How Much Does an AI Agent Actually Cost Your Business Each Month?

A live AI agent for a small business usually costs a few hundred dollars a month once you add seats, tokens, retries, babysitting hours, tool or MCP fees, and logging — not the $40 API invoice people screenshot. The cash vendors are the small slice. The hours you spend approving drafts are the line that actually moves the total.

I am William Spurlock, founder, AI Systems Architect, and Fractional AI CTO. I have built 600+ automations with 500+ still live, spent 20,000+ hours on agentic systems, and tracked 35,000+ hours saved for clients — that last number is aggregate busywork deleted across the book of work, not my personal hours. I hold the Make.com AI Automation certifications and have worked directly with the n8n team. None of that lets me invent your invoice. I will not.

Today is September 8, 2026. This page owns one ledger: the monthly run-rate for one live agent. It is not the payback test and it is not a shut-off essay — that is when an AI agent pays for itself and when it quietly burns cash. It is not hiring versus a whole automation program — that is what AI automation actually costs in 2026. If you still need the worksheet before you build anything, use the ROI of AI automation pillar. I am pricing the bill you will see in month three if the agent stays on.

Vendor list prices move. Every dollar below is a published rate I re-opened for this post, a dated government series, or a hedged Estimate. Re-open the vendor page the week you budget.

---

How much does it actually cost to run an AI agent for a small business each month? #

Most small-business agents I meter land in a planning band of about $150–$600 per month for one bounded job, once babysitting hours sit on the same sheet as tokens. A clean inbound-lead or ticket-label agent on a workhorse model can sit near the bottom of that band on cash vendors alone. The same agent jumps toward the top the week you keep rewriting drafts, or you park GPT-6 Astra on every form submit.

That band is a planning Estimate, not a quote and not a client invoice. Volume, model ID, retry rate, and how long you stare at the approval queue decide where you land.

I split the month into six lines so the API total cannot hide the rest:

Line	What you are paying for	Where the number comes from
Seats	The host and any human workspace this agent requires	Vendor plan page (n8n, ChatGPT Business, Claude)
Tokens	Input, output, cache, thinking	Model list price × tokens you actually sent
Retries	History replayed when a tool fails or the model tries again	Same token rates × extra turns
Babysitting hours	Approve, rewrite, re-prompt, explain the bot	Your loaded hourly rate × minutes logged
Tool / MCP fees	Search, scrape, extra API overages this agent calls	Firecrawl, search, CRM overage, hosted MCP
Logging	Traces you keep so you can see why a run cost $2	LangSmith, Helicone, or n8n execution retention

If you only export Anthropic, OpenAI, Google, or xAI at month-end, you are reading one line. I want all six.

My opinion, held loosely: if you cannot name those six lines for one agent, you do not have a monthly cost. You have a vibe.

---

Which monthly line items belong on one live agent? #

Only the seats, tokens, retries, hours, tools, and logs that exist because this agent is running. Do not dump the whole studio onto this sheet. Cursor seats you bought for site work do not belong here. The ChatGPT Business workspace you bought for the sales team does not belong here unless this agent is the reason those seats exist.

A run is one goal → tool loop → done or killed. A retry is the same run going around again. Babysitting is the standing labor of keeping the agent honest. I am not pricing cleanup-after-a-bad-send as a seventh line on this page — that is a failure reserve, and it lives on the payback post. This page is the bill you should expect in a normal month.

What I put on the sheet:

• One workflow host for this agent (I usually use n8n, the open-source workflow automation platform)
• One workhorse model on the hot path, plus a cheaper model for labels if I split the job
• Token volume × list price, then a retry adder
• Minutes a named human spends in the approval queue
• Every paid tool this agent calls through Model Context Protocol (MCP) or a direct API
• Enough logging to explain a fat run without buying an observability stack you will not open

What I leave off:

• Build hours from month zero (amortize those on the ROI pillar, not on month-three run-rate)
• Seats for people who never touch this agent
• A second agent's tokens
• "AI value" or brand-lift fiction

If two agents share one n8n Cloud project, I still put the Starter plan on the noisier agent's sheet, or I split the 20€. I do not pretend the host is free because "we already had n8n."

