When an AI Agent Pays for Itself — and When It Quietly Burns Cash

An AI agent pays for itself when the hours and errors it actually removes beat the full run cost: tokens, retries, failed sends, babysitting, and the human cleanup after a bad write. If you only watch the API invoice, a losing agent looks cheap for months.

I am William Spurlock, an AI Solutions Architect and Fractional AI CTO. I have built 500+ automations, spent 20,000+ hours architecting 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 a client payback number for this post. I will not.

The Aug 16 pillar — how to calculate the ROI of AI automation before you build anything — owns the workflow spreadsheet: hours saved × loaded labor, minus stack, build, and maintenance. This spoke owns a different ledger. An agent is not a 12-node n8n graph with a fixed bill per execution. It is a model in a loop. Every tool call resends history. Every retry resends history plus the error. Every "I'll just check Slack" minute is labor you never put on the invoice.

If you still need the category definition, start with what an AI agent is. If you already run ops agents, keep AI agents for operations next to this page — that post is the job list. This one is the money test.

Vendor list prices move. Every dollar below is a published rate I checked on August 25, 2026, a dated government series, or a hedged planning range. Re-open the vendor page the week you budget.

---

When does an AI agent pay for itself? #

An AI agent pays for itself when recovered labor and avoided misses exceed the run cost for 30 consecutive days, with retries and cleanup counted, and with a human still owning every irreversible send. A pretty demo week does not count. A week where you babysit every draft does not count.

I use a four-part test. Fail any one and I treat the agent as a science project, not a cost center.

Test	Pass	Fail
Job shape	Repeatable, bounded, measurable (inbox triage, lead draft, invoice chase, ticket label)	Open-ended "run the business" with no stop condition
Write surface	Drafts and read-only first; sends go through an approval queue	Live email, refunds, inventory, or payroll on day one
Meter	Tokens, tool calls, retries, failed sends, and babysitting minutes logged per run	You only see a monthly Anthropic / OpenAI / Gemini total
Payback window	Conservative hours × your loaded rate beat the full run cost inside 30–90 days	You need a heroic close rate or a made-up "AI value" line to make the math work

A run is one goal → tool loop → done or killed. A retry is the same run going around again because a tool failed, JSON came back dirty, or the model decided to "try another search." Babysitting is the time you spend prompting, approving, rewriting, or explaining the agent to your team. Cleanup is the time after a bad send: apology, CRM repair, refund, or a founder hour on the phone.

I do not need a finance model with twelve tabs. I need those four lines for one job. If you cannot fill them, you do not have an agent project. You have a ChatGPT tab with extra steps.

The jobs that usually pass the test first are the same ones I ship as scoped workflows: inbound lead drafts, invoice reminders, support labels. See the first AI automation every small business should build and manual workflows with the highest ROI when replaced by AI agents. An overnight lead desk is a good example of a bounded send surface — how to use an AI agent to handle inbound leads while you sleep — as long as booking and pricing stay behind a human gate.

My opinion, held loosely: most small shops should not start with an agent. Start with a fixed workflow. Promote the step to an agent only when the input is messy language and the output is a draft you can reject in under a minute.

---

How is agent payback different from a workflow ROI spreadsheet? #

A workflow ROI spreadsheet prices a known path. Agent payback prices a loop that can grow on every turn. Same owner, same loaded rate, different unit. If you drop an agent into the Aug 16 worksheet and treat it like a Make scenario, you will undercount the bill and overcount the hours saved.

I am not going to retell that calculator here. Use it for n8n / Make / Zapier graphs with a trigger, a few modules, and a write-back. Use this page when the model chooses the next tool.

Line	Workflow automation	Live agent
Billing unit	Execution, credit, or task	Tokens on every model call, plus tool I/O
Cost shape	Mostly flat per run	Grows as history, tool results, and retries pile up
Failure mode	A node errors; you see a red execution	A send looks "successful" and is wrong
Human time	Build + a few maint hours / month	Babysitting + cleanup on the fat-tail runs
Kill switch	Turn the workflow off	Need a token cap, a tool-call cap, and an approval gate

The hidden difference is replay. A workflow pays once for the run. An agent pays again for the system prompt, the tool schemas, and every prior observation each time it thinks. Anthropic's own engineering write-up on their multi-agent research system (published June 13, 2025) is blunt about the economics: in their data, agents typically use about 4× more tokens than chat, and multi-agent systems about 15× more than chats, and they say the task has to be valuable enough to pay for that (How we built our multi-agent research system). That is Anthropic's production research stack, not your inbox bot — treat the 4× / 15× as a warning, not a quote you put in a board deck as your multiplier.

