If you went looking for Dag Fjelstad on LinkedIn, you wouldn't find him. He exists only in the composite space between several real conversations and a body of research about how enterprise roles are quietly transforming under agent adoption. But the situation he describes is documented, the tensions are real, and his former career as a librarian felt too good not to give him.
We spoke over video. He was in what appeared to be a home office in Minneapolis, a single shelf of Dewey Decimal System reference guides visible behind him, which he later admitted he keeps "for the bit, but also not entirely for the bit."
You were a librarian before you were in product operations. How does that happen?
Dag: Shorter distance than you'd think. I was a cataloging librarian for six years. Subject headings, authority control, the whole system of deciding what something is so other people can find it later. Then a SaaS company needed someone to organize their reporting taxonomy, and it turned out that was just cataloging. With dashboards. I managed a team of seven analysts for about four years after that. Now it's me and two agents, and honestly the work is closer to librarianship than it ever was when I had the team.
Closer how?
Dag: When I had analysts, my job was allocation. Who works on what, are we on track, does this deck look right before it goes to the VP. Management stuff. Now my job is — okay, this is going to sound pretentious, but it's the honest answer — defining what words mean.
Before an agent can answer "how many active customers do we have in EMEA," someone has to decide what "active" means. Logged in within 12 months? Completed a transaction in four quarters? Had a support ticket? And "EMEA" — does that include Turkey this quarter? Because it didn't last quarter when we were using the old regional mapping. These aren't edge cases. This is every single question.1
That sounds like it should be straightforward.
Dag: Right? Five-minute conversation. Except those five-minute conversations are where the actual analytical work lives now. Running the query is trivial. Defining what the query should ask — that's the whole game. Over half of agent errors on enterprise SQL tasks turn out to be filter errors. Wrong conditions, wrong enumeration values. The agent's syntax is fine. It just doesn't know what you meant.2
You're describing ontology.
Dag: I am describing ontology, and I will accept the pretension charge. My old cataloging professor would be thrilled. She always said the most important intellectual work is the work that looks administrative. I think she was trying to make us feel better about our career prospects, but she was also right.
What does a typical day look like now versus two years ago?
Dag: Two years ago: standup with the team, review three or four analysis requests, check someone's SQL, sit in a stakeholder meeting, give feedback on a deck. Pretty recognizable management job.
Now: I spend my mornings writing what I've started calling "enclosure documents." Borrowed from lab automation — before you let an autonomous system run, you define the objective, the materials, the boundaries, the safety limits.3 I do the same thing for analytics. What question are we answering, what data sources are in scope, what's the metric definition and version, what would count as a suspicious result, and when should the agent stop and ask a human.
Afternoons I review what the agents produced. But here's the part that's hard to explain to my manager: reviewing means going back to the query, the joins, the filters. Not reading the summary slide.4
Why does that distinction matter so much?
Dag: If I just read the executive summary and say "looks good," I'm a rubber stamp. The organization gets to put my name on it, and I haven't actually inspected anything. There's research on this — people don't feel ownership of AI-generated work but they still put their name on it.5 I think about that more than is probably healthy. I don't want to be the person whose professional reputation launders agent output.
How does your organization see this work?
Dag: [long pause]
They see throughput. The agents produce more analyses faster. That's visible, that's on the dashboard, that gets mentioned in the all-hands. What's not visible is that I spent three hours last Tuesday in a meeting with finance and product arguing about whether "net revenue retention" includes expansion from upsells closed by partners. That meeting is the reason the agent's output next week will be correct. But it produces no slide. It's the work before the work, and it's harder than the work.
Do you feel like your job is more demanding now?
Dag: Unambiguously yes.
But less recognized?
Dag: Also unambiguously yes. And I don't think that's a coincidence. There's a McKinsey survey showing that 80% of people report personal productivity gains from AI, but the share of organizations attributing positive financial impact hasn't budged since 2025.6 I think the specification work is part of that gap. Someone has to do it. It's expensive. And it doesn't register as "AI productivity" because it looks like meetings and documents and arguments about definitions.