---

How much do seats cost when you only count this agent? #

Seats are the flat lines: a workflow host, and any chat workspace that is actually the runtime. For most small-business agents I ship, the host is n8n and the brain is an API key. A ChatGPT Business seat only belongs here if the agent lives inside that workspace.

I re-opened n8n's pricing page for this post. On the annual toggle, Starter is 20€ per month billed annually (2,500 workflow executions, 5 concurrent, 1 shared project, 7-day execution log retention). Pro is 50€ per month billed annually (10,000 executions, 20 concurrent, execution search, 30-day log retention on the comparison table). Those are published EUR stickers, not USD. Convert at the rate your card uses and label the dollars Estimate.

Seat	When it belongs on this agent	Published sticker I checked	How I treat it
n8n Cloud Starter	The agent is a webhook or schedule in n8n	20€/mo billed annually	Count it; FX is Estimate
n8n Cloud Pro	You blow 2,500 executions or need execution search	50€/mo billed annually	Upgrade line, not a vanity upgrade
n8n Community (self-hosted)	You already run a box	$0 license	Host + your time; do not call it free
ChatGPT Business	The agent is a Custom GPT / workspace runtime, not an API loop	Confirm OpenAI's live Business checkout (multi-seat minimum)	Only if this agent is why you bought the seats
Claude workspace seat	A human reviews drafts inside Claude, and that seat exists for this agent	Confirm claude.com/pricing	Same rule: this agent or leave it off

n8n bills executions, not steps. One inbound-lead run that hits eight nodes is still one execution. That is why Starter can hold a weekday agent at 50 runs/day: 50 × 22 weekdays ≈ 1,100 executions, inside 2,500. A chatbot that bills one execution per user message will eat Starter faster. Count the trigger, not the node count.

I do not put a Cursor seat on an inbox agent. I do not put a Grok Bot teammate seat on this sheet unless that teammate is the agent. One job. One host. One reason each seat exists.

---

How should you price tokens and retries for one month? #

Price tokens as list rate × tokens you sent, then add retries as extra turns that resend the transcript. Output is the line people notice. Replay is the line that grows. A flagship ID on a high-volume loop is how a "cheap agent" becomes a $200 model week you cannot explain.

I re-opened the official model pages for September 8, 2026. Standard tier, per million tokens, before batch or regional uplifts:

Model	Job I give it	Input / MTok	Output / MTok	Official page
Claude Sonnet 5	Default agent loop, drafts, tool use	$2	$10	Anthropic pricing
Claude Opus 5	Messy judgment, ugly contracts	$5	$25	Same Anthropic page
Claude Haiku 4.5	Labels, routing, cheap glue	$1	$5	Same Anthropic page
GPT-6 Astra (gpt-6-astra)	Hard OpenAI / computer-use steps, not the hot path	$10	$50	OpenAI GPT-6 Astra
Gemini 3.8 Flash (gemini-3.8-flash)	Volume language jobs through Dec 31, 2026 intro	$0.75	$3.75	Gemini API pricing
Grok 4.6 (grok-4.6)	Grok-native loops under 200k prompt tokens	$2	$6	xAI Grok 4.6

Anthropic's pricing page also states that Sonnet 5's $2 / $10 rate, first sold as an intro price through August 31, 2026, is now the standard price and the planned September 1 jump to $3 / $15 will not happen. Claude Opus 4.8 is a Fable fallback on my board, not the flagship I price against. I price Claude Opus 5 at $5 / $25.

GPT-6 Astra adds a long-context cliff: prompts over 272K input tokens reprice the full request at 2× input/cache and 1.5× output. Fast mode is 2×. Batch and Flex are half. Grok 4.6 has its own cliff: once the prompt hits 200k tokens, the whole request moves to $4 / $12. Gemini 3.8 Flash intro is $0.75 / $3.75 through December 31, 2026, then $1.50 / $7.50 on January 1, 2027. Those are published clocks. They are not reasons to put Astra on spam tags.

