disciplinary actions, dismissals and redundancy. So Jacob, maybe I could ask you to sort of set up this scenario for us.
Jacob Alston (16:36):
Sure, Debs. So I think picture a system analyzing access logs, communications, call recordings, performance history, and previous comparable disciplinary cases, which then produces a determination along the lines of gross misconduct established, dismissal recommended, complete with a confidence score and a timestamped evidential trail. Or take a redundancy selection. An AI system scores employees who are at risk of redundancy using activity and performance data and people who generated less data, perhaps because they were on protected leave or working reduced hours end up with disproportionately lower scores. The attraction for an employer is understandable. It gives you greater consistency across managers and sites, potentially less individual managerial bias in the process and gives you what looks like a solid evidential record.
Damian Babic (17:25):
And that’s a really important point is that a fair employment process though requires more than just arriving at a consistent answer. And ultimately an employer defending an unfair dismissal claim needs to show that its decision fell within a range of reasonable responses. Now in a misconduct case, the employer needs to investigate properly, give the employee a genuine opportunity to explain their position, and also have it considered. In a redundancy, the selection process needs to be objective. The employee needs to have a meaningful opportunity to challenge that decision. And in either case, there ultimately needs to be somebody who exercises judgment. Now, are there mitigating circumstances, for example? Is the outcome proportionate for this employee in this circumstance? Is there a reason why this case might need to be treated differently from comparable cases in the past? Now, if the AI has effectively reached that answer before that decision-making process has taken place, I think the employee has an obvious argument that the consultation or hearing was never genuinely capable of changing the outcome.
Jacob Alston (18:23):
If the process is unfair, that becomes particularly important as the unfair dismissal regime in the UK changes. From January of next year, the ordinary qualifying period of service for unfair dismissal reduces to six months and the cap on compensatory awards is removed. So more employees will potentially be able to challenge dismissal decisions and the financial exposure attached to an unfair dismissal claim will increase.
Deborah Kirk (18:47):
So there is a slightly counterintuitive governance point here. We normally tell clients that good AI governance means having a strong audit trail and that is right, but an audit trail only helps if what it records is a defensible process.
Jonathan Stephenson (19:02):
Yeah, exactly. And maybe the answer to that is that there are really two audit trails.
Deborah Kirk (19:06):
Right.
Jonathan Stephenson (19:07):
You need the technical record showing what the system did, what information went in, what output came out, and to the extent possible, how the recommendation was reached. But you also may need the human decision trail showing what the employer did with that information. Did the manager interrogate it? Did they consider information the model had missed? Did they consider the employee’s explanation or proposals? Did they depart from an AI-generated recommendation where it was appropriate to do so? And why did they ultimately reach the decision they did?
Damian Babic (19:41):
Yeah, and from an employment perspective, that second record may ultimately matter more in the end. A perfect technical record showing that an algorithm reached a conclusion and the manager simply accepted it could actually make the employer’s decision harder rather than easier to defend. The question is not just whether you can reconstruct what the machine did. I think it’s more whether you can show what the employer did with its output.
Jonathan Stephenson (20:04):
There’s also thinking about it, a confidentiality point here. Using external generative AI tools in disciplinary or investigation work can involve highly sensitive employee information, confidential business information, and potentially legally privileged material. And where an internal investigation is legally privileged, putting that material into an external AI tool can also create questions about whether that privilege is being preserved. So organizations need clear controls around what can and cannot be put into external AI systems, which tools are approved and how investigation material is handled.
Jacob Alston (20:42):
And Jonathan, on the employment side, there’s a similar consideration actually in relation to the duty of confidentiality that an employee would expect from their employer, which goes to that general breach of trust and confidence point. There’s also a scale thing to consider here. Acas, an independent advisory body, estimates that UK organizations already deal with around 1.7 million formal disciplinary cases and 375,000 odd formal grievances each year, at an estimated cost of more than two billion pounds. So even before AI changes the way that employees formulate their complaints, businesses are dealing with a very significant volume of workplace conflict. And it’s not just employers using these tools. Employees have fast and easy access to powerful generative AI too. So an employee can take a relatively short description of a workplace issue and use AI to turn it into a lengthy,legally framed grievance within seconds. It’s not that an AI-assisted grievance is less legitimate, but AI can increase the volume, the length, and the legal framing of workplace complaints, and employers need processes in place which are capable of identifying the underlying issues and responding to them properly.
