--- title: Where to Point Your First Agent · Atrium description: Find the bottleneck costing you the most, then build the agent that fixes it. A working guide to asking the right question before you build anything. --- Atrium a working guide --- # Where to point your first agent Most of the value in working with AI comes from asking the right question, and being precise about what you are trying to achieve. This guide is about finding that question for your business, so that when you sit down with Claude you already know what you are building and why it is worth building. --- § I  The Question ## Start with the bottleneck, not the tool You can spin up an agent for almost anything, which is exactly the problem. Most founders spend their first month automating something that was never costing them very much, conclude the technology is overrated, and stop. So before any of the setup, answer one question honestly: **which bottleneck in my business is costing me the most money right now?** Not the most annoying one. The most expensive one. They are rarely the same, and the annoying one is the one you will be tempted to fix first. Once you can name it and put a number on it, the conversation with Claude becomes straightforward, because you are no longer asking it to be clever. You are asking it to solve something specific that you have already measured. --- § II  The Diagnosis ## Where founders quietly lose the most These are the ones we see most often. Read them looking for yourself, and ignore the ones that are not your problem, however tempting they look. ### The follow-up that never went out You meant to circle back. Nineteen days passed. They signed with someone who simply showed up on day four. Nothing went wrong in the deal; it died of nothing happening. *Price it:* count the open conversations that have gone more than ten days without a touch. Multiply by your usual close rate, then by your average deal. That is the annual cost of one habit. ### The slow reply to inbound You paid for that lead through ads, content, or an event. Then the enquiry sat for a day because you were in back-to-backs. Buyers shortlist whoever answers first, so you have effectively paid full price for a fraction of a lead. *Price it:* your monthly acquisition spend, multiplied by the share of enquiries you answer late. ### The business that lives in your head The real state of your pipeline is spread across your inbox, two chat threads, and your memory. You cannot forecast it honestly, and you cannot hand any of it to a first hire without months of shadowing. *Price it:* the hire you have postponed, plus whatever a wrong forecast has already cost you once. ### The proposal that takes six days Quotes compete with everything else you do, so they go out late, and deals cool while they wait. *Price it:* average days to send, multiplied by how many deals a month, against the ones that went quiet in that window. ### The invoices nobody chases Chasing feels awkward, so it does not happen. That is your runway, resting in other people's bank accounts. *Price it:* the total sitting past 45 days. It is usually larger than founders expect. ### The customer who left quietly The signals were there for a quarter: usage sliding, the champion gone quiet, two support tickets with an edge in them. Reading them was nobody's job. *Price it:* last year's churn, and the retention number your next raise gets priced on. Pick one. The exercise is worthless if you pick three, because you will build three mediocre things instead of one that works. The one with the biggest number wins, even if it is the least interesting. --- § III  The Conversation ## Sitting down with Claude Now the setup, which is genuinely an afternoon. Get a Claude account with Cowork, connect only the tools this one problem touches (your CRM, your email, your calendar) through the official connectors you can revoke in one click, and give the agent its own folder. Then write one page about how you actually sell: who your customer is, what a good lead looks like, what you say first, what makes you walk away. Every agent you build will read it. Most people skip this and then wonder why the output sounds like nobody. When you start the conversation, bring the bottleneck and its number, not a request for a tool. The difference in what comes back is dramatic. Useful things to ask: frame it Here is what is costing me the most, here is roughly what it costs, here is how the work happens today. ask for options What are three different ways to attack this, and what does each one need access to? push back Which of these would you not build first, and why? What would make this a bad idea? get specific Write the playbook for the one we picked, in the five-question format below. plan the check How will I know in two weeks whether this is working? What number should move? That last one matters more than it looks. If you cannot say what number should move, you have not found the bottleneck yet, and you should go back to the previous section. --- § IV  The Playbook ## Five questions, answered plainly A playbook is the instruction sheet an agent works from. It is not code and it is not a prompt trick. Answer these five in plain language and you have one, and anyone can read the agent's whole job in thirty seconds. runs When it wakes up: a schedule, a trigger, or both reads Exactly which of your systems it may look at writes Exactly what it may change or create, and where authority Whether it acts alone, drafts for your approval, or only reads escalates The conditions under which it must stop and bring you a decision **An example, for the slow-inbound problem** **Runs:** the moment anything arrives from a form, an ad, or the inbox. **Reads:** the enquiry, the sender's website, and whether we have spoken before. **Writes:** a draft reply in my drafts folder, and a line on the record. Never sends. **Authority:** draft only. I send, edit, or delete. **Escalates:** anyone clearly outside who we sell to comes to me as a route-or-decline decision, not a draft. --- § V  The Rules ## Three rules that keep this safe I **Agents act only inside authority you set.** Start narrower than feels necessary. Widen a gate only after you have watched that specific job go right several times. II **Nothing reaches a customer without review.** Draft-only is not a phase you graduate from on a schedule. Some agents stay draft-only forever, and that is the correct outcome. III **Every action leaves a record you can read.** If you cannot reconstruct why something happened, you will not trust it, and untrusted automation gets switched off. The failure mode is never that the agent was not clever enough. It is that it was given room to act before anyone checked what it does with room. --- § VI  Honestly ## What is hard about this Standing the first agent up takes an afternoon. Making it good takes a few weeks of reading its output and correcting it, and that is the part nobody warns you about. The work is the tuning: noticing that a drafted reply is technically correct and sounds like nobody, and knowing what to change. Two other things worth knowing. The first agent will be too eager and needs narrowing, not encouraging. And your records will be messier than you think, so expect the early weeks to be as much about cleaning them up as about agents. None of that should stop you. It is genuinely doable alone, and you will understand your own business better at the end of it than any document could teach you. Atrium atriumagency.io