What hiring an AI agent actually involves for a small business
Vendor documentation puts real numbers on token costs and failure rates — and shows that most of the work is approval design, not model selection.
The phrase "hiring an AI agent" suggests something like recruitment — you pick a candidate, you pay a salary, the work gets done. The documentation published by the two largest vendors describes something considerably less tidy: a software component that runs commands and edits records on your behalf, bills by the token, and fails in ways that a small business owner has to design around in advance.
An agent, in the vendor sense, is a model with permission to act. Anthropic's Agent SDK documentation describes agents that "autonomously read files, run commands, search the web, edit code, and more," shipping with built-in tools including Read, Write, Edit, Bash, Glob, Grep, WebSearch and WebFetch. OpenAI's agent guides frame the same capability through what constrains it, distinguishing guardrails, which "validate input, output, or tool behavior automatically," from human review, which "pauses the run so a person or policy can approve or reject a sensitive action."
What neither vendor claims is that this works unattended out of the box.
What it costs, in the vendors' own numbers
Anthropic publishes per-token pricing rather than per-seat pricing, which makes small-business budgeting an estimation exercise rather than a line item. As listed on its pricing page, Claude Haiku 4.5 is billed at $1 per million input tokens and $5 per million output tokens; Sonnet 4.6 at $3 and $15; Opus 4.8 at $5 and $25. The company's own worked example for a customer support workload assumes roughly 3,700 tokens per conversation and puts the total at approximately $37.00 per 10,000 tickets using Haiku 4.5.
That figure is the floor, not the bill. Server-side tools carry separate charges: web search is priced at "$10 per 1,000 searches, plus standard token costs for search-generated content." Managed Agents adds session runtime at $0.08 per session-hour. Tool definitions themselves consume tokens before the agent does any work — the bash tool alone adds 325 input tokens per request on Opus 4.7 and 4.8, on top of a per-model tool-use system prompt.
There are discounts in the other direction. The Batch API carries "a 50% discount on both input and output tokens" for asynchronous work, and a prompt cache hit "costs 10% of the standard input price." A business whose agent re-reads the same policy document on every request is paying full price for something the vendor will cache at a tenth of the cost.
Where it fails
Anthropic's engineering write-up on advanced tool use is unusually direct about failure modes. "The most common failures are wrong tool selection and incorrect parameters, especially when tools have similar names," it states. The accuracy numbers it reports are worth reading closely, because they describe how much headroom the default configuration leaves: Opus 4.5 "improved from 79.5% to 88.1% with Tool Search Tool enabled," and Opus 4 "improved from 49% to 74%." Tool use examples, separately, "improved accuracy from 72% to 90% on complex parameter handling."
Read plainly, a tuned agent picking tools correctly around 88% of the time is the improved case. For a business deciding whether an agent should issue refunds or send client emails unsupervised, that is the number that matters more than any pricing table.
Context is the other constraint. Anthropic notes that at its own scale it has "seen tool definitions consume 134K tokens before optimization" — context spent on describing tools rather than doing work. And OpenAI warns that safety checks do not apply as broadly as their names imply: "Agent-level guardrails don't run everywhere." Input guardrails run "only for the first agent in the chain," output guardrails "only for the agent that produces the final output," and tool guardrails only "on the function tools they're attached to." The guidance is explicit that in a multi-step workflow, validation belongs next to the tool that creates the side effect.
The work is approval design
Both vendors converge on the same practical point from different directions. Anthropic's SDK exposes permission controls, allowlists, hooks that can "validate, log, block, or transform agent behavior," and a disallowed-tools setting. OpenAI's human-review pattern pauses a run before "side effects like cancellations, edits, shell commands, or sensitive MCP actions" and returns resumable state so a person can approve or reject.
That is the actual job. Choosing a model takes an afternoon. Deciding which actions an agent may take without a human, and building the pause where a human intervenes, is the part that determines whether the deployment is useful or expensive.
A note on the guidance landscape
Small businesses researching this will find a growing tier of sites that aggregate agent guidance between vendor documentation and consultancy. One such site describes itself as "The definitive resource for hiring AI agents" and organises material into twelve category guides — customer support, sales, coding, marketing, recruitment, data analysis, legal, personal assistants, pricing, build-versus-buy, security and compliance, and ROI measurement. Its coverage is uneven by its own admission: the build-versus-buy category page states that "This category guide is currently being developed," pointing readers to individual articles instead. The site attributes its content to "top YouTube experts, enhanced and verified for accuracy," and names no operator or individual author, which is a reason to treat its framework comparisons as a starting map rather than a settled verdict.
The general rule holds for any intermediary: cost and capability claims should be checked against the vendor page they derive from, because those pages change. Anthropic's own pricing table carries a note that Sonnet 5's introductory rate of $2 and $10 per million tokens runs "through August 31, 2026," after which it rises to $3 and $15.
Details verified against Anthropic's Claude platform documentation, Anthropic's engineering blog and OpenAI's developer documentation in July 2026. Prices and model availability change frequently; check the vendor pricing pages before budgeting.