Challenges of implementing AI in HR (and how to overcome them)

Your Slack channel is full of the same ten policy questions, and your exec team just asked why you haven't "turned on AI" yet. You know the honest answer. A generic chatbot dropped on top of your HR stack will quote a parental leave policy from 2022 or surface a salary band to someone who shouldn't see it.

So you hesitate. That's reasonable. Moving from experiments to real HR service delivery means getting past six obstacles that have nothing to do with which model you pick: undocumented processes, legacy integrations, employee trust, skills gaps, governance and ROI proof.

None of them are fatal. But each one has stalled a rollout somewhere, and most of them stall it before anyone notices.

What are the biggest challenges of implementing AI in HR?

The biggest challenges of implementing AI in HR are messy source data and undocumented processes, integration with legacy systems, employee and HR-team resistance, missing skills, governance that can't keep pace, and the difficulty of proving a return.

Here's the short version. The rest of this piece takes each one in turn.

Challenge 1: your data and processes aren't written down

The first obstacle isn't the model. It's that the AI can only work from what you've documented, and most People teams have never had to document at that level.

Your PTO policy is in Confluence. The updated version is in a Google Doc. The exception for the Berlin office is in a Slack thread from March. An AI agent reading all three will pick one, and it won't be the right one.

Then there are the processes that were never written down at all. How a relocation request actually moves between People Ops, payroll and IT lives in two people's heads. The AI can't automate a process it can't see.

This is the most common blocker, and not just in HR. In SSON Research & Analytics' 2025 survey of more than 350 shared services executives, the organizations adopting generative AI named data management (59%) and skills gaps (56%) as their main implementation challenges. Not the technology. The inputs.

How to get past it

Don't try to clean everything. Pull your last 90 days of HR requests and find the twenty topics that make up most of the volume. Those are the documents that need to be right before launch.

For each one, name an owner and a review trigger. When the policy changes, the document changes the same week. If your documentation is messy, your automation will be messy too.

For undocumented workflows, write down the top five as a simple sequence: who does what, in which system, and what "done" looks like. You'll need this anyway. It's the difference between an agent that answers a question and one that completes a request.

Challenge 2: integration and legacy systems

The second obstacle is that an AI tool with no connection to your systems can only talk. It can't act.

There's a real difference between a wrapper and an integrated agent. A wrapper passes text to an external model and reads from static files you uploaded. An integrated agent authenticates the person asking, reads live data from your HRIS, and writes the result back.

Connecting to a legacy HRIS, ATS or ITSM is where this gets hard. You'll find aging APIs, custom fields nobody remembers adding, and data that disagrees between systems. When the AI can't see the payroll record, it doesn't know the employee lives in Texas and quotes California overtime rules instead.

This is also why internal HR AI builds usually fail. Uploading the handbook to a custom GPT takes an afternoon. Wiring it to Workday, Slack and Jira with the right permissions takes a team, and then someone has to maintain it.

How to get past it

Ask every vendor three questions before anything else.

• Pre-built integrations or custom APIs? If connecting to your HRIS means a custom build, your time to value stretches out and your IT team owns a new liability.
• Can it write back? Reading from the HRIS is table stakes. Updating a direct deposit, opening the IT ticket, changing the address: that's where the time savings live.
• Is it tied to one model? The model landscape changes fast. You want a vendor that can switch models underneath without breaking your workflows.

Sports analytics company Hudl is a useful reference point for what a properly integrated agent does. Hudl saw 70% of requests handled by AI without escalation in six months, with 80% of requests meeting first-response SLA targets and onboarding coordination cut from 60 to 20 minutes per hire.

Challenge 3: employee trust and change resistance

The third obstacle is people, on both sides of the request.

Employees first. Trust is easy to lose. If someone asks about their parental leave entitlement and gets last year's policy, they won't ask again. They'll DM your team directly, and now you're running two service desks instead of one.

The HR team is the other half, and it's the half most articles skip. Stanford Digital Economy Lab's 2026 study of 51 enterprise AI deployments found that Legal, HR, Risk and Compliance functions were the most frequent source of resistance, at 35% of cases, ahead of internal end users at 23%.

Read that again. In the deployments that eventually worked, HR was more likely to be the blocker than the people it serves.

Some of that is healthy caution. Some of it is fear that automation makes the role obsolete. Both are worth taking seriously, and neither goes away by being ignored.

How to get past it

Get the first impression right. Launch on the narrow set of topics you've cleaned, not the whole knowledge base. An agent that answers twenty topics correctly beats one that answers two hundred badly.

Make the human path visible. Every answer should carry a one-click route to a person, and when it escalates, the whole thread should travel with it so nobody repeats themselves. Employees who know they can reach a human are far more willing to try the AI first.

Then give the HR team ownership. When People Ops owns the knowledge base, the escalation rules and the metrics, the tool stops being something done to them. It becomes theirs. That's when the shift from admin to impact becomes real, because AI handles the repetitive administrative work and the team gets its week back.

Be honest about roles too. Say out loud what the agent will and won't do, and what the team does with the reclaimed hours. Vagueness reads as a threat.

Challenge 4: the skills gap

The fourth obstacle is that running an AI agent well is a skill nobody on your team was hired for.

