Case Study — — by Mahmoud Zalt
A first-person account of what replacing a traditional operations team with AI employees actually looks like: the functions replaced, the holdouts, the surprises, and the results 6 months in.
We were not running a big team. The company was at early revenue with a founding team of two and a small cluster of outside help: one part-time content freelancer on retainer, a VA for calendar and research, an email marketing contractor we brought in once a month, and a support queue we were personally managing between the two of us.
The outbound pipeline was the thing we were the most honest about not doing well. Neither of us was spending the hours on it. We had a list and a sequence and it ran inconsistently because the humans running it had other priorities.
That was the state of the operations layer six months before we moved everything to AI employees on Sistava.
We did not start from a list of AI capabilities. We started from a list of things that were not getting done, and then asked four questions about each one. Any function that answered yes to all four went first.
Sales outreach answered yes four times, which is why it went first. Investor updates answered no to the third question immediately, which is why they never moved. The test took us about an hour and saved us from the mistake most people make, which is starting with the most visible function rather than the most suitable one.
We did not do this all at once. We moved one function at a time, evaluated the output for two to three weeks, and then moved to the next. Here is what happened with each one.
We hired an AI SDR first. Uploaded our ICP document, three examples of emails that had worked, and connected Gmail and our CRM. The AI started running outreach sequences the same day. Week two, it booked a demo that led directly to a paid contract. We had not closed an outbound deal in four months before that. The SDR now runs sequences consistently that our previous setup could not maintain. This was the clearest win.
We expected the content to be noticeably worse than our freelancer. It was not. The AI Content Marketer published 10 posts in the first month against 2 from the previous month. Quality required more editing early on, less by week three. The posts the AI writes now are well-structured, targeted at real search questions, and appropriately voiced for our brand. The total editing time per post is about 15 minutes. The freelancer took 2 to 3 hours of back and forth per post.
We handed the support inbox to an AI Support Agent and kept escalation routed to us. The AI handled 71% of tickets without escalation in the first week. Response time went from an average of 4.5 hours to under 2 minutes. One customer emailed to say the support quality had improved significantly. They did not know we had changed anything.
We went from 1 newsletter per month to 2 campaigns per week. The AI Email Marketer manages segmentation, writes the campaigns, and handles the follow-up sequences for new signups. Open rates went up in month two after the AI started testing subject lines systematically. The contractor we had used sent campaigns when we remembered to brief them.
Competitive research and prospect background preparation had been things we did inconsistently or skipped. The AI Research Analyst now prepares a competitor update every Monday and a background brief on every prospect before demos. We show up to meetings knowing things that were previously only visible to funded teams with research staff.
The word replaced makes this sound like a switch we flipped. It was closer to onboarding a new starter, compressed into a week. The same shape repeated for every function, and the weeks that went badly were the ones where we skipped a stage because we were impatient.
By the end of that week the function was running and we were reviewing rather than doing. The two functions where we rushed day three both needed a second pass a fortnight later, which cost more time than the day we saved.
If there is one transferable lesson in this whole account, it is that output quality tracked brief quality almost exactly. The functions we described well worked well in week one. The functions we described loosely produced generic work until we went back and fixed the description. Nothing about the technology changed in between.
The unexpected side effect is that we now have documentation of how our company operates, written for the first time because a machine needed it. If we hired a human tomorrow, those briefs would be the onboarding pack, and they are better than anything we would have written for a person.
The question people ask first is what happens when it gets something wrong. The answer is that the interesting part is not the mistake, it is what the mistake was allowed to touch. We spent more time on the permission boundaries than on the instructions, and it is the reason nothing has gone badly wrong in six months.
That last one is the discipline most likely to slip, and it is the one that catches quality drift. Drift is invisible day to day and obvious across a month, so the weekly slot matters more than how thorough it is. Ours takes about twenty minutes.
The first surprise was how much mental load the previous setup was carrying. Managing freelancers and contractors involves constant overhead: briefing, feedback, payments, availability windows, revision cycles. When the AI employees took over, we did not just recover hours. We recovered cognitive capacity. The overhead of managing the operations layer went from background noise that was always running to almost nothing.
The second surprise was that the consistency of the AI workforce exposed how inconsistent our previous setup had been. The sales SDR runs sequences every day. The Content Marketer publishes every week. The newsletter goes out on schedule. When you see that level of consistency for the first time, you realize how much the previous setup had been drifting without you noticing. The inconsistency was invisible because it happened gradually.
Accounts like this one tend to stop at the good news, so here is the counterweight, from a company far larger than ours. In February 2024 Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two thirds of its customer service chats, doing what it described as the equivalent work of 700 full-time agents, cutting resolution time from 11 minutes to under 2 and driving an estimated $40 million profit improvement. Those numbers are real and they came from Klarna itself.
