She Refused to Let AI Near Her Data. Now She Runs It.

For years, Liz Lukan kept AI out of ISS’s systems — on purpose. Then the technology met her conditions, and she became the person who governs it.

The Challenge

Liz Lukan had been saying no to AI for years, and she was right to.

As Director of Database Operations at ISS of New York, she’s responsible for the data of more than 2,200 people with developmental disabilities — medical records, identifying information, and Medicaid compliance. Over twelve years she helped build every part of the system, from bookkeeping to billing. When the founder asked about AI, when the development team raised it, her answer held: “No, Alan, no, no, just no. It’s not ready yet. I’ll let you know when it’s ready.”

That wasn’t reflex. It was judgment about what she was protecting. She’d heard the same horror stories everyone has — AI that hallucinates an answer, AI that wipes a database it was trusted to access. She wasn’t going to find out the hard way on her system.

“The integrity of the data in the system is my responsibility. If I unleashed something into that environment that destroyed stuff — that’s millions of bits of data, gone.”
She watched the industry do the opposite. “Everybody’s jumping on the AI bandwagon. They don’t ask if they should. What is the value in doing this?”

What was at stake: ISS was moving toward a self-service portal for participants and families — people often already frustrated with the process of making serious decisions about their own care.

Bolting a chatbot onto a thousand internal help documents without sufficient boundaries could surface multiple issues, such as sharing internal compliance notes to an external user, or pulling another state’s Medicaid rule into a New York answer.

The “no” from Liz was the thing standing between those users and an authoritative-sounding AI hallucination. As Liz put it:

“I will die on that hill.”

So, when Steve Sykora at Harmonic told her the technology had finally reached her bar, she didn’t take his word for it. “All right,” she said. “Prove it to me.”

The breakthrough came when we asked what it would take for the person protecting this data to say yes

Most teams put AI on their documents first and try to fence it in afterward. Liz’s conditions forced the opposite order.

ISS has a knowledge base with over a thousand help documents in a wiki called Bookstack — but several user groups who don’t see the same things. A participant asking how to submit a timesheet should never get a result related to internal IT procedures. Admin staff need business rules, not deep technical documentation. So before anything conversational got built, every page was classified, based on who’s allowed to see it, and what kind of content it is.

An early test proved why the order mattered. A portal-facing build was told, by prompt, to use portal content only. It answered from internal shelves anyway. The model didn’t honor the instruction.

The fix wasn’t a better prompt. It was making the type of audience a filter in the database query itself, so off-limits documents never reached the model in the first place. The boundary became structural, built into how documents are stored and retrieved rather than a request the model can ignore.
That distinction is exactly what Liz had held out for.

The skeptic tested it herself, and the boundaries held

When the team believed it was ready, Liz ran it through its paces herself. She asked the portal version specific questions pulled from every shelf it wasn’t supposed to touch.
“I have asked Kelly a very specific question from each of the shelves it shouldn’t have access to, and not once did I get a link that wasn’t on the public help website. Her answers might not always be right, but Kelly is staying in her lane.”

She’d set the conditions years earlier: it had to be contained; it had to know its boundaries. She tested those boundaries from every angle, and they held. She reported results that morning and told the team to move it to production before lunch. When someone’s conditions are specific and the system meets them, the approval process moves fast.

The result: she runs it

ISS now has versions of Kelly scoped to different audiences — public portal, admin staff, and IT — all drawing from one knowledge base, each limited to what its audience is allowed to see.

She controls access herself. With over a thousand pages in Bookstack, Liz controls access at the shelf level — not page by page. “Drilling down into individual pages just makes it easier to make a mistake and give access to someone who shouldn’t have it.” That’s not a preference. It’s the same logic that kept AI out of her systems for years. What changed isn’t her standards. It’s that the system finally meets them.

She set the rollout sequence. The simplest, most constrained audience went live first — the public portal version, with the tightest boundaries — while her team tested the more complex internal versions. Nobody had to bet the whole system on day one.

She maintains it herself. When she adds a procedure or a revised state memo, she syncs Kelly in three steps: pull from Bookstack, set the metadata, push it live. No developer ticket. No build cycle. She decides what Kelly knows, and when.

The metric is silence. “Success is quiet. Failure is loud and noisy and full of support tickets and screaming.” Since launch, no tickets reporting that Kelly exposed something she shouldn’t have. The questions she can’t answer are documentation that hasn’t been written yet — “not a technical problem,” as Liz puts it — not boundaries failing.
The support inbox has a first line that isn’t a person. Matt Baluyot’s team spent two to three hours every morning answering emails already answered somewhere in Bookstack. Stakeholders can ask Kelly first now. And staff who’d never risk looking foolish in front of a colleague will ask a machine without hesitating.

The documentation culture is shifting. Departments that used to guard their own processes are putting them into the shared system, because now there’s a reason to. Liz even uses Kelly to find things in the Bookstack she built herself: “I know it’s in there, but maybe I’ve forgotten what I called it or which shelf it’s on.”

Liz didn’t come around on AI. She set a bar, and the system finally cleared it. That’s a different thing, and a more durable one.

Why It Worked

The conditions determined the architecture. The metadata layer exists because Liz wouldn’t accept a prompt-based promise. Her resistance produced a better-built system than mere enthusiasm would have.

The boundary is structural, not conversational. When the model tried to answer from the wrong shelf — and it did — the filter caught it at the query level, before the question ever reached the model.

The skeptic got control, not a sales pitch. Liz wasn’t talked into trusting AI. She was given the means to verify it herself and the controls to govern it afterward. When the person most responsible for protecting the data is the one running the system, adoption accelerates.

It wasn’t plug-and-play. Throttling errors, case-mismatched metadata keys, documents pushed with empty content, a model that confidently answered from shelves it shouldn’t reach. Each got solved. This was real engineering, and Harmonic’s willingness to show Liz the messy parts is part of why she trusts the result.

If your documentation is good but your people still can’t find what they need

That gap between “the information exists” and “people can reach it when they need it” is common in organizations with complex operations, multiple audiences, and compliance requirements. If staff keep private document stashes because they don’t trust the central system, or the person responsible for your data won’t let AI near it, those are tensions worth talking about.

Resistance like Liz’s is not an obstacle to work around. Handled right, it’s what gets the system built correctly the first time.

Our aim is to understand your situation before suggesting anything.