AI automation that takes repetitive work off your team

• ·Your team answers the same questions every day, on WhatsApp, email or chat.
• ·Someone reads documents, forms or messages and types the data into another system by hand.
• ·You tried ChatGPT or a chatbot tool. It helped one person, or it worked in a demo and fell apart with real customers.
• ·You work in an area where a wrong answer has consequences, such as health, law or finance, so a person has to stay responsible.

Getting an AI model to answer well once is easy. Getting it to answer well for the thousandth customer, within a budget, without leaking data and with a way for a person to step in, is most of the work.

Off-the-shelf chatbots don’t know your rules, can’t reach your systems and fail quietly. A useful assistant needs access to your data, clear limits, a record of what it did and a review screen your team will use every day.

Every hour spent on repeatable work is an hour your team doesn’t spend on the customers and cases that need judgment.

When a customer is comparing options, a reply that comes the next morning competes with the ones that came in minutes.

An assistant that answers alone and gets something wrong in a sensitive area costs more than it ever saved. That risk is why many teams never start.

• ·The repetitive part arrives already drafted, and your team reviews, edits and sends.
• ·Answers follow your rules and your data, and you can see what the assistant did and why.
• ·You know what it costs to run each month, with limits so the bill can’t grow unnoticed.

1. 01A working assistant or automation connected to your data and systems, doing one clear job.
2. 02Human approval where it matters: drafts, edits, approvals and a clear path to hand a case to a person.
3. 03The channel your team and customers already use: WhatsApp, email, web chat or your internal system.
4. 04A test set built from your real cases, so a new prompt or AI model is measured before it goes live.
5. 05A log of what the assistant did, cost tracking and limits per user or task.
6. 06Documentation: what it does, what it refuses to do, and how to switch AI provider later.

1. Step 1: 01

   Short

   Pick one job

   One concrete task, the data it needs and what a wrong answer would cost. We start with the smallest version worth using.

2. Step 2: 02

   Early

   Test on real cases

   A first version runs against examples from your business, so its mistakes show up before customers see them.

3. Step 3: 03

   Most of the project

   Build around it

   Connections to your systems, the review screen, the channel, the logs and the limits.

4. Step 4: 04

   As long as agreed

   Pilot

   Your team uses it on real work and approves every output. Each mistake becomes a new test case.

5. Step 5: 05

   Short

   Hand over

   Code, prompts, tests and instructions in your accounts.

• ·One clear task, and someone on your team who knows what a good answer looks like.
• ·Real examples: conversations, documents or tickets, anonymized where needed.
• ·Accounts in your company’s name for the AI provider and the channel (for example WhatsApp Business), so keys and billing stay yours.
• ·Your privacy rules: what data may reach an AI provider, how long it’s kept, and what LGPD or GDPR require in your case.

Public work where the hard part was the system around the AI.

• Ludh (ludh.com.br)External site

  I built the agent orchestration for Ludh’s AI assistant: the AI drafts replies to patients, and the doctor reviews and approves each one on WhatsApp or the web before it’s sent.

• editkit-tsExternal site

  An open-source library I wrote for applying code edits that AI models generate, including edits that don’t match the file exactly.

How much does an AI automation cost?

There are two costs. Building it is priced per project, from a written proposal. Running it depends on volume and on the AI model, and I measure it while testing on your real cases, so you see a monthly estimate based on real numbers before the pilot starts.

How long until it’s working?

A first version running on your real examples comes early, because that is where the risks show up. The full build depends on how many systems it connects to. The proposal gives dates for your scope.

We use a specific CRM, helpdesk or ERP. Can it connect?

If the system has an API, an export or a webhook, it usually can. I check before the proposal, and I tell you when a connection isn’t possible or would cost more than the work it saves.

We already have a developer. Why bring you in?

A demo is quick to build. Making it hold up with real users takes work a busy developer rarely has time for: tests from real cases, review screens, retries, cost limits. I can build it and hand it over, or work with your developer so they own it afterward.

What if the AI gets something wrong?

It will, sometimes. No model is free of mistakes, so the system limits the damage: it answers from your documents, says so when the data isn’t there, and sends anything that matters to a person for approval. That is how Ludh works: no reply reaches a patient without the doctor’s approval.