AI in HR and customer support at SMEs: where it pays off (and where it becomes a legal risk)

Human resources and customer support are the two departments where AI touches people directly: candidates to assess on one side, customers to serve on the other. It's also the ground where the promise is strongest — less repetitive work, faster answers, a leaner team — and where, for the same reason, a mistake costs the most. A screening that unfairly rejects a candidate or a chatbot that gives a customer wrong information aren't just communication slips: they are, today, matters of legal liability. That's why we treat the two departments together: they share the same promise and the same risk bar, higher than any other department we've analysed so far.

The serious question, as always, isn't "does the tool exist?" but what is worth automating in a company with a small team — and, in these two departments in particular, what responsibility you take on when you do. We've gathered the 2026 field evidence and try to answer in plain language.

HR: where AI pays off, and why the SME is precisely the underserved segment

In human resources the use cases that deliver a real return follow the same pattern as the other departments — high-volume, repetitive activities where AI removes the manual work but doesn't take the final decision:

• CV screening and candidate sourcing — in place of manual sifting and keyword search. It's the single most adopted use case in recruitment.
• Onboarding automation — forms, calendars and role instructions generated in place of manual administrative work, with a marked cut in the intake load.
• Attrition prediction — signs of disengagement caught before they become resignations, in place of the exit interview when the person has already gone.
• Engagement analysis — automatic reading of surveys and internal communications in place of manual review of the questionnaires.
• Policy and benefits chatbot — self-service on leave, payslips and time off in place of tickets to the HR office for routine questions.
• Performance-review drafts — a synthesis of collected feedback into a first draft, which the manager reviews and makes their own.

Here's a figure that concerns SMEs specifically. AI adoption in HR isn't distributed by maturity, but by size: six in ten companies above five thousand employees use it, against a little more than one in three among small firms. Translated: the least-served segment today is exactly where an SME lives. It's not a lag to be made up in a rush — it's a space to start from with method while the big players are still tangled in the complexity of their own systems.

Customer support: the department where AI has the fastest return

Customer support is, of all of them, the department where AI reaches value first — because much of the work is made of repetitive requests with already-known answers. The cases that work:

• First-tier deflection — a conversational agent that closes simple requests on its own. On average, mature programmes absorb about 40% of first-tier contacts, and on the most schematic cases (password reset, order status, standard refunds) it goes well beyond. On nuanced complaints, instead, the rate collapses — and rightly so.
• Ticket triage and routing — automatic classification of the queue, with first-response times that shorten sharply.
• Agent-assist copilot — AI suggests the reply and retrieves the context while the agent talks to the customer, in place of manually hunting through notes and scripts.
• Automatic answers from the knowledge base — frequently asked questions served from the documentation, with first-response times going from minutes to seconds.
• Sentiment analysis — tone recognition to trigger escalation in place of manual sampling of calls.

The economic return here is tangible: the cost per request resolved with AI is orders of magnitude lower than that of a human agent. But the number we always give clients is another, the one on satisfaction: the perceived quality of a purely automated service stays slightly below the human one, and the gap closes almost entirely only when there's a clean handover to a human for the cases AI shouldn't handle. The operational lesson is clear: AI in support pays off when it's designed to know when to step aside, not when it tries to answer everything.

The aggregate numbers, read honestly

Between the two departments there's a maturity gap worth saying out loud. Support wins fast: it's one of the few areas where the first-year return is documented and strong, and it grows as the knowledge base fills out. HR is more uncertain: there's no equally clean percentage return, and the data mostly tell of a gap between ambition and measurement — nearly half of the HR leaders who invest in AI haven't yet defined how to measure its productivity. They invest driven by expectation, not by a number.

A warning we always give: much of these figures are self-reported by vendors and analysts, not independently verified. They should be read as direction, not as numbers to put in a business plan. And the overall picture calls for caution: independent analyses agree that only a minority of AI projects genuinely hold the promised return, and that the large majority of projects produce no measurable impact on the accounts. Anyone who cites you only the good numbers is selling, not advising.

The practical reading is the usual one: start from the department where the return is fastest and most verifiable — in fact, support — and use that measured result to justify the next step. In HR it's worth moving, but with a success criterion decided before the tool, otherwise the project dies from confused objectives.

The risk that makes HR and support unlike any other department

Here lies the real reason we treat these two departments with more caution than all the others. In sales or marketing an AI mistake is a clumsy email; in administration it's a payment to review. In HR and in support a mistake touches a person — a candidate, an employee, a customer — and the law knows it.