What I actually put in an agent ledger that the workflow sheet does not have:

• Tokens per run (input + output + cache writes, not just "API this month")
• Tool calls per run and repeat-read ratio (same CRM record fetched three times)
• Retry count and retry tokens (history replay, not one extra completion)
• Failed-send count (wrong email, wrong Slack, wrong CRM field)
• Babysitting minutes (approve, rewrite, re-prompt)
• Cleanup hours (the Tuesday after a bad Friday send)

Skip those and your "ROI" is the subscription plus a vibe. Hiring-vs-stack cash still belongs in what AI automation actually costs in 2026. This page is the agent-specific burn.

---

Where do tokens actually go on a live agent? #

On a live agent, most of the bill is input you already paid for last turn: the system prompt, tool definitions, and the growing transcript — not the clever sentence at the end. Output is visible. Replay is where the cash goes.

As of August 25, 2026, the list prices I use when I sketch a hot path (per million tokens, standard tier, before batch or cache discounts):

Model	Job I give it	Input / MTok	Output / MTok	Source (checked Aug 25, 2026)
Claude Sonnet 5	Default agent loop, drafts, tool use	$2	$10	Anthropic model pricing
Claude Opus 4.8	Rare hard reasoning, ugly contracts	$5	$25	Same Anthropic page
GPT-5.4 mini	Labels, routing, high-volume sub-steps	$0.75	$4.50	OpenAI GPT-5.4 mini
GPT-5.5	Flagship OpenAI step, not the hot path	$5	$30	OpenAI GPT-5.5
Gemini 3.5 Flash	Volume language jobs, grounded extract	$1.50	$9	Gemini API pricing
Gemini 3.1 Pro	Harder reasoning when I want Google	$2 / $4	$12 / $18	$2/$12 at ≤200K input; $4/$18 above — Google Cloud Agent Platform pricing
Llama 4	Local / self-hosted batch if you already run the box	Infra + your time	Infra + your time	No API sticker; do not pretend it is free

Anthropic's pricing page also states that the $2 / $10 Sonnet 5 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 (same pricing page). That is good news for a workhorse loop. It is not a reason to put Opus 4.8 on every form submit.

Where the tokens actually sit on a run I log:

Bucket	What it is	Why it grows
System + tools	Instructions + every tool schema	MCP servers with fat descriptions can eat a five-figure token block before the user message. Anthropic's MCP code-execution note describes tool-definition bloat in the 150,000-token range collapsing to ~2,000 when the agent loads only what it needs (Code execution with MCP, 2025).
Transcript replay	Prior thoughts + tool results	Each new call resends the pile. Ten steps is not 10× one step.
Tool results	CRM rows, PDFs, search dumps	A 2-hour call transcript pulled twice is 50,000 extra tokens in Anthropic's own MCP example on that same page.
Output + "thinking"	The visible answer	Usually the smallest slice unless you left reasoning maxed on a flagship model.

A labeled planning example — not a client result. One inbound-lead agent, Claude Sonnet 5, 8 tool steps, 4,000 new tokens of real work, ~28,000 tokens of replayed history by the last step. Rough input: 32,000 × $2 / 1M ≈ $0.06. Output: 1,200 × $10 / 1M ≈ $0.01. Call it **$0.07** if the run is clean.

Same job on Claude Opus 4.8 is about 2.5× on input and 2.5× on output at the published $5 / $25 vs $2 / $10. Same job with a retry that replays the full 32,000 is another $0.06+ before the model writes a word. Fifty clean runs a day is still pocket change. Fifty messy runs with three retries and a 40,000-token CRM dump is how a "cheap" agent becomes a $200 week you cannot explain.

If you wire tools through MCP, read Model Context Protocol explained for the mechanism and keep the tool list short. More servers is not more value. It is more schema tax.

