We were mid-way through building out our customer support team when ChatGPT 4.0 became a household name. Up until then, our hiring plan was straightforward. Train junior reps on simple tickets, let them graduate to complex cases over six months, and backfill with new hires as we grew. Then, in a single week, the investor questions shifted. No one asked about our training process or our ticket backlog anymore. They wanted to know why we weren’t “just using AI” to handle the bulk of support.
It wasn’t an unfair question. Large language models were suddenly doing, in seconds, what it took a human three minutes to complete. They could summarise a customer complaint, suggest a solution from our documentation, and draft a polite response. On the surface, this was a cost-saver and a speed upgrade. But it also triggered a set of challenges we weren’t prepared for—challenges not about the technology itself, but about the human architecture around it.
We assumed people would welcome the change. Instead, some saw it as a threat. Junior hires quietly asked if their contracts would be renewed. Senior hires wondered aloud why they should spend time mentoring people whose roles might disappear. We had been on track to grow headcount by 40 percent in the next two quarters. Within two weeks of implementing the LLM, we froze hiring entirely. The technology hadn’t just automated tasks—it had rewritten the emotional contract between us and our team.
From a purely operational perspective, the situation was logical. The AI could handle repetitive queries, flag exceptions, and leave human reps free to focus on complex, high-value interactions. But “freeing people up” is a vague promise. In reality, without a clear redefinition of roles, the workday became a grey zone. The easy tasks were gone, but no one knew what the new priorities were. The result was an undercurrent of uncertainty that eroded morale.
The first week we went fully live with AI-assisted support, our customer satisfaction score jumped. Average resolution time dropped. Investors nodded in approval. But inside the company, something else was happening. Participation in training slowed. Team members hesitated to take ownership of complex cases. One rep put it bluntly: “If the AI does all the easy stuff, that means we only get the hardest cases—and if we mess those up, the fallout is worse.” It wasn’t resistance to change; it was recognition that the gap between entry-level and high-skill work had suddenly widened, leaving no safe middle ground for skill development.
This was the first real breakdown point. In the old system, new hires learned through repetition and gradual exposure to complexity. They could build confidence on low-stakes cases before moving to high-pressure situations. With LLMs in place, that ladder vanished overnight. Now, onboarding meant diving straight into edge cases, with no runway to practise the basics. The AI had made the work faster, but it had also made it harder for humans to learn.
Our moment of clarity came during a Friday review call. Our lead support rep, who had been with us since the MVP days, said, “If you want us to use the AI, fine. But tell us where we fit in after it does the first draft.” The question wasn’t “Will AI take my job?” It was “What does my job become when AI is in the room?” We had been so focused on efficiency metrics that we forgot to redesign the human role to make sense to the human.
I’ve since come to believe that LLMs don’t automatically eliminate jobs. What they eliminate are poorly designed jobs—roles built entirely on repetitive, predictable, text-based tasks. Those can and will be automated. But roles anchored in judgment, negotiation, and creativity under constraint? Those will not only survive but may become more valuable, because the AI accelerates the low-value tasks that previously consumed that person’s bandwidth. The founder’s work, then, is to re-scope roles so that humans are consistently playing at the top of their skill range.
In practice, this means you can’t simply “drop in” AI and expect the org chart to hold. Workflows need to be rebuilt from scratch. If you keep the same structure, you risk creating dead zones—positions where employees feel redundant but remain on payroll. That’s a morale killer and a financial inefficiency rolled into one. Instead, think of LLMs as a catalyst for organisational surgery. You need to decide, deliberately, what the AI owns, what the human owns, and where the two interact.
There’s also a training paradox to navigate. On the surface, if AI can do a task, you might think you can reduce training. In reality, you have to increase it. Reviewing, correcting, and contextualising AI outputs is a skill in itself—and it’s a different skill than doing the work from scratch. In our case, we had to retrain reps to stop treating the AI’s suggestions as final. Early on, we had a case where the AI recommended a refund outside our policy, and the rep approved it without checking. The error cost us more than the AI subscription for the month. That was the wake-up call: AI fluency isn’t about knowing prompts. It’s about knowing where the AI is likely to fail, and how to catch those failures before they hit a customer.
This isn’t just a customer support issue. I’ve seen the same pattern play out in marketing teams, legal review, and product documentation. An LLM drafts a press release; the human edits it. The AI summarises a contract; the lawyer verifies it. The machine handles volume; the person ensures nuance. When the boundaries between those two roles are clear, the system works. When they’re not, you get either blind trust in AI or pointless double-checking of everything, both of which erode the efficiency gains you were chasing.
One of the less discussed impacts of LLMs is on career pathing. If the “entry-level” work is automated, where does a new hire start? In creative roles, this might mean starting with mid-level complexity tasks, which can overwhelm someone fresh out of school. In technical fields, it might mean there’s no safe space to practise before working on high-stakes systems. Founders need to think about what the new version of “junior” looks like. It might involve simulation environments, sandboxed projects, or AI-assisted shadowing before touching live work. If you don’t create that bridge, you risk a talent pipeline collapse—not because there aren’t enough candidates, but because you’ve removed the places where they learn.
Looking back, if I could redo our LLM integration, I’d begin with a mapping exercise before touching production systems. I’d write out exactly which tasks the AI owns, which tasks require human judgment, and where the two hand off. I’d make “AI oversight” an explicit skill in the job description, not an informal expectation. I’d run failure drills where the AI makes the wrong call, just to make sure humans know how to step in confidently. And I’d communicate openly with the team—not as a blanket reassurance, but as a shared understanding—that the AI is here to remove the work that never should have been theirs in the first place.
Because the truth is, AI adoption won’t wait for your team to feel ready. But your team will wait for you to explain where they stand. And if you don’t, the best people will leave before the technology ever has a chance to replace them. That’s the irony—without clear human role design, you lose talent not to automation, but to uncertainty.
In other industries, I’ve seen founders handle this transition well. A design agency in Singapore replaced its first-draft copywriters with LLMs but immediately re-scoped those roles into “creative concept developers,” training them to brief the AI with more nuance and spend more time on visual-story integration. A logistics startup in Riyadh used LLMs to automate status updates but retrained their coordinators to focus on relationship management with high-value clients. In both cases, the AI was framed not as a job killer but as a job shaper. The teams knew what was leaving their plate and what was coming onto it. The morale stayed intact because the career story still made sense.
The founders who get this right will treat AI not as an off-the-shelf upgrade but as an inflection point in organisational design. They will see that efficiency gains on paper mean nothing if the human system around the technology breaks. And they will understand that the biggest risk of LLMs in the workplace isn’t displacement—it’s disengagement.
If you’re a founder staring at a roadmap with “integrate LLM” somewhere in the next quarter, don’t start with the tech stack. Start with the people stack. Decide who owns what, what skills you’ll invest in, and how you’ll rebuild the career ladder your AI just tore out. Because the companies that survive this shift won’t be the ones that deploy LLMs first. They’ll be the ones that deploy them without losing the humans who make the rest of the system work.
WhatsApp
Instagram
Threads