Damian Babic (21:49):
So I think where we are is that AI governance in the workplace really increasingly runs in both directions. Employers need rules for the technology they use to make or support decisions about employees, but also sensible policies governing employees’ use of external AI tools for work. Now that includes protecting confidential information and trade secrets, setting expectations around approved tools, and making clear how AI may be used in work product or employment processes.
Deborah Kirk (22:17):
And that very neatly takes us back, I think, to where we started. The employment and technology questions can’t really be separated here. The fact that usage of data or automation is permitted does not make the resulting employment decision fair, of course. Equally, a sensible employment objective does not remove the need to think about the data, the technology, and the regulatory framework supporting it. So before we wrap up, team, why don’t we pull out the practical points here? Damian, over to you.
Damian Babic (22:50):
Sure. So look, I think the first one is genuine human decision making. For consequential employment decisions, a qualified manager really needs to actively review the AI recommendation, apply independent judgment, and make a documented decision setting out their own reasoning. Now, the aim I don’t think is to create a human authored memo that simply repeats the machine’s output. It’s to be able to demonstrate what factors the decision maker considered, what the individual circumstances were that really mattered, and why the ultimate decision was reached.
Jonathan Stephenson (23:19):
Happy to take the second, and I think that is do the governance work before the tool goes live. That means understanding the use case and lawful basis, completing a DPIA or other assessment where required, understanding what data the tool actually needs, testing for bias and accuracy, and where relevant, understanding whether historic training data may embed patterns that the tool could reproduce. For third party tools, businesses should also think about what information and audit rights they need from the provider, how problems will be escalated, and whether they have enough visibility to defend the decisions made with that tool.
Jacob Alston (23:58):
I think the third point is making sure that employees have a meaningful route to challenge AI-assisted decisions. So that means being transparent about how relevant performance or monitoring metrics are used and making sure that an employee can raise context that the system may have missed. Importantly, that challenge process needs to be capable of changing the outcome, otherwise it’s human involvement in name only.
Damian Babic (24:19):
I think finally, govern AI in both directions. So policies should really address how the employer uses AI in employment decisions, but also how employees may use external AI tools in their work. That includes clear rules around confidential information, trade secrets, and sensitive employee data, as well as sensible guidance on when AI-generated material might need human checking.
Jonathan Stephenson (24:42):
Yeah. And for me, the common thread across all of this is really simple, and I’ll say what I said before, and that’s adding a human to the workflow is not the same thing as adding human judgment.
Deborah Kirk (24:52):
And I guess also a good audit trail only really helps if it records a good decision-making process, including showing that the human element truly was adding human judgment and not a tick box to show that a human was involved. So AI can make employment decisions faster and more consistent, and there are clearly real opportunities here, but the technology does not remove the need to understand individual circumstances, exercise judgment, and of course be able to explain the outcome and how you got there. So that is why the twin lens and our two worlds collaborating on this matters so much.
(25:33):
Employment law tells you whether the workplace decision is fair and defensible, data protection and AI regulation regulate whether the data and automation supporting that decision are being used lawfully and appropriately. And increasingly, of course, those are not separate exercises. They’re just two legal pillars encompassed in one question that our clients want answers to. How the system is designed affects the employment decision and how the employer uses the output affects the regulatory analysis. And that’s why these issues, like so many, need to be looked at together. Damian, Jacob, thanks so much for joining us.
Jacob Alston (26:10):
Thanks, Debs. Great to be here.
Damian Babic (26:11):
Thanks, Debs.
Deborah Kirk (26:14):
And if any of the themes we have discussed touch on what you are working through within your organizations, we would, as always, be pleased to continue the conversation. We’re very easy to find, but for now, thanks for listening to “SkadBytes,” and we will see you next time.
Voiceover (26:29):
Thank you for joining us for today’s episode of “SkadBytes.” If you like what you’re hearing, be sure to subscribe in your favorite podcast app so you don’t miss any future conversations. Additional information about Skadden can be found at Skadden.com.
Listen here or subscribe via Apple Podcasts, Spotify, YouTube or anywhere else you listen to podcasts.
See all episodes of SkadBytes: Tech Innovation Meets Regulation