By May 2025 the same CEO said the cost-cutting had gone too far, that the full tilt toward AI support had produced lower quality work on harder cases, and that the company would recruit human agents again. Both halves of that story are true at once, and reading only the first half is how a team ends up making an expensive mistake. The lesson we took is not that AI support fails. It is that the easy 70 percent and the hard 30 percent are different problems, and a plan that only accounts for the first one will break in public.
It is also why our support employee escalates rather than improvises. The goal was never to remove humans from the hard cases. It was to stop humans spending their day on the cases that did not need them.
Three functions stayed human. Closing complex deals: the relationship and judgment in the final stages of a sale require a human who can read the room and make concessions in real time. Product decisions: what to build next, what to cut, how to prioritize. These are judgment calls with long-term consequences that require the founders' direct involvement. Investor communications: the relationship with investors is one of the highest-stakes judgment-intensive functions in an early company.
The pattern is consistent: anything that is a process running on communication moved to AI. Anything requiring real-time human judgment, ongoing relationship, or creative originality stayed human.
The headline is that the operational cost is a fraction of the freelancer and contractor setup. The more useful version is what actually left the budget and what quietly arrived, because the second list is the one people forget when they build the business case.
If you are building a case for this internally, do not lead with the saving. Lead with the work that currently does not happen at all, because that is where the honest gain is and it is the part a spreadsheet comparing salaries will miss entirely.
Six months in, the AI workforce is handling all five functions without a regression in quality. The operational cost is a fraction of what we were paying for the freelancer and contractor setup. The output volume is higher across every function. The overhead of managing the operations layer is close to zero.
What we would do differently: we would have started with the AI SDR earlier. The four months before we moved outbound to AI were four months of inconsistent prospecting that cost us pipeline we cannot recover. The support move should also have happened sooner. Handling support tickets personally for that long was the single biggest waste of founder time in the entire pre-AI period.
The question we get most often from other founders who hear about this setup is: does it actually work or is it a hack that falls apart under pressure? Six months in with real customers and real revenue, the answer is that it works. It is not a hack. It is a different kind of company structure.
The fastest way to judge that for yourself is not to read another account of someone else's setup. It is to give one of these employees a real task from your own week and see what comes back.
Everything above compresses into a short list. If you do only these five things, you will get most of the result and avoid the failure modes that made our own transition slower than it needed to be.
If you want the roles rather than the philosophy, the AI-native startup stack lists every function you can run this way right now, in the order most founders hire them. It is close to the order we ended up following, though we worked that out the slow way.
Hiring and onboarding an AI employee on Sistava takes 20 to 30 minutes. The transition period where output requires daily review is typically 1 to 2 weeks. By week three, most founders are reviewing weekly rather than daily. Full transition for a function is about one month.
Quality varies by function. For high-volume, process-driven work like support and outbound, quality holds or improves because the AI is more consistent than a human who has many competing priorities. For creative work like content, quality depends on the briefing quality and how much feedback is given early on. By month two, most founders report the gap has closed. Complex or unusual cases are the exception, which is why escalation rules matter.
You correct it once with explicit feedback and the employee adjusts. Sistava's guardrails catch high-risk actions before they execute, so most errors are recoverable. The audit log shows every action, so nothing is hidden. The more important control is what the mistake was allowed to touch: anything involving money, a customer commitment, or a public channel should require a human yes.
Most customers do not notice. Support response times drop significantly and accuracy stays high, which customers experience as an improvement. For outbound, prospects respond to well-personalized emails the same way they respond to human-written ones. The minority who ask directly get an honest answer.
No. The AI-native structure is most visible in small companies because they are building from scratch, but established companies are running specific functions on AI employees too. The economics work at any size where you have process-driven functions that consume human capacity.
The one that answers yes to four questions: is it mostly communication, can you describe what good output looks like, is a mistake recoverable, and is a person currently cutting corners because of the volume? For most small companies that is outbound sales or first-line support. Avoid starting with anything where the first mistake is irreversible.
Two, and they are related. The first is a thin brief: output quality tracks brief quality almost exactly, and adjectives like professional and friendly are not a brief. The second is skipping the week where everything is held for approval, which is the week you find out which parts of your brief were vague. Klarna's public reversal on AI-only customer support is the large-scale version of the same lesson: plan for the hard cases separately, because they behave nothing like the easy ones.
Six months on, the thing I would tell my earlier self is that this is a management change wearing a technology costume. The hard parts were writing down what good looks like, deciding what a mistake is allowed to touch, and reading the output every week when nothing appears to be wrong. Get those three right and the rest is setup.