A labeled planning example — Estimate, not a client result. One inbound-lead agent, Claude Sonnet 5, about 8 tool steps, ~32,000 input tokens and ~1,200 output tokens on a clean run.

Math	Number	Label
Weekday runs	50/day × 22 days = 1,100	Estimate
Clean input	32,000 × 1,100 × $2 / 1M ≈ $70	Estimate at published $2
Clean output	1,200 × 1,100 × $10 / 1M ≈ $13	Estimate at published $10
Clean token month	~$84	Estimate

Same volume on Claude Opus 5 is about 2.5× ($5 / $25 vs $2 / $10) → ~$210 Estimate. Same volume on GPT-6 Astra at $10 / $50 → ~$418 Estimate before Fast mode or the 272K cliff. Same volume on Gemini 3.8 Flash intro → ~$31 Estimate. Same volume on Grok 4.6 under 200k → ~$78 Estimate. Same classify step on Haiku 4.5 is pocket change. Pick the ID on purpose.

Retries are not "one extra completion." Most frameworks replay the transcript. If 25% of runs retry once and replay ~32,000 input tokens:

275 × 32,000 × $2 / 1M ≈ $18 plus a little extra output. I budget ~$20 Estimate on a messy-but-normal month for this sample. A week of empty-search loops can double that. Cap identical tool calls. Cap tokens per run. Do not "try once more" into a void and then act surprised at the invoice.

---

How do you price babysitting hours on the monthly bill? #

Babysitting is operating cost. Price it at a loaded hourly rate times the minutes you actually spend in the queue. If you skip this line, you will swear the agent costs $80 while you became its unpaid ops lead.

The U.S. Bureau of Labor Statistics Employer Costs for Employee Compensation for March 2026 puts private-industry total compensation at $46.60 per hour worked — $32.60 wages and $14.01 benefits. That is a dated public floor for "this hour was not free," not a quote for your shop. Use your bookkeeper, VA, or founder rate. The June 2026 ECEC series was still pending as of this writing; I am not going to pretend a newer dollar exists on September 8.

Queue habit	Minutes I log (hedged)	At $46.60 (BLS, Mar 2026)	At $120 founder opportunity
45-second approve × 40 drafts	~30 min	~$23	~$60
4-minute rewrite × 12 drafts	~48 min	~$37	~$96
Two "did the bot send this?" Slack threads	15–30 min	~$12–$23	~$30–$60
A 20-minute daily habit × 22 days	~7.3 hours	~$342	~$880

Those minute ranges are planning Estimates from queues I have sat in, not a named-client timesheet. Twenty minutes a day does not feel like a job. At month-end it is the largest line on the sample bill below.

I count babysitting when a human:

• Opens the approval queue
• Rewrites a draft before send
• Re-prompts because the tool auth died
• Explains the agent to a teammate
• Reviews the weekly log

I do not count staring at Slack because you are bored. I do count the 4 p.m. "just checking." If median approve time is still four minutes in month three, the agent is a first-draft button. Keep it if the first draft still saves time. Put those minutes on the bill either way.

---

What do tool, MCP, and logging fees add each month? #

Tools and logs are small until the agent reads the open web on every turn, or you buy an observability plan you will not open. First-party CRM and mail APIs are often already on the books. Hosted scrape, search, and trace products are the adders.

Model Context Protocol (MCP) is Anthropic's open standard for exposing tools to a model over a persistent connection. The protocol is not a subscription. The servers you attach can be. Fat tool schemas also tax tokens — that cost shows up on the model line, not here.

Tool stickers I re-opened:

Tool	When this agent pays it	Published sticker	Source
First-party CRM / mail API	Read/write your own records	$0 extra if you already pay the SaaS	Your existing stack
Firecrawl Hobby	The agent scrapes or searches public pages	$19/month, or $16/month billed annually (5,000 credits, 5 concurrent)	Firecrawl pricing
Firecrawl Free	Under ~1,000 credits / month	$0, 2 concurrent	Same page
Gemini grounding / search	Google Search tool on Gemini calls	Confirm the live Gemini pricing page (free bucket, then a per-1,000 adder)	Gemini API pricing

Firecrawl scrape is 1 credit per page on that page. Search is 2 credits per 10 results. JSON format adds credits. If the inbound-lead agent only reads your CRM, this line is $0. If it "just checks the company's homepage" on every new lead, Hobby at $19 is the honest card. Standard at $99/month is a different agent.

