Not Evil, Not Magic: How HR and IT Can Use AI Safely with Employee Data
Ask the internet whether AI belongs anywhere near employee data and you get two answers. One says it is far too risky. The other says it is fine to connect to everything. Neither answer comes from someone who knows your company, your systems, or your tolerance for risk.
Meanwhile, HR is the team being asked which tools should connect to the company’s AI, often without the vocabulary to discuss it with IT. The real answer sits in the settings: who can get in, what each tool can read, and what gets logged. Most of those decisions already live somewhere in your human resources tech stack.
On November 4, Rebecca Taylor of AllVoices hosted Brandon Ashley, senior director of service desk operations at ePlus, for the third session of a series on who is driving HR in the age of AI agents. The two worked together when Rebecca was in talent acquisition and Ashley was the IT partner readying laptops for new hires on Friday nights. Ashley sees what most security people do not: what employees actually do with new tools, from access requests to can-I-use-this tickets. The series opened with getting HR data out of spreadsheets and continued with using ER data to spot patterns early. Rebecca called this the conversation underneath both.
Rebecca opened with a poll on where each company stands with AI. At the last count shared in the session, 75% of attendees were piloting AI tools and 25% had them fully rolled out. Nobody picked the option for AI that is officially banned and unofficially everywhere, a pattern common enough to have a name: shadow AI.
What is the difference between an enterprise AI account and a personal one?
Control. With an enterprise account, the company decides who gets in, sees what they did, and knows where the data goes. With a personal account, nobody at the company knows any of that.
Ashley started with the personal side, where most employees begin. The problem is not the tool. It is that the account sits outside everything the company can see.
You don’t know where your data is going to be stored, who has access to the prompts and the stuff that you’ve uploaded in the personal accounts. You might be able to know, but you don’t know. And your company definitely doesn’t know. - Brandon Ashley, senior director of service desk operations at ePlus
Upload personal information about employees, or a customer’s credit card details, and the company can be in serious trouble if it leaks. An enterprise account replaces that uncertainty with specific guardrails. IT controls who has access to what, and it can connect the tool to work systems. Ashley’s example was Copilot pulling in Outlook email and the files shared with you to draft a new document, something a personal account will never reach.
Single sign-on is what makes that control practical. When an HR generalist moves into a recruiting role, they still need some of the same data, but not all of it. Security profiles built behind single sign-on let IT adjust access as people move jobs or departments, and pull all of it back when someone leaves. Logging adds the second layer, and Ashley explained why IT cares so much about it.
If I’m working Monday through Friday in New Hampshire, we can see logs on that. If all of a sudden there’s logs at two in the morning, it’s like, okay. Well, that’s a little bit unusual. But if that’s coming from China, okay. We might have a problem here. - Brandon Ashley
In HR terms, you would not want proprietary company information or employee records sitting in someone’s personal email either. A personal AI account is the same risk in a newer format.
Personal AI account vs. enterprise AI account
| Personal account | Enterprise account |
|---|---|
| Nobody at the company knows where prompts and uploads are stored | IT sets guardrails on who can access what |
| Access lasts as long as the person keeps the login | Access runs through single sign-on and is pulled when someone moves or leaves |
| No record the company can review | Logs show who accessed what, and when |
| Reaches nothing inside the company | Can connect to work email, files, and approved tools |
| Fine for learning prompts and brainstorming | Where work that touches company data belongs |
Why do employees use personal AI accounts at work?
Because most people learn AI on their own time, and the habit follows them into work. Ashley pointed out that everything on LinkedIn is about AI right now, so people feel they need to learn it. They watch videos, try inexpensive tools, and start building on what they learn.
It starts out as a learning thing, and then you end up trying to bring it into your work because we’re all so busy, and we can utilize AI. - Brandon Ashley
That path is hard to argue with, which is why banning personal accounts rarely ends the practice. Ashley’s experience is that a flat no with no reason attached sends people around IT rather than away from the tool. A reason for every limit does more than the limit itself.
Rebecca raised a related debate from the creator side of LinkedIn. Should you do your work in a personal AI account so that you can take it with you when you leave, or contribute it to what some companies now call the company brain? Ashley drew the line between the output and the person.
The tool and the work that the tool creates belongs to the company. So if you’re in a creative role creating content for your company, that stuff belongs to the company. Now the way of thinking and the way you got to that, that’s on you. - Brandon Ashley
That distinction gives HR a cleaner policy conversation than a ban does. Ashley still uses a personal account for brainstorming and best-practice questions, the kind of work that never touches company data.
What does IT need from HR before approving an AI tool?
A narrow scope, a business case, and an early conversation, ideally before anyone has signed. Rebecca asked Ashley for the most surprising AI request ever to land on the service desk. The surprise was not the request. It was the way it arrived.
