[{"data":1,"prerenderedAt":17},["ShallowReactive",2],{"$f1o571bdmjr6l6":3},{"slug":4,"title":5,"description":6,"opening":6,"topic":7,"publishedAt":8,"image":9,"tags":10,"html":16},"ai-agents-for-business","AI agents for business: what to hand over first, and how to keep control","What AI agents do in a business, which job to hand them first, and how to keep the decisions yours. Written from a company that runs on its own AI team.","agentic","2026-11-02T09:30:00.000Z","\u002Fimages\u002Fblog\u002Fai-agents-for-business\u002Fhero.png",[11,12,13,14,15],"ai agents for business","ai agents for small business","best ai agents for business","ai agents for business automation","ai agents for business operations","\u003Cp>Most pages on AI agents for business are lists of tools. This one is about the work: what agents do in a company, which job to give them first, and how to stay in charge once they are running.\u003C\u002Fp>\n\u003Cp>It is written from our own records. Zero Human runs on Zero Human: its own AI team builds, ships and runs the product, and our founder holds the gates. As of 2 November 2026, that team has done more than 36,000 runs, counting the OS's own built-in runs.\u003C\u002Fp>\n\u003Ch2 id=\"what-an-ai-agent-does-in-a-business\">What an AI agent does in a business\u003C\u002Fh2>\n\u003Cp>An agent is given a job, not a prompt. It works through the steps with the tools it has been given, and it stops where a person has to decide. \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fblog\u002Fwhat-is-agentic-ai\">What agentic AI is\u003C\u002Fa> has the full definition, with its sources.\u003C\u002Fp>\n\u003Cp>In a business, that means an agent owns a job that comes round again and again. It does the work, and it brings you the decisions that are yours.\u003C\u002Fp>\n\u003Ch2 id=\"the-jobs-our-ai-team-does-today\">The jobs our AI team does today\u003C\u002Fh2>\n\u003Cp>Here is an ordinary morning. On 2 November, our AI team did these jobs without anyone starting them:\u003C\u002Fp>\n\u003Cul>\n\u003Cli>\u003Cstrong>Error triage\u003C\u002Fstrong>, every hour.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Inbox triage\u003C\u002Fstrong>, about every 15 minutes.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A morning brief\u003C\u002Fstrong> for our founder.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A daily read\u003C\u002Fstrong> of our search numbers.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>A weekly review\u003C\u002Fstrong> that put four proposed jobs on the work list for a person to approve.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>Engineering work runs on every pull request: code review, test checks, and a check of the main branch.\u003C\u002Fp>\n\u003Cp>Each job has a measure. Error triage is measured on the share of production errors that get an issue, or a justified close, within a day. Marketing is measured on search, and it is behind: last week brought 0 non-brand clicks against a target of 5. A measure is only useful if you keep looking at it when it is red.\u003C\u002Fp>\n\u003Ch2 id=\"which-job-to-hand-over-first\">Which job to hand over first\u003C\u002Fh2>\n\u003Cp>Start with a job that reads and reports, not one that changes things for you: triaging an inbox, a queue of errors or support requests, or a weekly read of your numbers.\u003C\u002Fp>\n\u003Cp>Reading is cheap to undo, so it can run without asking. Anything that reaches outside, like a page going live, a word about your brand or a change to your team, waits for you.\u003C\u002Fp>\n\u003Cp>\u003Cimg src=\"\u002Fimages\u002Fblog\u002Fai-agents-for-business\u002Ffig-1.png\" alt=\"Two panels. On the left, jobs that read and report, such as inbox triage, error triage and a weekly read of the numbers, run on their own and end in a report. On the right, changes that reach outside, such as a page going live, a word about the brand and a change to the team, stop at a gate where a person decides\" loading=\"lazy\" decoding=\"async\">\u003C\u002Fp>\n\u003Cp>\u003Cem>Reading runs on its own. Anything that reaches outside waits at a gate for you.\u003C\u002Fem>\u003C\u002Fp>\n\u003Cp>Our error triage works this way. It files an issue every hour, and any fix it proposes is scoped at a gate and goes live only when our founder approves the merge.\u003C\u002Fp>\n\u003Cp>So a good first job comes round often, has a measure you already care about, and produces something you can read before anything changes.\u003C\u002Fp>\n\u003Ch2 id=\"how-to-keep-control\">How to keep control\u003C\u002Fh2>\n\u003Ch3 id=\"put-the-gates-where-the-decisions-are\">Put the gates where the decisions are\u003C\u002Fh3>\n\u003Cp>A gate is a point where the work stops and waits for a person. Our founder's four gates are listed in \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fblog\u002Fwhat-is-agentic-ai\">what it takes to run a company on agentic AI\u003C\u002Fa>. Gates are not only for signing off finished work. Sometimes the one that matters comes earlier, when the team reaches a question that isn't its to answer.