---

Why do retries and failed sends turn a cheap agent into a money pit? #

Retries do not add one extra completion. They resend the whole transcript, then often take another tool swing, then fail the same way. A failed send is worse: you pay the tokens and you pay a human to unsend the damage.

I watch two different failures. People mash them together and then wonder why the bill and the inbox both look cursed.

Failure	What you see	What you pay
Hard retry	Tool 500, timeout, bad JSON, empty search	Tokens for the replayed history + the new attempt
Soft loop	Tool returns { "ok": true, "results": [] } and the model tries again with a synonym	Same as a retry, except nothing is red in your logs
Failed send	Email / Slack / CRM write that went out wrong	Tokens + apology + CRM repair + trust
Ghost success	Execution marked success because HTTP was 200	You think the agent is "working" while it writes garbage

The mechanism is simple. Most agent frameworks retry by replaying conversation history, not by re-running the one broken step in a vacuum. Step 4 fails, and you pay for the system prompt, tools 1–3, and the error text again. A 20% miss rate on a 5-step agent is not a 20% surcharge. In my logs it often lands closer to 1.5×–2.5× tokens on the messy days. That range is a planning estimate from how replay works, not a lab paper you should paste into a board deck.

Soft loops are the ones that quietly empty a budget. The model gets an empty result, assumes it searched wrong, searches again, and again. Nothing threw. The circuit never opened. Avraam Mavridis describes this as a semantic loop — same tool, same argument fingerprint, same empty result — and the transport breaker never trips because the HTTP layer is fine (Agentic Patterns: Circuit breakers). I agree with the pattern even if you never read that post: cap identical tool calls, cap tokens per run, cap wall-clock.

Failed sends are where "quietly burns cash" stops being a metaphor.

Miss	Token cost	Human cost I actually see
Wrong lead email	Pennies	20–60 minutes rewriting, plus a colder lead
Duplicate invoice chase	Pennies	An annoyed customer and a collections mess
Bad CRM overwrite	Pennies	An hour reconstructing the record; sometimes a lost deal
Refund or inventory write	Pennies	Real money out the door

I do not let an agent send on the first week. Draft only. Approval queue. That is slower. It is also how you avoid paying cleanup at a founder rate for a $0.08 run.

Kill rules I put in n8n / the agent loop before go-live:

• Max N tool calls per run (I often start at 8)
• Max identical tool + args (I start at 2)
• Max tokens or USD per run (a number you can defend in a cost review)
• No send if confidence is low or the tool result is empty
• Page a human, do not "try once more" into the void

If you cannot name those five, you do not have an agent. You have an unbounded tab.

---

How much does human cleanup really cost? #

Human cleanup is usually the largest line on a losing agent, and it never appears on the model invoice. Price it at a loaded hourly rate, not at "I just fixed it between meetings."

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. Sales-and-office loaded cost in the same ECEC occupational split sits lower than management; management, professional, and related is much higher (ECEC Table 4, March 2026). Use your bookkeeper, VA, or founder rate. The BLS number is a dated public floor for "this hour was not free," not a quote for your shop.

Cleanup event	Hours I plan (hedged)	At $46.60 (BLS private-industry, Mar 2026)	At $120 founder opportunity
Rewrite a bad outbound draft before it sends	0.15–0.4	~$7–$19	~$18–$48
Apology + correct a sent email	0.5–1.5	~$23–$70	~$60–$180
Rebuild a smashed CRM record	0.75–2.0	~$35–$93	~$90–$240
Customer call after a wrong invoice chase	0.5–2.0	~$23–$93	~$60–$240
Refund / inventory undo	1.0–4.0 + cash	Labor plus the cash	Same, plus your night

Those hour ranges are planning estimates from jobs I have unwound, not a guaranteed client case study. One ugly Friday can erase a month of token "savings."

The math that founders skip:

Cleanup cost / month = (bad sends × hours each × loaded rate) + cash put back

If an agent drafts 200 emails a month and 4% need a real repair at 45 minutes, that is 6 hours. At $46.60 you are at ~$280 of cleanup — already bigger than a workhorse model bill for the same volume. At a founder rate it is worse. The API invoice might still say $38.