Logging stickers I re-opened:

Logger	Published floor	What you get	Source
n8n Cloud Starter logs	Included in the 20€ plan	7-day execution retention, 2.5k saved executions	n8n pricing
LangSmith Developer	$0 / seat, then pay-as-you-go	1 seat, 5,000 base traces / month	LangSmith pricing
LangSmith Plus	$39 / seat / month	Unlimited seats, 10,000 base traces / month, then LCU $1.50 / LSU $1.00	Same page
Helicone Hobby	$0	10,000 requests / month, 1 GB, 1 seat, 7-day retention	Helicone pricing
Helicone Pro	$79 / month	Unlimited seats, then usage-based	Same page

For one inbound-lead agent at ~1,100 runs, Helicone Hobby or LangSmith Developer often covers the month if you log one trace per run. I do not buy Helicone Pro or LangSmith Plus for a single small-business agent unless I need the extra retention, seats, or deployment SKU. n8n's seven-day log is enough to debug last Tuesday. It is not a 90-day audit trail.

My rule: log enough to explain a fat run. Do not subscribe to a $79 or $39 plan because a launch video said "observability."

---

What does a sample monthly bill look like for one inbound-lead agent? #

Here is one labeled month for one agent: inbound leads → draft reply in a queue. No auto-send. Claude Sonnet 5 on the loop. n8n Cloud Starter as the clock. Firecrawl Hobby only because the agent reads a public page per lead. Logging stays on Helicone Hobby / n8n's seven-day executions. This is a planning sheet, not an invoice I sent a client.

Assumptions I am willing to defend as Estimates:

• 50 weekday runs, 22 weekdays → 1,100 runs
• Clean run ≈ 32,000 input / 1,200 output on Sonnet 5
• 25% of runs retry once
• A named human spends 20 minutes/day in the queue
• Loaded rate $46.60/hr (BLS private-industry, March 2026)
• EUR→USD on n8n at a rough card rate

Line	Basis	Amount	Label
n8n Cloud Starter	20€/mo billed annually (n8n.io/pricing)	~$23	Estimate (FX on a published EUR sticker)
Tokens (Claude Sonnet 5)	1,100 clean runs × ~32k / 1.2k at $2 / $10	~$84	Estimate (volume) at a published rate
Retries	~275 extra transcript replays	~$20	Estimate
Firecrawl Hobby	Public-page reads (firecrawl.dev/pricing)	$19	Official monthly list
Logging	Helicone Hobby $0 / 10k requests, or n8n 7-day logs	$0	Official floor if you stay under the cap
Cash vendors subtotal	Seats + tokens + retries + tool + logs	~$146	Estimate
Babysitting	7.3 hours × $46.60	~$342	Estimate (hours) at a dated BLS rate
Full monthly run-rate	Cash + babysitting	~$488	Estimate

Swap the model, keep the volume:

Hot-path model	Token + retry Estimate	Full run-rate with the same $342 babysitting
Gemini 3.8 Flash intro	~$31 + ~$8 ≈ $39	~$423 Estimate
Grok 4.6 under 200k	~$78 + ~$18 ≈ $96	~$480 Estimate
Claude Sonnet 5 (sample)	~$84 + ~$20 ≈ $104	~$488 Estimate
Claude Opus 5	~$210 + ~$44 ≈ $254	~$638 Estimate
GPT-6 Astra	~$418 + ~$88 ≈ $506	~$890 Estimate

The seat, the scrape card, and the human minutes barely move when you change models. The token line does. That is why I will not put Astra on this job. Astra is a named flagship step. This job is a weekday loop.

