Service — AI Agents
Custom AI Agents
Built Around Your Process
August builds custom AI agents for companies in Miami, Buenos Aires and Madrid — software that carries out a real task end to end instead of another dashboard someone has to remember to open. We start by mapping the process and telling you which parts are worth automating, including the ones that are not.
01 — Definition
What a custom AI agent actually is
An AI agent is software that completes a task rather than answering a question. It reads the input, decides what the case is, gathers what it needs from the systems you already run, and either finishes the job or escalates it to a person with the context attached. Four things have to be true for one to be worth building.
It happens often enough
Automation earns its cost on repetition. A task that runs twenty times a day pays for itself in weeks; one that runs twice a month almost never does, no matter how tedious it feels. The first thing we look at is not how annoying the work is but how often it happens.
It needs a decision, not a rule
If the task can be written as an if/then table, it does not need a language model — it needs a script, and a script is cheaper and more predictable. Agents earn their place where the input is messy: free text, a photo, a conversation, a document that never arrives twice in the same shape.
The data is reachable
An agent is only as good as what it can look up. Before writing anything we check that the information the task depends on can actually be read — and that it is the same in every system that stores it. This step is where most automation projects quietly fail, so we do it first rather than last.
The blast radius is known
Every agent we build has an explicit line between what it decides alone and what it must ask about. Reversible, low-stakes work runs unattended. Anything touching money, a promise to a customer or a permanent record stops for a human. The line is agreed before the build, not discovered after.
02 — Método
How we work
Three stages, in this order. The first is a deliverable on its own: if you stop after the diagnosis, you still leave with a written map of your own process and an honest read on what is worth automating.
01
Process diagnosis
We sit with the people who do the work and write down what actually happens, not what the process document says happens. Deliverable: the mapped process, the volume of each step, the systems involved, and a shortlist of candidates ranked by payback — with the ones we recommend against and why.
02
Build and pilot
One agent, one task, running alongside the person who does it today rather than replacing them on day one. We compare its output against theirs on real cases until the disagreements are understood. Nothing goes unattended until we can explain every case where it was wrong.
03
Operate and extend
Once it is stable it needs an owner, not just a deployment: logs someone reads, a threshold that raises a hand when behaviour drifts, and a monthly review of what it handled and what it escalated. Extensions come from that log — from what the agent actually met, not from a roadmap written in advance.
03 — Use cases
What we are usually asked to build
Concrete shapes rather than categories. Each of these is a task someone is doing by hand right now, often at a volume that quietly consumes a role.
Lead qualification on WhatsApp
Answers the first message in seconds at any hour, asks the three questions your sales team always asks, and hands over a qualified conversation with the answers already recorded — or books the call directly.
Shared inbox triage
Reads everything arriving at a shared address, classifies it, answers what is routine with the answer your team would have given, and routes the rest to the right person with a summary attached.
Reporting across platforms
Pulls the numbers from the platforms that never agree with each other, reconciles them against one definition, and writes the recurring report — including the sentence explaining what moved and why.
Document and form intake
Takes the invoice, the form or the photo of a receipt that arrives in a different shape every time, extracts the fields, checks them against what is already on file, and flags only what does not reconcile.
Monitoring and alerts
Watches a set of sources — competitors, prices, listings, reviews, a regulator — and tells you what changed and whether it matters, instead of sending a digest nobody reads.
Internal knowledge lookup
Answers the question a new hire would otherwise interrupt someone to ask, from your own documents, with the source cited so the answer can be checked rather than trusted.
04 — FAQ
Frequently asked questions
What is a custom AI agent?
A custom AI agent is a piece of software that carries out a task on its own, end to end, using a language model to make the judgement calls a rules-based script could not. It reads the message, decides what it is about, looks up what it needs, and either completes the task or hands it to a person with the context already gathered. The difference from a chatbot is that a chatbot answers; an agent does the work and leaves a record of what it did.
Why build one instead of buying a tool?
Because a tool makes you adopt its process, and the process is usually the part of your business that actually differentiates you. Off-the-shelf platforms are the right answer for problems everyone has in the same shape — email, calendars, invoicing. They are the wrong answer when the work depends on your criteria, your data and your handoffs. We build on top of what you already run rather than asking you to migrate to it.
What kinds of work do you automate?
The repetitive, judgement-light half of a role that a person currently does between the tasks that need them. Qualifying and routing inbound leads, answering the same product questions on WhatsApp at three in the morning, pulling numbers from several platforms into one report, classifying and drafting replies to a shared inbox, monitoring a set of sources and flagging what changed. The test we apply is simple: if a competent person could do it from a written instruction, an agent can usually do it too.
What happens when the agent gets it wrong?
It gets it wrong sometimes — that is a property of the technology, not a bug we will promise away. What we design for is the cost of being wrong. Anything reversible and low-stakes runs on its own. Anything that touches money, a commitment to a customer or a permanent record stops and asks a person first. Every run is logged with its input, its decision and its output, so a mistake can be found and corrected rather than discovered months later.
Do we need our data in order first?
Less than people fear, but not none. An agent needs to be able to reach the information the task depends on and it needs that information to be true; it does not need a data warehouse or a migration project. In practice the first engagement usually surfaces two or three places where the same fact is stored differently in two systems. Finding those is part of the work, and it tends to be worth doing whether or not the agent ever ships.
How do you charge, and how long does it take?
The entry point is a scoped diagnosis: we map the process, agree on what the agent will and will not decide, and come back with the build cost and the timeline before anything is built. Small, single-task agents are a matter of weeks; anything that touches several systems is longer and we say so upfront. We would rather tell you a process is not worth automating than sell you the build.
05 — Contact
Start with the process, not the tool
Tell us which task is eating the most hours and we come back with whether an agent is the right answer for it — and what it would take. If the honest answer is that it is not worth automating, that is the answer you get.