I count a send as "bad" if a human had to touch the record after the fact. Tone edits in the approval queue are babysitting, not cleanup. Cleanup is the work after the write left the building.

Do not put "brand lift" in the savings column to offset this. If you need a fiction line to make the agent look profitable, the agent is not profitable.

---

What babysitting hours never show up on the invoice? #

Babysitting is the standing labor of keeping the agent honest: prompt tweaks, approval clicks, "why did it do that" Slack, and the weekly log review. If you do not meter it, you will swear the agent is free while you become its unpaid ops lead.

I split human time into three buckets so it does not all vanish into "I was already at the laptop."

Bucket	What it is	Put it in the ledger?
Build	First prompts, tool wiring, test cases	Yes — amortize over the months you will keep the agent
Babysitting	Daily approve / reject, re-prompt, explain to staff	Yes — this is operating cost
Cleanup	After a bad write	Yes — this is the failure reserve you actually spend

Babysitting that never hits the invoice:

• Re-reading every draft because you do not trust the agent yet (fair in week one; a smell in month three)
• Rewriting the system prompt every time a customer uses a new phrase
• Sitting in Slack answering "did the bot send this?"
• Re-running a job by hand when the tool auth expired
• Teaching a new hire the agent's quirks instead of documenting the kill switch

A labeled week — again, not a client ROI. Approval queue of 40 lead drafts. 45 seconds to accept, 4 minutes to rewrite. If 30% need a rewrite, that is 30 × 0.75 min + 10 × 4 min ≈ 62 minutes. At $46.60 that is ~$48. At $120 it is ~$124. Add two 15-minute "the CRM field moved" fire drills and you have burned a token budget's worth of labor before the model bill shows up.

The pass test I use: median approval time under 60 seconds by week three, and rewrite rate under 20%. If you are still rewriting half the drafts, the agent is a fancy first-draft button. Keep it if the first draft still saves time. Do not call it autonomous.

If you are a solo operator, this labor is you. Running a one-person business with AI is about replacing a part-time hire. This page is the warning: an agent that needs a part-time you to babysit it did not replace anyone. It changed the costume on the same hours.

---

Which agent jobs pay back in the first 30 to 90 days? #

The jobs that pay back are narrow, high-frequency, and draft-first: the agent removes repeatable minutes without holding an irreversible pen. I want volume you can count and a send I can block.

These are planning cases, not invoices and not named-client results. Hours are conservative. Model $ assumes a workhorse (Claude Sonnet 5, GPT-5.4 mini, or Gemini 3.5 Flash) on the hot path. Loaded rate in the "why it clears" column is illustrative — swap yours in.

Job	Why it can pay	Silent-burn risk if you botch it	30–90 day tell
Inbox / lead triage + draft reply	Same shape every day; human sends	Auto-send on day one; wrong offer in writing	Rewrite rate dropping; first-response time down
Support ticket labels + suggested macros	High volume, cheap model, human still replies	Agent closes tickets or refunds	Queue time down; reopen rate not up
Invoice reminder drafts	Recurring, templated, easy to meter	Chase the wrong invoice or a paid account	Days-sales-outstanding moves; complaint count stays flat
Meeting-note → CRM fields (draft)	Painful copy-paste; read-only source	Overwrite live fields without a diff	Minutes per meeting down; field-error rate < 2%
Weekly ops digest from real systems	Replaces a Monday status scramble	Digest invents numbers from a stale export	You stop building the slide by hand
FAQ draft from your own docs	Bounded corpus; easy reject	Agent answers from the open web	Edit time per draft keeps falling

What those rows share:

• A stop condition (draft ready, label applied, digest sent to you)
• A human gate on anything a customer or a ledger will see
• A log of tokens, retries, and minutes

What I do not put in the 30-day payback column: "strategic advisor," "AI employee," "autonomous SDR that books itself," or anything that needs Claude Opus 4.8 or GPT-5.5 on every turn. Flagship is a scalpel. Volume is a workhorse.

If the work is already a clean field-to-field copy, skip the agent. A fixed n8n graph is cheaper. Agents earn their keep on messy language. Workflows earn theirs on structured clicks. That split is the whole point of AI automation vs regular automation.