On support there's already a precedent. A court held a company liable for the wrong information its chatbot gave a customer, rejecting the defence that "the chatbot is a separate entity". The lesson is definitive: you can't offload onto an AI the responsibility for what it says to your customers. And the risk isn't theoretical: a chatbot left free to answer "invents" a far from negligible share of its replies, a share that collapses near zero only when it's rigorously grounded in company sources — your documentation, your policies — rather than in the model's generic knowledge. That's why a serious support-AI project isn't "change the model": it's building around it an architecture that keeps it anchored to what the company officially states.

On HR the deadline is regulatory. The EU AI Act classifies as high risk AI systems used for recruitment, promotion, termination and worker monitoring. For these applications the full obligations have been postponed, but it's a window, not an exemption: whoever introduces automated screening today without thinking about it will find themselves unprepared when the deadline arrives. And the underlying problem is documented, not hypothetical: screening algorithms have shown they can discriminate systematically against entire categories of candidates. The practical consequence for an SME is simple: any use of AI in recruitment must be accompanied from the outset by risk assessment, bias testing and human oversight — not as a formal box-tick, but because it's the control that makes the choice defensible.

The quickest way to tell whether a vendor is serious, in these two departments, is to reverse the order of the questions. Not "how much will it save me?", but first "how do I prove this decision is right and that I'm the one who answers for this reply?". If the tool has no clean answer on grounding in the sources (support) and on human oversight and traceability (HR), the return doesn't matter: it's a risk waiting to present itself.

Augmented copilot or autonomous agent: here the choice weighs more than elsewhere

As in the other departments, tools split into two families. The augmented copilot stays an assistant: it proposes the reply to the customer, drafts the review, prepares the screening — but the human approves and remains accountable. The autonomous agent aims to close the process on its own. In HR and in support, though, "closing the process on its own" means talking directly to a person or deciding about them: exactly the point where legal risk is highest.

For an SME the recommendation is clear-cut: start with the augmented copilot. Have AI handle autonomously only the simplest, most reversible requests — where a mistake is corrected without harm — and keep the person in the process for everything that touches a decision about the employee or a sensitive reply to the customer. Autonomy isn't a goal to chase: it's a permission to grant case by case, where the outcome is verifiable.

And in Italy?

The Italian picture is that of a market still behind but accelerating fast: the share of SMEs using at least one AI technology has grown a lot over the past year, but stays low in absolute terms, and the distance from large enterprises on "structured" adoption is marked. It's honest to add a limitation: on the HR and support departments specifically, granular Italian data are scarce — much of the evidence on return comes from English-speaking markets. We don't hide it, because it's exactly the kind of detail that distinguishes an analysis from a slogan. The advantage, for an Italian SME, is that the market's lag is also a space: few competitors have yet set up these two departments well.

Where to start, in practice

If HR or support is the department you want to begin with, the sensible path is short and ordered:

• In support, start with deflecting simple cases — routine requests and already-known answers deliver the fastest return and the lowest risk.
• Anchor the chatbot to your sources from day one — your documentation and your policies, not the model's generic knowledge. It's what keeps the answers faithful and the liability manageable.
• In HR, treat automated screening as high risk — human oversight, bias testing and decision traceability from the start, not as a later addition.
• Prefer augmentation to autonomy — a copilot that makes the team faster, with the human approving whatever touches a person.
• Define the success criterion before the tool — a number that says whether it's working. Without one, the project dies from confused objectives, not from the limits of AI.

Even before choosing the department, though, it's worth knowing where you are: our AI-readiness assessment helps you understand where to start with more return and less friction, and which controls to put around the first workflow. And since in HR and in support compliance isn't a detail but the pivot, our compliance overlay explains how we wire the controls — grounding in the sources, human oversight, traceability — to every design.

We've turned the first step into a self-serve, free assessment: a few questions and an indication of where to start, with what controls around it. Take the AI-readiness assessment — then, if it makes sense, let's talk.

This article is for orientation and does not constitute legal advice. The return, adoption and deflection figures cited come from market analyses and from largely self-reported industry sources, not independently verified: they should be read as indications of direction and not as guarantees of results. Every use of AI in staff recruitment or in customer support must be assessed against the controls, the data and the context of the individual company, and — for recruitment systems — in light of the applicable regulatory obligations.