Most companies that experiment with AI agents start in engineering. It makes sense — the tooling is native, the feedback loops are tight, and the output is measurable. But the interesting question isn't what happens when your dev team uses agents. It's what happens when every department does.
We run a company where agents work across engineering, design, operations, finance, and growth. Not as experiments. As staff. They have access to our systems, they produce real output, and they make mistakes that cost real time to fix. Here's what we've learned.
The org chart flattens overnight
Traditional companies have layers because humans can only manage so many direct reports. A VP manages directors who manage managers who manage ICs. Each layer exists to compress information upward and distribute decisions downward. Remove the information bottleneck and the layers lose their purpose.
When agents handle execution, one person can direct the output of what used to require a team. Not because the person works harder — because the coordination overhead disappears. The agent doesn't need a 1:1. It doesn't need context about the company's Q3 priorities whispered through three levels of management. It reads the spec and builds.
What this means in practice: our engineering "team" is one human and several agents. The human makes architectural decisions, reviews output, and handles ambiguity. The agents write code, run tests, fix linting errors, and draft documentation. The same pattern repeats in every function. One human director, N agent executors.
Finance and ops are easier than you think
People assume agents work best on technical tasks. In our experience, the highest-leverage agent deployments are in operations and finance — domains with clear rules, structured data, and repetitive workflows.
Bookkeeping follows rules. Tax deadlines are deterministic. Invoice reconciliation is pattern matching. An agent that monitors bank transactions, categorizes expenses, flags anomalies, and drafts quarterly reports doesn't need creativity. It needs accuracy and consistency — exactly what agents are good at.
The same applies to ops work like vendor management, contract tracking, and compliance monitoring. These are tasks where humans add value through judgment on exceptions, not through executing the routine. Let agents handle the 95% that follows the rules. Humans handle the 5% that requires a phone call.
Growth becomes a machine, not a grind
Content creation, prospect research, outreach drafting, analytics — growth work is a natural fit for agents because the feedback loops are measurable and the tasks are decomposable.
Our content pipeline works like this: a human provides weekly inputs — what shipped, what we're thinking about, any customer interactions worth sharing. An agent turns those inputs into blog drafts, social posts, and newsletter content. The human reviews, edits for voice, and approves. Total human time: about 45 minutes per week for a publishing cadence that would otherwise require a full-time content person.
Prospect research is similar. An agent can scan forums, track keyword mentions, identify people complaining about problems you solve, and draft personalized outreach. The human reviews the shortlist and decides who to contact. The ratio of output to human effort is absurd compared to doing it manually.
The failure modes are different per department
Engineering agents fail on ambiguity. When the spec is clear, they're excellent. When it's vague, they build confidently in the wrong direction. The fix is better specs, not better agents.
Finance agents fail on edge cases. The 95% of routine transactions are handled flawlessly. But a refund that crosses fiscal years, or a foreign currency payment with unusual fees — these require human judgment. The fix is clear escalation rules: when the agent encounters something outside its training distribution, it stops and asks.
Growth agents fail on voice. They can produce technically correct content that reads like it was written by a committee. The fix is a strong editorial layer — the human's job isn't to write, it's to ensure everything sounds like it came from a person with opinions.
The meta-pattern: agents fail at the boundaries of their domain. Not in the middle where rules are clear, but at the edges where judgment, taste, or politics matter. The org design challenge is putting humans exactly at those boundaries.
What changes about management
Managing agents is not managing people. You don't motivate agents. You don't develop their careers. You don't navigate interpersonal dynamics. What you do is:
- Define clear interfaces. What goes in, what comes out, what quality bar must be met. Agents work best with explicit contracts, not implicit expectations.
- Build review loops. Every agent output gets checked before it affects the real world. The review can be light (scan and approve) or heavy (rewrite from scratch), but it always exists.
- Monitor for drift. Agents don't get bored, but they do get stale. The prompt that worked last month might produce worse output this month because the context changed. Regular calibration is the equivalent of a performance review.
- Invest in observability. Every agent action should be logged. When something goes wrong at 3am — and it will — the audit trail is the only thing between you and a mystery.
The economics
A full-time employee in Sweden costs roughly €5,000-8,000 per month when you include taxes, benefits, and overhead. An agent doing equivalent output in a specific domain costs €200-800 per month in API calls and compute. Even accounting for the human time required to direct and review agent work, the economics are dramatic.
But cost isn't the real advantage. Speed is. An agent doesn't have a two-week notice period. It doesn't need onboarding. It doesn't take vacation. When you need to scale a function up for a product launch and back down afterward, agents make that possible without the human cost of hiring and layoffs.
The company that figures out how to run every department on agents isn't just cheaper. It's faster, more responsive, and more resilient. The humans in that company aren't doing less — they're doing different work. Higher-leverage work. The kind of work that actually requires a human brain.
Start with the boring stuff
If you're thinking about expanding agents beyond engineering, don't start with the exciting applications. Start with the boring ones. Expense categorization. Meeting scheduling. Report generation. Status tracking. These are tasks where failure is low-cost, the rules are clear, and the time savings are immediate.
Once you've built the muscle for directing agents in low-stakes domains, move to higher-stakes ones. Content creation, then customer communication, then financial operations. Each step up requires better review processes and clearer escalation rules. But the fundamental pattern is the same: define the interface, let the agent execute, review the output, improve the spec.
The end state isn't a company with no humans. It's a company where every human is a director — shaping intent, making judgment calls, and ensuring quality — while agents handle the execution that used to require headcount. That's not a future prediction. It's what we're running today.