---

Which agent jobs quietly lose money for months? #

The jobs that quietly lose money are unbounded, low-frequency, or allowed to write. They look impressive in a demo and then sit on a flagship model all month while you clean up the exceptions. The invoice stays small enough that nobody investigates. The labor does not.

Job	Why it looks smart	How it burns	Kill or shrink
"AI employee" with Slack access and a smile	Demo answers anything	Soft loops, tool-schema tax, founder Slack all day	Demote to one job with a cap
Unbounded web research on every lead	Feels thorough	Multi-step browse; Anthropic's 4× / 15× warning applies hard here	Cache a short dossier; do not re-research known accounts
Flagship model on labels and routing	"Best quality"	You pay Opus 4.8 / GPT-5.5 rates for work GPT-5.4 mini can do	Route the hot path down
Auto-send sales email from a thin CRM	Speed	Failed sends + cleanup at founder rate	Draft + approve; add a competitor/price gate
Nightly "optimize the business" agent	Strategy theater	Reads everything, writes a novel, no decision changes	Weekly digest, human picks one action
Agent with 20 MCP servers attached	"It can do anything"	Tool definitions eat the window before the user speaks	Three tools. Then five. Not twenty.
Low-volume specialty work (one weird contract a month)	Hard problem	Build + babysitting never amortize	Do it yourself or hire the hour
Retry-with-no-cap into a flaky API	"It will get it"	Replay tokens + duplicate writes	Circuit breaker, then page a human

The pattern I keep seeing: the owner bought autonomy and received a second inbox. The agent produces work-shaped text. Someone still has to decide. That someone is you. If the decision time did not fall, you did not buy payback. You bought a content generator with API fees.

A quiet-burn smell test I run at 30 days:

• Token $ is "fine" but you cannot name tokens per successful job
• Retry rate is unknown
• Rewrite rate is still > 30%
• Someone on the team has a side ritual of "checking the bot"
• One bad send already happened and you called it a learning moment instead of a ledger line

If three of those are true, the agent is burning cash even if the card statement looks polite.

---

How do I measure agent payback without a fake ROI slide? #

You measure agent payback as successful jobs minus full run cost, using logs you can audit on a Tuesday — not a vendor multiplier and not a story about "10x." If a number has no source, it does not go in the ledger.

I keep one row per agent per week. That is the whole dashboard.

Field	How I fill it	Cheat I reject
Successful jobs	Count of runs that hit the stop condition and needed no cleanup	Counting every execution as a win
Hours returned	Time study on 10 jobs, then × volume. Conservative.	"It feels like I got my mornings back"
Token $	Vendor usage export, split by agent / API key	One blended OpenAI bill for the whole company
Retry $	Tokens on runs with retry_count > 0	Ignoring them because "that's just how agents work"
Babysitting hours	Timer or a tally in the approval tool	Founder time at $0
Cleanup hours + cash	Ticket after each bad send	"We'll train it out" with no dollar
Net	(hours × your loaded rate) − token − retry − babysitting − cleanup	Adding "innovation value"

Hours × rate is the same idea as the parent pillar. I am not pasting that worksheet. The extra columns are the point of this post.

A labeled arithmetic example you can replace — not a client result:

• 180 successful lead drafts in a month
• Time study: 4.0 minutes saved per draft after approval (you still spend 0.8 minutes approving)
• Hours returned: 180 × 4.0 / 60 = 12.0
• Loaded rate: $46.60 (BLS private-industry March 2026 — swap this)
• Labor value: ≈ $559
• Token $: $42 (workhorse, logged)
• Babysitting beyond the 0.8 min already netted out: 1.5 hours → ≈ $70
• Cleanup: one bad send, 1.0 hour + $0 cash → ≈ $47
• Net ≈ $559 − $42 − $70 − $47 = $400

That agent is paying. Same shape with 6 minutes of rewrite per draft and three cleanup events is a different story. Do the second sheet. Do not ship the first sheet as if it were destiny.

Rules for the time study: watch the real queue, include the approval click, exclude wait time while you were in a meeting. If you cannot watch 10 jobs, you are not ready to claim hours.

I want a weekly digest more than a slide. Execution count, fail rate, token $, rewrite rate, cleanup events. If you cannot get that out of n8n or your agent host, you cannot defend the spend.

