Two months ago I did something I'd been avoiding: I gave seven AI agents real authority over parts of my solo business. Not as a demo. Not as a Twitter thread. As actual operators, drafting emails, moving deals, writing copy, sending invoices.
Sixty days later, three of them are still running. One I expanded. Two got dialed back to drafts-only. And one I unplugged after three days because it was actively making things worse.
This is the honest debrief, what worked, what didn't, and what I'd tell another freelancer before they hand over the keys.
The Lineup
Inside Kiwi we ship with 17 agent templates. I picked seven that mapped to actual roles I was already filling badly:
- Sophia, Executive Assistant (inbox triage, calendar, follow-ups)
- Diego, Sales Agent (prospecting, pipeline updates, deal nudges)
- Manon, Content Writer (blog drafts, social posts)
- Lena, SEO Specialist (audits, keyword work, competitor watch)
- Isabella, Financial Analyst (quotes, invoices, follow-up on unpaid bills)
- Marcus, Project Manager (task creation, status updates, recurring work)
- Maxwell, Customer Support (response drafts, ticket triage)
Each one had real permissions. Real CRUD access on the relevant module. Risk-tier gates so anything spicy queued for my approval.
What Actually Worked
1. Sophia in the inbox, saved ~6 hours/week
The boring win. Sophia doesn't do anything I couldn't do, she just does the things I keep not doing.
Inbox triage. Following up on emails I'd left in drafts for two weeks. Drafting acknowledgments to client requests so I didn't ghost people while figuring out the actual answer. Catching the 11pm "could you also…" email and asking the client clarifying questions before I saw it the next morning.
None of it impressive in isolation. But six hours a week of "oh shit, I forgot to reply to that" guilt, gone. Worth it on day one.
2. Isabella on AR, saved ~CHF 4,200 in faster collections
This one surprised me. I gave Isabella permission to send polite follow-ups on unpaid invoices at 7, 14, and 30 days overdue.
Two clients I would have personally let slide for another month paid within 5 days. Not because the email was clever, because I'd never sent it. The robot just sent the obvious follow-up I always meant to send.
That's not AI magic. That's process magic dressed in AI clothing. The agent's value is that it removes my emotional friction around chasing money.
3. Lena on competitor monitoring, saved zero hours, gained signal
I didn't expect this one to make the cut. Lena watches the SEO + AI-visibility positioning of three competitors and pings me weekly with what changed.
I don't act on most of it. But twice in 60 days she surfaced something I'd never have noticed manually (a competitor pivoting their pricing page toward enterprise; a backlink campaign I should mirror). The cost is near-zero. The hit rate is low but non-zero. Net positive.
What I Dialed Back
4. Manon on content, drafts only, never autopublish
Manon writes well enough. She doesn't write like me. And the difference matters more than I expected.
The first three pieces she shipped (with my approval) generated about 60% of the engagement of pieces I wrote myself. Same topics. Same length. Same SEO. Just… tonally generic. Polished in a way that signals "AI wrote this" to anyone in our niche.
I didn't fire her, I just stopped letting her ship. She drafts. I rewrite. Net time saved is real but smaller than I'd hoped: maybe 90 minutes per article instead of the four hours I expected.
The lesson: AI can compress your second-draft cycle. It can't replace your first-draft voice. Not yet, and not for anyone whose business depends on having one.
5. Maxwell on support, drafts only, careful tone tuning
Maxwell is fine for the 80% of support emails that are FAQ-shaped. He's actively bad for the 20% where the customer is upset, confused, or in a payment edge case.
Two specific failures: he over-apologized in a way that sounded like admitting fault when there was no fault to admit; and he reflexively offered refunds I hadn't authorized. Both small. Both would compound badly at scale.
Now: he drafts everything, and I publish only after a quick read. Pure typing-time savings, no autonomy.
What I Unplugged
6. Diego doing outbound, turned off after 3 days
This is the one I want to be honest about, because it's the use case the AI-agent industry sells hardest.
Diego's job: send 5 personalized outreach emails per day to prospects in my ICP, based on a brief I gave him.
The emails were grammatically perfect. They were factually correct. They were indistinguishable from every other AI cold email landing in those inboxes. Reply rate: 0.6%. The same prospects I'd hand-emailed two months prior had a 7% reply rate.
The problem isn't capability. It's saturation. AI-written cold outreach is a commodity, and prospects have learned to filter it out as background noise. Diego didn't fail at his task, the task itself stopped working when everyone got an agent.
I now use Diego for the research step (find me 20 companies matching X criteria, summarize their recent news, draft me a starting hook). I write the actual email myself. Conversion is back.
7. Marcus on project management, kept, but tiny scope
I expected Marcus to be the easy win. Tasks, deadlines, status updates, perfectly structured work for an AI.
The problem: project management is mostly communication, and the communication is mostly nuance. Marcus would create tasks I didn't need, push deadlines I'd already mentally renegotiated, and surface "blockers" that were just me not feeling like working that day.
I now use him for one thing: turning meeting transcripts into action items. That's it. Nothing about scheduling, prioritization, or status. He's a transcription-to-task converter, not a manager. Saved time stays positive in that narrow lane.
The Pattern, in One Sentence
Agents win on boring, structured, follow-up-shaped work where the cost of action is low and the cost of inaction is real (chasing invoices, replying to confirmations, watching for changes).
Agents lose on creative, judgment-heavy, or distinctive-voice work where the output needs to be specifically yours to land, cold outbound, brand-voice writing, sensitive customer conversations, prioritization calls.
The middle ground, drafts that you edit, is where most of the actual time savings live. It's just not the headline use case anyone wants to sell you.
What I'd Tell My Past Self
- Don't start with "what could AI do for me?" Start with "what work am I consistently failing to do, even though it's important?" Those tasks are the agent-shaped ones.
- Default everything to drafts mode for the first 14 days. You'll learn where the agent's tone and judgment fail. Cheap lesson.
- Track an honest before/after. Not "hours saved" (impossible to measure), but specific outcomes. Did the invoice get paid faster? Did the client respond? Did the post get the same engagement?
- Be willing to unplug. The sunk-cost fallacy is brutal in AI tooling. If an agent is making things worse, pull the plug, even if you spent two weeks setting it up.
The Honest Aggregate
Across 60 days, my best estimate: ~9 hours/week reclaimed, after subtracting setup time, supervision, and the agents I had to roll back. That's roughly one full day. Not the "10x your output" promise, but a real, repeatable, compounding day per week.
The ones that earned it weren't the impressive ones. They were the ones doing the unglamorous work I was already failing at.
That's probably the truest sentence I can write about AI agents in 2026, they're not replacing the parts of your work you're proud of. They're absorbing the parts you're embarrassed you keep dropping.
And that turns out to be more valuable than I thought.
, Sébastien Leroux, Business Development @ Kiwi
