A conversation I've had 8 times this quarter. Different company. Same sentences.
Founder of a 40-person B2B SaaS, roughly $6M ARR, leans in on the Zoom and says:
"Our biggest competitor has the same ARR. Same ICP. They run with 22 people. They're hiring nobody. They're pulling ahead on every inbound metric. I can't figure out what they're doing."
I can. I've shipped their stack for 4 different clients this year.
It's not a secret. Two or three AI agents wired into n8n. Three tripwires per agent. One human owner per agent. Total cost: around $600 a month. A cheaper line item than your Salesforce seat count.
If you run a 20 to 200 person company, you are not competing against Klarna or Midjourney. You are competing against your same-size peer who quietly installed this stack 6 months ago and is now posting 2x your revenue-per-head without raising a round.
The gap is already open. By Q4 2026, roughly half your segment will have shipped their first production agent. The other half will be raising awkward bridge rounds. This is the quarter the split happens.
This issue: what peer-scale teams are actually shipping right now, the exact stack and prompt I'd have you run in 30 days, and the failure mode I got burned by so you don't have to.
The scar I'm not letting you repeat
First, the honest part.
Week 2 of an agent I shipped last year, the thing hit a low-confidence case and didn't flag it. Output looked plausible. It wasn't. A human caught it 6 hours later because of a cost-ceiling tripwire, not the confidence tripwire I'd built. If the cost ceiling hadn't fired, that mistake would have compounded for 3 days before anyone looked.
Lesson paid for in production: never trust a single guardrail. Every agent I ship now has three minimums or it does not ship.
Confidence threshold. LLM returns confidence < 0.8 → route to human.
Cost ceiling per run. Token spend > $X → halt + alert.
Human-in-loop on errors. Any error branch pages the human owner in Slack within 60 seconds.
⠀That's the whole moat between "our agent works" and "our agent embarrassed us in front of our biggest account."
If you're about to ship your first agent and don't have these three written down, stop reading, fix that, then come back.
Peer-scale receipts
Not Klarna. Not Midjourney. Companies at your weight class.
6-person marketing agency Self-hosted n8n under $25 per month. Saved 20+ hours per week across the team on lead handling, follow-up emails, and internal updates.
18-rep B2B distributor Deployed agents to strip data entry out of the sales flow. Rep meeting bookings +40% in 90 days. The reps didn't work more hours. They stopped doing work that was never theirs.
8-loan-officer mortgage brokerage Was burning 120 admin-hours per week across the team. Blended cost $35/hr. That's $218,400 per year in pure admin labor. One agent deployment took 70% of those tasks off the plate. $152,000 saved annually. One agent. Sub-$200/month run cost.
Unbabel (growth-stage, 400-ish) Used n8n as the spine. Cut manual operational work 51%.
The SMB base, in aggregate
U.S. Chamber of Commerce 2026 SMB survey: 91% of companies deploying AI agents report positive ROI.
50 to 249 employee businesses are adopting AI at 2x the rate of microbusinesses. The middle of the ICP curve is moving fastest. You are exactly in that curve.
Humans collaborating with AI agents produced 73% higher productivity per worker than humans collaborating with other humans. Read that twice. Adding one agent beats adding one hire on output per person.
Sales teams save 3 to 6 hours of admin per rep per week. Ops teams cut cycle times 20 to 40%. Support hits 30 to 50% ticket deflection.
AI agents start at $20/month. The SDR you've been trying to hire for 4 months costs you $8–12K a month loaded. The replacement-grade agent costs you a latte a day.
⠀
The pattern scales in both directions (this is the part your board needs to hear)
Klarna runs 853-agent-equivalent volume at 7,000+ headcount. Midjourney hit $200M ARR with 11 people. Those are the ceiling references.
The floor looks different and it looks like you. An 8-officer mortgage broker saves $152K with one agent. A 6-person agency saves 20 hours a week for under $25 a month.
Same pattern at both ends. n8n as the spine. One Claude or GPT-5 call at the judgment step. Three tripwires. A human owner per agent.