What I want you to take from the table:

1. The API screenshot is not the bill. $84 of Sonnet is not $488 of run-rate.
2. Babysitting is the swing line on a small-business agent that still needs a human gate — and it should still need a human gate.
3. Tools only show up when the agent leaves your own APIs. Delete Firecrawl and the cash subtotal drops by $19. Leave it on because "research feels smarter" and you bought a hobby plan for vanity.
4. Logging can be $0 at this volume. I still log. I do not upgrade the logger to feel professional.

If your agent is a support-labeler on Haiku 4.5 with no web scrape and a 10-second approve, the cash month can sit under $50 Estimate and babysitting can sit under $100 Estimate. If your agent is an always-on researcher on Astra with unbounded search, I will not put a pretty total on this page. That is a different animal. Cost that agent as itself.

Print this table. Put your run count in the token row. Put your minutes in the babysitting row. That is the monthly number I will argue about on a strategy call. I will not argue about a screenshot of the Anthropic usage tab.

---

Frequently Asked Questions #

How much does it actually cost to run an AI agent for a small business each month? #

A bounded small-business agent usually lands in a planning band of about $150–$600 per month once seats, tokens, retries, babysitting hours, tools, and logs sit on one sheet. Cash vendors alone can sit near $80–$160 on a workhorse loop. The hours in the approval queue are what push you toward the top of the band. Those totals are Estimates, not a quote.

What is usually the largest line on a live AI agent bill? #

Babysitting hours are usually the largest line on a draft-and-approve agent, and they never appear on the model invoice. At the March 2026 BLS private-industry rate of $46.60 per hour, a 20-minute daily queue is about $342/month Estimate. Tokens on Claude Sonnet 5 for the sample volume above are about $84 Estimate.

Do I need a ChatGPT Business seat to run an AI agent? #

No. An API loop in n8n does not require a ChatGPT Business seat. Add that workspace only if the agent is a ChatGPT Business runtime, and only count the seats this agent forced you to buy. Confirm the live OpenAI Business checkout — I am not pasting a stale Team-era sticker as if I opened it today.

How much do retries add to an AI agent's monthly token bill? #

Retries add a planning surcharge I often see around 15–40% of the clean token line on a messy-but-normal month, because the transcript is replayed. In the sample, a 25% single-retry rate added about $20 Estimate on Sonnet 5. Soft loops with empty search results can exceed that without a red error in n8n.

Should babysitting hours count as part of the monthly agent cost? #

Yes. If a human has to approve, rewrite, or explain the agent, those minutes are operating cost. Skip them and you will compare a $23 n8n plan to a part-time hire and call it a win. Put your own loaded rate in the cell. The BLS $46.60 figure is a dated floor, not your shop.

What logging do I need for one small-business agent? #

Enough to explain last week's fat run: n8n's seven-day executions, or Helicone Hobby at $0 / 10,000 requests, or LangSmith Developer at $0 / 5,000 traces. I do not start on Helicone Pro at $79/month or LangSmith Plus at $39/seat for one inbound-lead agent. Upgrade when you hit the free cap or you need longer retention.

Can one inbound-lead agent stay on n8n Cloud Starter? #

Yes, if you stay under 2,500 executions a month. Fifty weekday runs is about 1,100 executions, inside Starter's published 2,500 cap on n8n.io/pricing. A per-message chatbot is a different count. Blow the cap and Pro is 50€/mo billed annually, not a moral failing.

Should I put GPT-6 Astra on a high-volume agent loop? #

No. GPT-6 Astra's published standard rate is $10 / $50 per million, with a 272K input cliff. On the sample volume that is about $418 Estimate in clean tokens before retries. Keep Astra for hard OpenAI or computer-use steps. Put Sonnet 5, Haiku 4.5, Gemini 3.8 Flash, or Grok 4.6 under 200k on the weekday loop.

---

If you want a number for your agent, not this sample, book an AI automation strategy call. I will cost one live job — seats, tokens, retries, babysitting hours, tools, logs — against your run count and your loaded rate. I will not turn that call into a payback essay, and I will not sell you an AIO audit. I have done this across 600+ automations with 500+ still live. The useful output is a monthly sheet you can defend, not a screenshot of an API dashboard.