The 20-Person Company Math: Per-Seat Is Wrong Even at Small Scale
A 20-person company — say a growing professional-services firm, an e-commerce brand, or a B2B SaaS startup — looks at the AI tools and sees per-seat bills everywhere:
- ChatGPT Team: $25/user/month × 20 = $500/month ($6K/year)
- Microsoft 365 Copilot: $30/user/month × 20 = $600/month ($7.2K/year)
- ChatGPT Plus + Custom GPTs: $20/user × power users only — still per-head
- Specialized SMB AI (customer support, sales, marketing tools): $30–80/user × headcount
In a 20-person company, the per-seat math hides under "looks affordable." But it's still the wrong shape. Most of those seats are bought for occasional use — the founder's daily ChatGPT habit, a couple of analysts doing weekly research, the marketer drafting copy on Thursdays. Two or three power users generate 80% of the actual AI work. The other 17 seats are insurance against future use that mostly doesn't happen.
A flat-rate, usage-based, or self-hosted alternative costs a fraction. And for the customer-support, lead-qualification, and bookkeeping workloads that actually move the needle for small business, it scales without a per-employee surcharge as the team grows.
The math is the post.
What the Latest Models Actually Cost in 2026
Token pricing across the major providers, approximate as of mid-2026:
| Model | Provider | Input ($/MTok) | Output ($/MTok) | SMB sweet spot |
|---|---|---|---|---|
| Claude Haiku 4.5 | Anthropic | $1 | $5 | Customer-support Q&A, ticket triage |
| Claude Sonnet 4.6 | Anthropic | $3 | $15 | Sales follow-ups, content drafts |
| Gemini 3 Flash | $0.35 | $1.05 | Cheapest high-volume routing | |
| GPT-5 mini | OpenAI | ~$1.50 | ~$6 | General-purpose mid-volume tasks |
| Llama 4 (self-hosted) | Meta (open weights) | ~$0 | ~$0 | Flat-rate $50–200/mo VPS for the whole company |
For a small business, the frontier models (Opus, GPT-5) are usually overkill for the volume — the work is concentrated in Haiku / Sonnet / Flash territory.
A Real Workload: Customer-Support Automation for an E-Commerce SMB
A 20-person e-commerce company handles roughly 5,000 customer-support tickets per month — order status questions, return requests, product questions, sizing help, shipping issues. About 60% can be answered by an AI agent with access to the order database and a small product knowledge base; 40% need a human. For a deeper per-ticket cost breakdown — including a side-by-side against Intercom Fin ($0.99/conversation), Ada, Drift, Ultimate.ai, and Forethought at SMB / mid-market / enterprise scale — see What AI Customer Support Actually Costs in 2026.
The AI handles ~3,000 tickets per month, averaging 600 input tokens (ticket + context) and 400 output tokens (drafted response). That's 1.8M input + 1.2M output tokens per month — a workload that costs less than a single per-seat license at frontier-model pricing.
What it costs by deployment shape
| Deployment | Pricing shape | Monthly cost | Annual | Scales with… |
|---|---|---|---|---|
| ChatGPT Team | Per-seat ($25/user × 20) | $500 | $6,000 | Headcount (linear) |
| Microsoft 365 Copilot | Per-seat ($30/user × 20) | $600 | $7,200 | Headcount (linear) |
| Specialized AI customer-support tool | Per-conversation ($0.10–0.30/ticket) | $300–900 | $3,600–10,800 | Ticket volume (and quickly overpriced) |
| Direct API — Claude Haiku 4.5 | Token-based | ~$8 | ~$96 | Actual work |
| Direct API — Gemini 3 Flash | Token-based | ~$2 | ~$24 | Actual work |
| ibl.ai self-hosted (Llama 4 on VPS) | Flat license + small VPS | ~$100–250 | ~$1,200–3,000 | Whatever you put on it (flat) |
The ibl.ai row for an SMB doesn't need an H100 — a small VPS ($20–50/month) running a quantized open-weight model handles the entire customer-support workload, plus the founder's daily ChatGPT-style chat, plus the marketer's content drafts, plus sales follow-ups — all on the same flat-rate infrastructure with no per-seat fee.
Why Per-Seat Pricing Fails Even at 20 People
Three reasons:
1. Usage is hyper-concentrated in a small company. In a 20-person org, two or three people use AI heavily, five or six use it weekly, the rest barely. Per-seat invoices the whole 20.
2. Growth penalizes you twice. Hire 5 more people and the per-seat bill goes up linearly even if those new hires don't touch AI for their first three months. Usage-based or flat-rate stays flat as the team grows — until the work grows.
3. The cost-savings story unwinds. Small businesses adopt AI to do more with less. A per-seat bill that scales with headcount is the opposite of that — every new hire makes AI more expensive instead of more useful.
What Stays the Same, What Changes
A small business doesn't have an IT department. The ibl.ai option for SMB is the same managed platform — chat UI, agent dashboards, model routing, integration with Gmail / Slack / QuickBooks / Shopify — without the per-seat charge. The runtime can run on a small VPS the business owns (typically $20–50/month) or in a managed cloud the business doesn't have to think about.
What disappears: the per-seat line item that grows with headcount. What appears: a flat-rate AI capability the company controls, with the model choice the workload actually needs (Haiku for high-volume customer-support, Sonnet for sales follow-ups and content, an open-weight model for the bulk routine work).
Run the Numbers for Your Business
Two interactive tools sized for small business:
- AI Cost Calculator — Small Business — compares flat-rate vs per-seat subscriptions for your headcount and workload
- AI Readiness Assessment — Small Business — 5-question quiz to score your readiness to deploy AI agents
For the eight most-deployed SMB agents (customer support, sales, bookkeeping, social media, scheduling, hiring, inventory, website management), see Claw Agents Small Business: 8 AI Agents for Growing Companies.
For connecting your existing SMB tools — QuickBooks, point-of-sale, CRM, scheduling — to your AI agents via MCP, see MCP Guide — Small Business.
Why Family-Owned and New York Matters Here
A small business doesn't need an investor exit clock dictating its software roadmap. ibl.ai is family-owned and operated from New York, NY — a long-term partner that doesn't price its products like a VC-funded growth machine. The runtime is open source. The business owns the configuration. The math works at 5 employees or 500.