Here's the piece most 20 to 200 person leaders miss. Your marginal leverage is higher than Klarna's, not lower. A Fortune 500 can absorb a bad hire. You cannot. A well-built agent on your team replaces the hire you couldn't afford, never takes PTO, and costs 1/20th of the fully-loaded salary. The ratio of leverage to headcount-at-risk is best at your scale.
You are not too small for this. You are the scale where the math compounds fastest.
The stack that keeps working
I've shipped 20+ of these for North American teams. Boring on purpose. Boring is what survives week 2 in production.
Judgment prompt (copy this, edit the bracketed pieces):
You are a lead qualifier for [YOUR COMPANY], B2B [YOUR ICP].
Given the discovery notes below, return ONLY this JSON:
{
"fit_score": 0-100,
"confidence": 0.0-1.0,
"top_objection": "one sentence",
"suggested_next_step": "book_call | nurture | disqualify | human_review",
"reasoning": "≤50 words"
}
If confidence < 0.8, suggested_next_step MUST be "human_review".
If the notes reference [legal review | procurement freeze | RFP],
suggested_next_step MUST be "human_review" regardless of fit_score.
Notes:
{{discovery_notes}}Five nodes. One prompt. Three tripwires. This is what the 18-rep distributor and the 8-officer broker actually run. Nothing fancier. Nothing built on a research-paper framework. This is the whole shape of the thing.
Why most 20–200 person teams still won't ship this year
Three anti-patterns I hear on discovery calls every single week. Tuned for your scale:
"We'll hire a Head of AI first." No you won't. The role doesn't exist at your size, and every operator you'd want is already at a Series C or running their own agency. Fix: pick one existing team lead. Give them one agent to own. Ship in 30 days. Revisit hiring in Q3.
"We need clean data first." No you don't. You need one use case where your current data is good enough. Inbound lead qualification does not require a data warehouse. A CRM record plus 200 words of discovery notes is enough for an agent to fit-score.
"Let's wait for Claude Opus 5 / GPT-6." The version of you who said this in 2024 is now 18 months behind the peer who shipped on GPT-4. Your Q4 2026 self will thank your Q2 2026 self for not waiting again.
Tool of the week
n8n's Evaluations capability + a 10-example agent scorecard.
You don't need a data team. Pick one role. Pull 10 real past outputs from that role: 10 leads the team qualified, 10 tickets they resolved, 10 briefs they wrote. Score the agent against those 10 human baselines every Monday.
80% on volume + 95% on accuracy = the agent is hired. Below that = it is a science experiment and you kill it.
Two hours of setup. Ends 6 weeks of "is this thing actually working?" in standup.
The ask
If you run a 20 to 200 person company and you've been stalling on this for 90+ days, reply to this email with one sentence:
Which role on your team would be the easiest to hybridize first?
I'll pick 3 replies this week and send back a 1-page architecture teardown. Exact n8n nodes. Handoff rules. Eval loop. Prompt scoped to your ICP and your current stack. Free. No pitch attached.
I do this openly because the founders I pick tend to come back as clients in Q3 anyway. No reason to hide the work. You can steal the teardown, run it yourself, and never talk to me again. That's fine too.
What I'd build this week if I were you
Budget: $600/month and one existing team lead's attention for 3 weeks. That's the whole spend.
Pick your single most expensive repetitive process. For most 20–200 person teams that is:
Inbound lead qualification (sales)
Content repurposing long-form → LinkedIn/X (marketing)
First-touch support triage (customer success)
Weekly metrics rollup for the leadership standup (ops)
⠀Ship the 5-node stack above. Use the prompt template. Confidence threshold 0.8. Cost ceiling $5 per run. Put your best ops-brained team lead on it as the owner.
Ship Monday. Review Friday. That's your first hire.
They don't ask for equity. They don't need health insurance. They work weekends. They cost $20 a day.
And your 22-person competitor already has three of them on the roster.
— Jet