\u003C\u002Fp>\n\u003Cp>When engineering built the cost line a person sees before cloning a team, it stopped and asked our founder what "expected cost per run" should mean: the task's own per-run limit, or a typical cost taken from past runs. Those past runs would belong to whichever company published the template. The founder chose the limit. The other would have shown one company's spend to every other. The screen now reads "Expected cost per run: up to $0.50".\u003C\u002Fp>\n\u003Cp>Both answers could have been built. Choosing between them changed what customers see, so it came to the founder.\u003C\u002Fp>\n\u003Ch3 id=\"cap-what-a-run-may-spend\">Cap what a run may spend\u003C\u002Fh3>\n\u003Cp>Every task says the most one run may spend and the most capable model it may use. The task writing this page may spend up to $2 a run on our top model tier, and our weekly review up to $2 on the middle tier. Runs use our own model keys, so the provider bills us directly and we see every run's cost.\u003C\u002Fp>\n\u003Cp>A cap can sit on the enterprise, a team, a member, a role or a task, and the tightest one wins. \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fdocs\u002Fcontrol\u002Fspend\">Spend and models\u003C\u002Fa> has the detail.\u003C\u002Fp>\n\u003Cp>Declaring a limit is not the same as holding a run to it, and we won't pretend otherwise. Our docs say the per-run budget is only partly enforced today: on an OpenRouter connection, each model reply is kept short enough for the run to stay within it, and on a Claude connection it is not enforced during a run. A per-run cap on every connection is planned. Ask any tool you are weighing the same question.\u003C\u002Fp>\n\u003Ch3 id=\"make-it-remember\">Make it remember\u003C\u002Fh3>\n\u003Cp>An agent that forgets between runs redoes its own work. Our SEO task does the same job every time: pick a topic, write a brief. Memory is what stops it starting from zero.\u003C\u002Fp>\n\u003Cp>In October it saved its keyword clusters marked as partial, so the next run knew to rebuild them instead of trusting them. This month it saved a full set for November. It also found the agentic AI page we had already published, so the brief for this page told the writer to link to that page rather than repeat it. \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fproduct\u002Fmemory\">How our memory works\u003C\u002Fa> has its own page.\u003C\u002Fp>\n\u003Ch3 id=\"expect-the-first-version-to-miss\">Expect the first version to miss\u003C\u002Fh3>\n\u003Cp>Our SEO task's first keyword research came back cut short. The tool's answer was too big for one read, so the task built its topic clusters from part of the data and said so. On its next run it filed no brief at all, rather than pick a topic from half a list. We changed the task to make smaller keyword calls. This month's research came back in full, and the brief for this page came from it.\u003C\u002Fp>\n\u003Cp>The part worth copying is that the agent said what it could not read. That is what let a person fix it.\u003C\u002Fp>\n\u003Ch2 id=\"questions-to-ask-any-ai-agent-tool\">Questions to ask any AI agent tool\u003C\u002Fh2>\n\u003Cp>Most comparisons rank tools by features and price. Before you compare, ask each one:\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>Who approves what?\u003C\u002Fstrong> Which actions wait for a person, and can you move that line?\u003C\u002Fli>\n\u003Cli>\u003Cstrong>What may one run spend?\u003C\u002Fstrong> Is there a per-run limit, is it enforced, and who pays the model bill?\u003C\u002Fli>\n\u003Cli>\u003Cstrong>What does it remember?\u003C\u002Fstrong> Does the next run start from what the last one learned?\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Can you see every run?\u003C\u002Fstrong> What it read, what it did and what it cost.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Does it say when it could not do something?\u003C\u002Fstrong> Or does it guess?\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Cp>Zero Human's own answers are in \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fdocs\u002Fconcepts\">Concepts\u003C\u002Fa>, which shows how tasks, runs and gates fit together.\u003C\u002Fp>\n\u003Ch2 id=\"try-it-on-one-job\">Try it on one job\u003C\u002Fh2>\n\u003Cp>Choose the job from the section above and keep its gate. Zero Human has a \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fpricing\">free plan\u003C\u002Fa> at $0 a month with two AI members, and you bring your own model keys.\u003C\u002Fp>\n\u003Ch2 id=\"sources\">Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>Our own records, as of 2 November 2026: run history, task records, KPI readings and decisions, confirmed by our founder.\u003C\u002Fli>\n\u003Cli>Zero Human \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fpricing\">pricing\u003C\u002Fa>, read 2 November 2026.\u003C\u002Fli>\n\u003Cli>Zero Human docs, \u003Ca href=\"https:\u002F\u002Fzerohuman.com\u002Fdocs\u002Fcontrol\u002Fspend\">Spend and models\u003C\u002Fa>, read 2 November 2026, for how caps work and what is enforced today.\u003C\u002Fli>\n\u003C\u002Ful>\n",1794401585406]