You mentioned you used to have seven analysts. What happened to them?
Dag: Three were redeployed to other teams. Two left the company. Two moved into more specialized roles — one does data engineering now, one does customer research. None of them do what I'd call traditional analyst work anymore.
And here's what worries me: they used to catch each other's mistakes. Peer review. Someone would look at a query and say, "you're using the wrong fiscal year boundary." That feedback loop is gone. I'm the only one reviewing now, and I'm reviewing more output with less help. There's a useful analogy from aviation — regulations require recent takeoffs and landings before a pilot can carry passengers. Review competence has to be maintained through practice, not just credentialed once.7 I'm not maintaining mine. I'm spending it down.
What does "done" mean now?
Dag: [laughs] That's the question I spend about 40% of my time on. "Done" used to mean an analyst finished the deck and I approved it. Clear enough. Now "done" means the specification was correct, the agent ran within the enclosure, the output matches the evidence chain, and someone with authority has granted the organization permission to treat this number as settled.8
That last part — the permission to treat it as settled — that's the scarce resource. Not compute, not data. Finality.
Do you have the authority to say "not done yet"?
Dag: Sometimes. When I do, it's the most valuable thing I contribute. When I don't — when someone senior needs the number by Thursday and the specification isn't tight — I stop being a quality gate and become an invisible repair layer instead.9 Those are two very different jobs wearing the same title.
If you could change one thing about how your organization understands your role, what would it be?
Dag: I'd want them to understand that the clean number in the deck exists because someone spent two days deciding what the question actually meant. And that if that person leaves, or burns out, or gets overloaded, the numbers will still come out. They'll just be wrong in ways nobody catches until the quarterly review.
The cataloging work was always like that, too. Nobody notices the classification system until a book is misfiled and you can't find it. By then the librarian who built the system has been laid off because the catalog "runs itself."
Dag paused here, then added: "You can use that as the ending if you want. It's a little dramatic. But it's also just... what happened to a lot of library systems."
We left it there.
Footnotes
-
Research on enterprise analytics agents identifies hidden conventions — fiscal-year boundaries, identity rules, channel definitions — as the primary source of specification difficulty. See KB: enterprise-analytics-and-sql-agents. ↩
-
The EntSQL benchmark found that 54.6% of failed predictions were WRONG_FILTER errors (missing conditions, wrong enumeration values) and 14.4% were WRONG_SCOPE errors. See KB: enterprise-analytics-and-sql-agents. ↩
-
The "enclosure" concept is drawn from autonomous-laboratory research, where physical autonomy requires prior human definition of objectives, materials, equipment, and safety limits. See KB: autonomous-laboratories-and-physical-agents. ↩
-
Research distinguishes between approving an agent's prose summary (weak oversight) and reviewing the query, schema assumptions, and metric definition (substantive review). See KB: human-in-the-loop-and-work-design. ↩
-
Draxler et al., "The AI Ghostwriter Effect," ACM Transactions on Computer-Human Interaction 31(2), 2024. https://doi.org/10.1145/3637875 ↩
-
McKinsey, "The State of AI: Global Survey 2026," August 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai ↩
-
Aviation regulations require recent takeoffs and landings before a pilot carries passengers; the principle that review competence must be maintained through practice, not merely credentialed once, applies to analytical oversight. See KB: human-in-the-loop-and-work-design. ↩
-
Process intelligence research identifies four layers of organizational process — recorded, documented, operational, and delegable — with completion requiring a known final state that can be validated. See KB: process-intelligence-and-workflow-legibility. ↩
-
The "exception topology" framework describes two possible futures for this kind of work: an empowered profession that controls finality, or an invisible repair layer that receives responsibility after meaningful intervention is no longer possible. See KB: exception-topology. ↩