---

When should I kill an agent instead of tuning it? #

Kill the agent when the job is wrong, the write surface is too hot, or two full prompt/tool passes did not move rewrite rate and retries. Tuning a bad job is how silent burn becomes a personality trait.

I give a new agent two weeks of draft-only and one more week of guarded sends if the draft week was clean. After that I use a kill table, not hope.

Signal (look at 14 days)	Tune once	Kill or demote
Rewrite rate	20–35%, falling	Still > 40% after two prompt passes
Retry / soft-loop rate	< 10% of runs, and you added a cap	> 20% or uncapped identical tool calls
Cleanup events	Zero, or one with a patch	Two or more customer-facing misses
Babysitting	Falling toward < 1 hour / week	Still a daily ritual
Token $ / successful job	Stable or down	Climbing while volume is flat
Owner feeling	Mild annoyance	You dread the approval queue

Demote means: strip tools, force a cheaper model, turn sends back to drafts, or replace the loop with a fixed workflow. Kill means: off. Export the logs. Keep the prompt in git if you want the scar tissue.

What I do not treat as a reason to keep it:

• "We already spent the build week"
• "The vendor said it learns"
• "Next month's model will fix it"
• "It's only $60 on the card" (while you spend four hours a week on it)

Sunk build hours are sunk. The parent pillar will tell you to amortize build over the life of a working workflow. An agent that fails the table above has no remaining life. Stop amortizing fiction.

If the job is still good and the agent is messy, I often win by deleting tools, not by adding a second agent. Two agents that retry each other's empty results is how you buy the 15× token world Anthropic warned about without getting the research benefit.

---

How should I route models so the hot path stays cheap? #

Put a workhorse on the loop and a flagship on the exception. The hot path is labels, drafts, and tool choice. The exception is the one ugly document or the one send you cannot cheaply undo. If every step is Claude Opus 4.8 or GPT-5.5, you are paying research-stack prices for inbox work.

Routing I use in August 2026. Re-check list prices the week you ship.

Step	Default model	When I escalate	When I refuse
Classify / route / extract a field	GPT-5.4 mini or Gemini 3.5 Flash	Messy handwriting, mixed languages, legal-ish text	Flagship on every inbound email
Draft a reply in my voice	Claude Sonnet 5	A high-value account or a complaint	Opus 4.8 for "thanks, we got your form"
Choose the next tool	Claude Sonnet 5	Tool list is large and I already trimmed schemas	Adding more MCP servers instead of routing
Hard reasoning (contract clause, nasty edge)	Claude Opus 4.8, GPT-5.5, or Gemini 3.1 Pro	Once per run, after a cheap model flagged "unsure"	The whole loop
Local batch / private text	Llama 4 on a box you already operate	You have a person who can keep it up	"Free Llama" with no operator

Price gaps that make this worth doing (standard list, Aug 25, 2026):

• Claude Sonnet 5 vs Claude Opus 4.8: $2 / $10 vs $5 / $25 per MTok (Anthropic)
• GPT-5.4 mini vs GPT-5.5: $0.75 / $4.50 vs $5 / $30 (OpenAI mini, GPT-5.5)
• Gemini 3.5 Flash vs Gemini 3.1 Pro (≤200K): $1.50 / $9 vs $2 / $12 (Gemini API, Google Cloud)

A 5×–6× output gap on the OpenAI pair is not a rounding error when the agent retries. One flagship replay of a fat transcript can cost more than a hundred mini classifications.

Prompt caching helps when the system prompt and tool schemas are stable. Anthropic lists cache hits at $0.20 / MTok for Sonnet 5 and $0.50 / MTok for Opus 4.8 (pricing page). Cache writes cost extra. If you rewrite the system prompt every afternoon, you never stay warm.

Llama 4 is a cost cap, not a personality. If you do not already run inference, the "savings" show up as your weekend. I only put it in the model table for shops that already have the box.

My opinion: Sonnet 5 or GPT-5.4 mini should be the default agent brain in 2026. Escalate by policy, not by vibe. If you cannot write the policy in five lines, you will escalate everything.