The Friday afternoon email saying, hey. We just bought this. Can you roll this out next week? That doesn’t go well. - Brandon Ashley
Behind every rollout sits a process: securing the data, putting up guardrails, tracking who holds admin access. A tool bought first and explained later turns a simple ask into a project. Brought in earlier, IT can sometimes point to an application the company already owns that does the same job, or retool one, and skip the rollout entirely.
Ashley’s first question is often whether the tool is AI at all.
I feel like AI is a marketing term right now. It’s everywhere. Everything is AI. Everything has AI in it because I feel like every company needs to have a product with AI because they don’t wanna feel like they’re being left behind. - Brandon Ashley
Some products have a small piece of AI built into a larger tool, and not everything sold as AI is AI. So does scope. A salesperson will show you everything the product can do, and you do not need all of it. Go into the call knowing the few things you need, which extras would be nice to have, and what to ask for. Ashley suggested getting a list of questions from IT before the sales call. Vendors may not know every answer on the spot, but they can send white papers and documentation afterward.
What IT wants to see, in Ashley’s words, is that you have done your homework. This is not an impulse buy or the next shiny thing. There is a business case and some thought behind it. Rebecca has seen IT leaders join buying conversations earlier than before, with buyers arriving at calls carrying questions their IT team wrote. Ashley’s view is that a written list of requirements helps the vendor too, because it rules out what will not fit.
What should IT check before a tool connects to your company’s AI?
Whether access can be controlled, separated, revoked, and logged, starting with single sign-on. Rebecca asked what Ashley looks at early in a purchase, beyond what a buyer hears in a demo.
I think any of the enterprise options are going to support single sign on, but that’s a bare minimum. - Brandon Ashley
Beyond that, IT wants to control access by application, limit features it does not want in use, and make sure shared documents reach the people who should see them and nobody else. Ashley’s test is whether users can be separated or whether the connection is all or nothing. Done with thought, a connection can be secure. Opened to everyone without it, it is where companies get in trouble.
The checks Ashley walks through change with every case, which is why the relationship matters more than any checklist. The recurring ones:
- Security. Whether the tool can separate users and limit what each one sees.
- Logging. Whether every access can be recorded and reviewed later.
- Fit with the current environment. How the tool behaves alongside the systems already in place.
- Other integrations. Which tools have to connect for it to work at all.
- The data. What the tool will actually be reading, and whether it is any good.
That last point led Ashley to suggest a slow approach: add one application at a time. Just because a tool can integrate with everything does not mean it should. If the data you send it is junk, it will not help you, so leave it out until you need it. Rebecca has seen what junk data does.
I would connect something that was a less than clean database to an AI tool, and then I’ll ask questions. And I would get the most unhinged answers. - Rebecca Taylor, head of community and content at AllVoices
The bad answers had a use. They showed where the system needed cleanup before the AI could be trusted. Rebecca framed it as the difference between a system of record and a system of action. AI is only as good as the data it sits on, so the security of the tools that hold your HR data matters before any AI is involved, and more once it is.
Is it safe to connect your HRIS to an AI tool?
It can be, when the connection is scoped and the permissions inside the HRIS are right first. For a while, one school of thought in HR held that you should never connect a human resource information system to AI at all. Rebecca understood why, and argued that the better question is what has to be true before you do.
Much of that connecting now happens through MCP, the Model Context Protocol, an open-source standard for connecting AI applications such as Claude or ChatGPT to external systems. Many HR vendors now offer one. Rebecca noted that it requires an enterprise account, which brings the connection under the same governance as everything else IT manages.
Rebecca’s biggest worry is the step teams skip. Admin lists in an applicant tracking system or an HRIS drift over time. Someone gets admin access for one hiring push, and nobody ever takes it back.
If you’re considering how to connect some of your tools, first, go into the tools that you have and see who your admins are, what level permissions they have. - Rebecca Taylor
A connected AI tool honors the permission levels and guardrails that each platform gives it. Someone without access to employee relations information will not get it, no matter what they ask. Someone who accidentally kept a seat that allows it will. Ashley added why that risk stays hidden until the connection goes live.
They might not even realize they have that access still. So they’re not going into the tool directly. But once you integrate in, then all of a sudden, strange things can start to happen. - Brandon Ashley
Global admins nobody knows about are the same problem multiplied. The audit has nothing to do with AI itself, which is what makes it easy to skip.
Rebecca described how AllVoices approaches the same question for employee relations data, which is especially sensitive. As its MCP connection develops, the team decides at the guardrail level what goes to an AI tool like Claude and what someone must always do inside the platform. Part of that comes from asking customers whether they would approve an action taken from Claude or keep every action in the platform.
What should HR do when IT says no to an AI tool?
Ask why, respectfully, and take the answer back to the vendor. Rebecca hears the complaint often: HR knows it should be trying new tools, and IT keeps saying no. Ashley agreed IT has to act as a gatekeeper sometimes, and explained why a bare no fails everyone.