It's not coding. Nobody needs to write Python. But someone has to decide what the agent is allowed to answer, review the conversations it gets wrong, keep the knowledge base current, and read an audit log when something goes sideways. That's a job, and in most People teams it currently belongs to nobody.

The SSON figure above puts it at 56% of shared services organizations naming skills gaps as a main challenge, right behind data. Our experience is that the two are the same problem. Bad data persists because nobody owns the tool.

How to get past it

Name one owner. Not a committee. One person in People Ops whose job includes the agent's accuracy, with a few hours a week protected for it.

Train for three things: writing and maintaining policy content the way an agent reads it, reviewing flagged conversations, and understanding what the permission model does and doesn't allow. A day of onboarding covers most of it.

Then lean on your vendor. The right partner runs the model, the integrations and the security review. Your team runs the logic. That's the split that works at 500 to 5,000 employees, where there's no budget for an AI platform team.

Challenge 5: governance that can't keep up

The fifth obstacle is that AI in HR moves faster than the rules around it.

Here's the version of this that's already happening. A manager builds a helpful agent in a no-code tool, shares it with the team, and three months later it's answering benefits questions for half the company with no review, no access control and no log. Cornell's CAHRS working group reported in 2026 that member companies' governance councils were struggling to govern the rapid, democratized creation of AI agents by employees. The fix they described was simple: a formal access request and a production process for any agent shared beyond a local team.

Governance also covers the part most HR leaders worry about most: privacy, bias and wrong answers. We've covered those risks in depth elsewhere. As implementation obstacles, they come down to three controls.

• Permissions the agent inherits from your systems. A manager who can see performance scores in the HRIS can ask about them. An individual contributor can't, and the agent should decline rather than guess.
• An escalation path for anything sensitive. Employee relations, accommodations, grievances and anything emotional don't get an AI answer. They get routed quietly to a human on the People team, with the context attached.
• A log of every interaction: who asked, what the agent read, what it answered, what it did. That's your evidence when someone disputes an answer, and it's what your security team will ask for first.

How to get past it

Decide the tiers before the tool arrives. It's a one-page exercise, and it settles most governance arguments in advance.

Then put AI in workforce operations on the same footing as any other system that touches employee data: a named owner, a security review, and a rule that nothing new goes company-wide without going through the access request.

Challenge 6: proving the return

The sixth obstacle is the one that decides whether your pilot gets a budget line next year. Most teams can't show what the AI changed, because they never measured where they started.

This is common. Insight222's 2025 survey of 372 companies found that 57% of people analytics and senior HR leaders don't measure AI's impact on outcomes like productivity or engagement. If more than half of the function isn't measuring, "we can't prove it" isn't a personal failing. It's the default.

The trap is measuring the wrong thing. "Questions answered" is a vanity metric. If someone has to ask four times to get one right answer, your volume looks great and your employee experience is poor.

How to get past it

Take a baseline before launch. Two weeks of request volume by topic, median time to first response, median time to resolution, and a rough estimate of HR hours per week spent on Tier 0 and Tier 1 requests. You can't show a change without a before.

Then track three things that matter.

• Deflection rate: the share of Tier 0 and 1 requests resolved with no human touch.
• SLA adherence on what does escalate, since the point is that humans now have time for the hard ones.
• Hours back to the team, translated into what those hours went to instead.

For a sense of scale, Kinfolk's own benchmark is up to 80% of Tier 0–1 employee requests handled without a human, and roughly 45 days a year reclaimed for HR teams. Your numbers will be your own. The point is that you'll have them.

Frequently asked questions

What is the hardest part of implementing AI in HR?

The hardest part is usually the data, not the model. AI agents answer from your documentation, and most People teams have policies spread across wikis, drives and Slack threads with no clear owner. Cleaning the twenty highest-volume topics and assigning owners before launch solves more problems than any vendor feature.

Why do HR AI pilots fail?

HR AI pilots fail when the tool can only answer but can't act, when it launches on messy documentation and loses trust in the first week, or when nobody set a baseline so there's nothing to show at the end. The fix for all three happens before launch: clean sources, real integrations, and metrics agreed in advance.

How do you overcome employee resistance to AI in HR?

You overcome employee resistance by getting the first interaction right, making a human always one click away, and being transparent about what the agent does with their data. For the HR team, resistance drops when they own the knowledge base, the escalation rules and the metrics rather than having IT hand them a finished tool.

What skills does an HR team need to run AI?

An HR team needs someone who can maintain policy content the way an agent reads it, review conversations the agent got wrong, and understand what the permission model allows. It's not a technical role. A single owner in People Ops with a few protected hours a week is enough for most teams of 500 to 5,000 employees.

How do you measure the ROI of AI in HR?

You measure ROI by taking a baseline before launch (request volume, time to first response, time to resolution, HR hours on routine requests) and tracking deflection rate, SLA adherence and hours returned to the team afterwards. Questions answered is a vanity metric; requests resolved without a human is the number that holds up in a budget review.

The obstacles are real, and none of them are the model

Every challenge on this list is about the work around the AI: what it reads, what it can touch, who trusts it, who runs it, who governs it and who measures it. That's good news. Those are things you control, and most of them you can fix before you sign anything.

Your team wasn't hired to run a ticket queue. Get the six obstacles out of the way and it doesn't have to.

If you want to see how an AI service desk handles requests inside Slack with your permissions, your escalation rules and a full audit log, we'd be glad to show you.

Book a demo