---

FAQ: Agent payback vs silent burn #

Is an AI agent cheaper than a part-time hire? #

Sometimes on the repeatable slice — never as a blanket swap for a person who handles exceptions. BLS private-industry loaded cost was $46.60 per hour in March 2026 (ECEC). A workhorse agent loop is usually tens of dollars of tokens plus your babysitting. If the hire's job is judgment and angry customers, keep the hire. If the hire's job is labeling tickets and drafting the same email, model the hours, then compare. I will not invent a payback month for your shop.

How many tokens does a typical business agent burn per run? #

There is no honest single number. A clean 6–10 step inbox agent might sit in the low tens of thousands of tokens; a research-style loop can run far past that. Anthropic reported that agents in their data used about 4× chat tokens, and multi-agent systems about 15× (June 13, 2025). Log your p50 and p95 tokens per successful job for a week. Budget from the fat tail, not the pretty median.

Should I put Claude Opus 4.8 on every step? #

No. Opus 4.8 is $5 / $25 per million tokens; Sonnet 5 is $2 / $10 as of August 25, 2026 (Anthropic pricing). I keep Opus 4.8, GPT-5.5, and Gemini 3.1 Pro for the rare hard step. Labels, routing, and most drafts stay on Sonnet 5, GPT-5.4 mini, or Gemini 3.5 Flash. Flagship-on-everything is how a "cheap" agent stops being cheap.

What is a retry loop and why does it cost money? #

A retry loop is the agent calling the same tool, or the same goal, again after a miss — and paying to resend the whole transcript each time. Soft loops are worse: the tool returns empty success and the model just tries a new query. Cap identical calls, cap tokens per run, and page a human. A 20% miss rate is not a 20% surcharge; replay makes it look more like a 1.5×–2.5× token day in my logs. That last bit is a planning estimate, not a universal constant.

How do I know if my agent is failing silently? #

Look for success flags with empty results, climbing tokens at flat volume, and humans who "just check the bot." HTTP 200 is not a business win. I want rewrite rate, retry count, cleanup events, and tokens per successful job in one weekly digest. If you only have a green execution list, you are flying on dashboard theater.

Can I run agents on Llama 4 to cut cost? #

Yes, if you already operate the box and the job is batch / private text you can evaluate. Llama 4 does not have a sticker like Sonnet 5. You pay GPU, electricity, and the person who keeps the process alive. For a 3-person shop with no inference habit, a workhorse API is usually cheaper than a new operations surface. Do not call local models free.

What is a reasonable monthly model budget for one agent? #

I plan allowances, then replace them with a week of logs — often $25 / $60 / $120 per agent as a first sketch for small-volume workhorse loops. High-volume support or anything that browses the web will blow that. Multi-agent research is a different sport; Anthropic's 15× note is the reason I do not put "research everything" on a $60 cap. If you cannot name the budget before go-live, you will discover it on the card.

Do I count founder babysitting time as a cost? #

Yes. Founder hours are usually the most expensive line you have. Price them at an opportunity rate you write down, or at least at the BLS loaded floor so the hour is not $0. Approval clicks in week one are normal. A daily rewrite ritual in month three is operating cost, and it belongs next to the token $ or the agent is lying to you.

When is a fixed n8n workflow cheaper than an agent? #

Whenever the input is already a field and the path does not branch on language. Copy this status to that column, send this template when a date hits, ping Slack on a failed payment — that is a workflow. I reach for an agent when the input is messy text and the output is a draft. Mixing them is fine. Pretending every workflow needs a model loop is how you buy retries you did not need.

How long should I give an agent before I shut it off? #

Two weeks draft-only, then a kill review. If rewrite rate and retries have not moved after two prompt/tool passes, demote or kill. One more week of guarded sends is earned, not owed. Sunk build time is not a reason to keep a babysitting hobby. Export the logs, keep the prompt, turn it off.

---

An agent that pays is boring on purpose: one job, a workhorse model, a cap, a human gate, and a Tuesday digest. An agent that burns cash is usually a demo that never got a ledger.

If you want that ledger built against your actual queue — n8n, MCP tools, model routing, kill switches — book an AI automation strategy call. I will tell you whether you need an agent at all, or a fixed workflow that does not retry itself into a hole.

Book an AI automation strategy call when you are ready to meter the loop instead of hoping the invoice stays small.