Then you become the person who always says no, then you become the department that, oh, don’t get them involved. They’re just gonna say no, or they’re just gonna stop it. And then everyone just goes around you, and then we find all these things after, and then it’s just so much worse. - Brandon Ashley
So when the answer is no, ask what the concern is. Ashley’s point was that IT should be able to tell you. Maybe the company already has an application that does the same thing. Maybe security cannot approve it for a specific reason, and the idea is dead. Or maybe the concern is one the vendor hears all the time. Take the security team’s questions back and ask for a response. Many vendors will send a white paper or set up a call to answer the security team directly.
Rebecca described the vendor side of that exchange: a channel just for security questions, collected from demos and follow-up emails with each prospect’s IT team. The goal on both sides is to get everyone who wants to reach a yes to a yes. Rebecca also turned the complaint back on HR.
IT is saying no. That can probably be as lazy of an answer as HR said no, which I hear people say all the time. - Rebecca Taylor
Neither department is just saying no. There is context and there are reasons, and the way through is to ask about them.
A small company may not have an IT team to ask at all. Ashley’s suggestion there was an integration partner. Some firms do these integrations every day and can run an initial AI evaluation, plan which systems to connect first, or handle the whole integration for you.
How can HR make it easier for IT to say yes?
Build the relationship before you need the approval. Rebecca brought Ashley a question from HR peers: what can HR do to make tool shopping and rollout easier, before throwing something at IT at midnight on a Friday? After a nod to bagels and baked goods, Ashley went back to the relationship.
Ashley also explained how IT tends to answer a request.
I like to lead with a qualified yes. So if you come to me, I’m probably gonna say yes, but, or yes if. - Brandon Ashley
The conditions are easier to meet when you have done the research. A request from someone IT has never heard of, asking for a rollout by next week, raises red flags and slows everything down. A request IT already knows about, with a vendor someone in IT has met, moves much faster. Familiarity with each other and with the vendor is most of the work.
It can feel as if you need every answer before you approach a partner. Rebecca admitted the temptation, and said a good IT relationship removes it.
I could go to you and I could say, Brandon, I don’t have the language for the thing I’m trying to do. - Rebecca Taylor
HR plays interpreter for employees all the time. A good IT partner does the same for HR: tell them the need, and they will tell you what to ask and what information to gather.
Bring IT in early, and know the job the tool is for
Bring IT in early. Rebecca has heard the same rallying cry at a legal conference, at a compliance conference, and across HR: every function wants to be brought in early. HR wants to hear about a manager’s problem before it becomes an investigation, and IT wants to hear about a tool before it becomes a Friday rollout. Ashley pointed out that it was true long before AI, back when the two of them handled onboarding and offboarding. Laptops do not prepare themselves overnight.
The second habit is knowing what you are trying to accomplish. Rebecca argued for anchoring every evaluation in the larger job to be done, not a single task, and using it as the North Star that keeps the buying committee pointed the same way. Ashley explained why that anchor matters.
Once you start to dig into these things, you can go sideways. You can go all these different paths, and you might lose focus on what the initial request was for. - Brandon Ashley
The same curiosity applies to vendors. If an answer sounds like sales or marketing copy, dig deeper. Rebecca noted that vendors usually have more detail and are waiting to see what level you are ready for. AllVoices publishes how it handles that question in its approach to responsible AI: human review at every step, strict data handling, and no decision made by AI alone.
See how AllVoices handles employee relations data
AllVoices keeps employee relations cases in one platform, with human review built into every step its AI takes.
Request a demoNone of this requires HR to become technical overnight. It requires asking early, asking why, and asking the people whose job is to know. AI is neither evil nor magic. It is as safe as the access and logging you build around it.
Frequently asked questions
No. Ashley’s concern was that nobody, including the company, knows where prompts and uploads in a personal account are stored or who can see them. An enterprise account with single sign-on, access controls by role, and logging is where work involving employee data belongs, once IT has set it up properly.
Shadow AI is employees using AI tools the company has not approved for work, usually through personal accounts. It tends to start as self-directed learning and drift into daily work. Ashley’s experience is that a flat no without a reason pushes the practice out of sight rather than ending it.
No. Rebecca explained that a connected AI tool honors the permission levels and guardrails of the platform it connects to. That makes the platform’s admin list the real control. Anyone who still holds an old admin seat can reach that data through the AI tool, so audit permissions before you connect.
The company, in Ashley’s view. The tool and the work it produces belong to the employer, including content created in a creative role. The way of thinking and the method the employee used to get there stay with the person. Ashley still uses a personal account for brainstorming that never touches company data.
Work with an integration partner. Ashley pointed out that some firms do these integrations every day. They can run an initial AI evaluation, recommend which systems to connect first and which to hold back, and build a plan, or handle the entire integration for you.
Before the first sales call. Ashley suggested getting a list of questions from IT to ask the vendor, then returning with the answers, a narrow scope, and a business case. IT can also tell you whether the company already owns a tool that does the job, which can avoid a rollout entirely.



