- 11 min read
AI Agent Bill Shock: What an Agent Really Costs to Run, and Who Pays When It Makes a Deal You Never Approved

Contents
- 11 min read
Two warnings landed in the same week. On 13 October a consultancy told IT Brief Australia that AI agents are about to give small businesses “bill shock”, and four days earlier a Clayton Utz partner told SmartCompany that when an agent makes a commitment, the business wears it. Put those together and you get the real question: what does an AI agent cost to run, and what does it cost when it gets something wrong? Both are knowable before you switch one on. Digital One Agency designs and builds AI agents for business automation with the spending caps and authority limits written in from day one, and the fastest way to see what yours would cost is to book a meeting.
Why agents bill differently from the chatbot your staff already use
A chatbot seat is a fixed monthly fee. You know the number. Chris Petzoldt of Pretian Squared put it plainly in the IT Brief piece: “Seats are simple. You know what you’ll pay.” Agents are paid for by consumption: tokens, credits or per-task charges. In his words, they “run tasks, call tools and burn compute whether or not the output was worth it.”
The mechanism is worth understanding, because it explains every surprise invoice. An agent does not answer once. It reads its instructions, looks at the job, calls a tool (your CRM, your inbox, a supplier’s website), reads the result, decides what to do next and goes round again. On each lap the model re-reads everything so far: the instructions, the tool descriptions, the conversation, whatever documents it pulled in. A job that takes eight laps can consume eight times the reading, and the reading gets longer every lap.
That is why the same “answer a customer email” job can cost a fraction of a cent or several dollars depending on how the agent is built.
What one agent task actually costs: a worked example
Here is a calculation from published prices, so you can see the shape of it. Anthropic’s API pricing page lists, in US dollars per million tokens, US$10 in and US$50 out for its top model, US$2 in and US$10 out for its mid-tier Sonnet 5.5, and US$0.10 in and US$0.50 out for the small Haiku 5.5.
Take a modest customer-service agent. It reads an enquiry, checks the order system, looks up a policy, drafts a reply and logs the outcome. Say it makes eight model calls, each re-reading around 20,000 tokens of context and writing about 500. That is 160,000 tokens in and 4,000 out per enquiry.
- On the top model: US$1.60 in plus US$0.20 out, so US$1.80 per enquiry.
- On the mid-tier model: US$0.32 plus US$0.04, so US$0.36 per enquiry.
- On the small model: US$0.016 plus US$0.002, so under 2 US cents per enquiry.
At 2,000 enquiries a month, that is roughly US$3,600, US$720 or US$36. Same job, same volume, a hundredfold spread, and the model choice is the biggest line in your running cost.
Now compare that with what businesses actually spend. IT Brief cites Ramp data from more than 70,000 US businesses showing median AI spend of about US$12 per employee per month. A ten-person firm at the median spends around US$120 a month. One mid-tier agent handling 2,000 enquiries costs six times that. Agents can absolutely pay for themselves, but they are a different budget category from seats, and nobody should find that out from the invoice.
Bill shock is usually a loop, not a price
The scary bills rarely come from normal operation. They come from an agent stuck in a loop: retrying a tool that keeps failing, re-reading a giant attachment on every lap, or chasing a goal it cannot reach.
Run the numbers on one bad job. An agent that loops 60 times with its context swollen to 100,000 tokens reads six million tokens. That is about US$12 on the mid-tier model and US$60 on the top model, for a single enquiry that probably still failed. If a broken integration triggers that on 200 jobs overnight, the top-model bill is US$12,000 by breakfast.
Every one of those failures can be prevented in the build:
- A hard step limit. No task gets more than, say, 15 laps. After that it stops and hands to a person.
- A per-task and per-day spend cap. Set at the provider account and enforced again in your own code, so one failure cannot override the other.
- Context hygiene. The agent keeps a summary instead of re-reading a 40-page PDF on every lap.
- Prompt caching and batching. The fixed part of the prompt can be cached; the pricing page lists cache hits on the mid-tier model at 0.05 times the normal input price. Work that is not urgent, such as overnight document processing, can go through the batch service at a 50 per cent discount.
- The smallest model that passes the test. Routine classification and routing rarely need the top model. Escalate to a bigger one only for the hard cases.
Petzoldt’s advice from the IT Brief piece is the same at a higher level: track total monthly AI spend, read the pricing metric before signing up, and “set usage limits before rolling out automated agents more widely.”
Want this built or fixed for your business?
Digital One Agency builds AI systems, apps, platforms and AI-search visibility for Australian businesses. Start with a free AI audit or book a meeting and we will tell you what it would take.
The second bill: when the agent commits you to something
The cost that does not appear on the provider invoice is liability. The SmartCompany piece is blunt. Simon Newcomb, head of AI at Clayton Utz, said “The AI agent is the business,” and that there is no legal difference between the two. University of Melbourne law professor Jeannie Paterson went further: “There’s no doubt at all that the responsibility lies with the person that’s deployed the agent.”
The ACCC told SmartCompany that a business can be liable for misleading information it provides “no matter how it provides that information,” including through AI chatbots. The cases cited are not hypothetical. Air Canada was ordered to compensate a customer after its chatbot wrongly told him he could claim a bereavement fare after the fact, and the tribunal rejected the argument that the chatbot was responsible for its own words. In Singapore, the Quoine case involved 13 automated trades executed at about 250 times the market price after a 2017 software fault, and the exchange was ordered to pay US$4 million into court as an interim payment.
So the practical cost of a badly scoped agent is whatever it promised. A refund it should not have offered. A discount outside policy. A delivery date your warehouse cannot meet. A statement about a product that turns out to be misleading under Australian Consumer Law.
Scope the agent’s authority like you would a new hire
Newcomb’s framing is the most useful one we have seen: treat an agent “almost like a parallel workforce” that needs training and delegated authority. That translates into build decisions:
- Write down what it may say and do. Answer delivery questions: yes. Offer refunds: only up to a set amount, only on orders that meet the policy, otherwise escalate.
- Give it the narrowest access that does the job. Read-only on the order system unless writing is the point. No access to pricing it is not allowed to change.
- Put money and commitments behind a human gate. Anything that commits the business to spend, discount or deliver above a threshold waits for a person to approve.
- Log every action with the reason. When a customer disputes what the agent said, you need the transcript and the tool calls, not a guess.
- Re-test when the model changes. Vendors update models regularly. Newcomb recommends testing continuously as models change, because behaviour that passed in March may not pass in November.
- Update your terms. Newcomb also flagged the reverse risk: customers using their own AI agents to buy from you, then trying to unwind purchases the agent got wrong. Your terms should say who is responsible for orders placed by a customer’s agent.
If a supplier builds the agent, read the contract with the same care. Newcomb’s point was that contracting out an agentic system is, in a sense, outsourcing your governance. Ask who carries the cost if the agent misbehaves, what transparency you get into its decisions, and whether you can see and change its limits yourself. We covered the technical side of keeping agents inside their lane in what to control before you put an AI agent in your business, and the security controls in our secure AI system development work.
The budget conversation to have before you build
Before any agent goes live, you should be able to answer four questions in a sentence each:
- What does one task cost at normal volume, and at ten times normal volume? If the builder cannot give a per-task figure, they have not designed the agent yet.
- What is the worst a single failed task can cost? That is your step limit times your context size times the model price. It should be a small, known number.
- What is the agent allowed to commit the business to without a person? In dollars, not adjectives.
- What does it replace or earn? Petzoldt’s point was that larger firms tie AI spend to commercial returns. An agent that saves two hours a day of enquiry handling, or answers after-hours leads that used to go to voicemail, has a value you can set against the running cost.
How Digital One Agency builds agents that do not surprise you
We build custom agents for Australian businesses: email and enquiry handling, AI voice agents for the phone line, SMS follow-up, and back-office automation that connects to the systems you already run. When we scope an agent, we work out a per-task cost at your real volumes, match the model to each step rather than running one expensive model for everything, put step limits and spend caps in code, write an authority policy for what the agent may promise, and log every action so you can see each decision it made.
On price: running costs depend on volume and model choice, which is why we work them out from your numbers rather than quoting a flat figure. Build cost depends on the systems the agent has to connect to, and we scope that with you before anything is committed. If you are weighing a custom agent against an off-the-shelf tool, our AI development solutions page sets out when each makes sense.
The next step is simple. Book a meeting with Digital One Agency and bring one process you want an agent to handle, plus a rough monthly volume. We will tell you what it would cost to build, what it would cost to run, what limits we would set, and whether it is worth doing at all. If you are not sure where to start, the free AI audit looks across your business and identifies the processes where an agent would pay back fastest.
Frequently asked questions
How much does an AI agent cost to run per month?
It depends mostly on the model and how many steps each task takes. Using published per-token prices, a customer-service task of eight model calls can cost under 2 US cents on a small model, about 36 US cents on a mid-tier model and about US$1.80 on a top model. At 2,000 tasks a month that is roughly US$36, US$720 or US$3,600 before any caching or batching discounts.
What causes AI agent bill shock?
Usually a loop rather than the price itself. An agent that keeps retrying a failing tool, or re-reads a large document on every step, can consume millions of tokens on one job. Hard step limits, per-task and daily spend caps, and summarising context instead of re-reading it stop that from happening.
Is my business liable if our AI agent promises a customer something?
Generally yes. A Clayton Utz partner told SmartCompany there is no legal difference between the agent and the business, and the ACCC said a business can be liable for misleading information no matter how it is provided, including through AI chatbots. Limit what the agent may offer and put commitments above a threshold behind a human approval. Get legal advice for your specific situation.
Should we use the most powerful AI model for our agent?
Not for every step. Routine classification, routing and lookups rarely need the top model, and the price gap between model tiers can be a hundredfold. A well-built agent uses a small model for routine steps and escalates to a larger one only for the hard cases.
What should be in a contract with a company that builds our AI agent?
Clear terms on who carries the cost if the agent misbehaves, what visibility you get into its decisions and logs, how its limits are set and changed, and how it will be re-tested when the underlying model changes. Outsourcing an agentic system is partly outsourcing governance, so the contract has to hand some of that control back to you.
Need AI Voice Agents in Australia or the UK?
Digital One Agency is a senior, founder-led engineering team based in Australia (Gold Coast HQ) with active UK clients in London, Manchester and Edinburgh. We build production AI voice agents that answer phones, qualify leads, and book appointments — built bespoke, not a SaaS reseller.
- Free 30-minute discovery call with a senior engineer
- Fixed-scope quote with milestone billing — no hourly rates
- AU & UK references on request
- Bespoke code you own — not a SaaS reseller, not a ChatGPT wrapper
Book your free discovery call → Or send us your project details
Share
- Get in touch
Kickstart a project today!
Fill out the form, and a member of our team will reach out to discuss your project and answer any questions you might have.