AI automation

What you get

Software that does the repetitive part of a task and stops where a decision starts. A model reads the message, the record or the report, prepares the next step, and puts it in front of someone on your team. That person approves it, changes it or rejects it, and the system records which.

This is a new service for us, so we have no client results to quote yet. This page describes how the work runs, not what it has achieved elsewhere.

What can be automated

WhatsApp and email reply drafting. Incoming messages are matched to the customer record and answered in draft, using your own price lists, policies and past replies as the source.

Lead follow-up and CRM updates. New enquiries are logged, assigned and followed up on the schedule you set. Call notes and email threads become structured CRM fields, so nobody retypes them.

Business reports and internal workflows. Figures from several tools are written up as a report someone can check against the source. Approvals, handovers and reminders get routed and chased.

Custom AI agents connected to your tools. Where a task spans several systems, an agent can look things up and prepare actions across them. It gets named permissions for named tools, and nothing else.

How a project runs

Pick one workflow. We sit with the people who do the work and write down the steps, the exceptions and what a good result looks like.

Test on your real material. We run past messages and records through a prototype connected to nothing live, and your team marks the output right or wrong. That marked set becomes the test the system has to pass, and it runs again before every later change.

Run beside your team. The system goes live in draft mode. Every output goes to a person, and we read what they changed.

Widen with care. Once a category of work is reliably right, you decide whether it needs a lighter check. New workflows start again at the first step.

Where a person approves

Anything that leaves the business goes past a person by default: a customer reply, a quote, a follow-up, a report sent outside the team. So does anything hard to undo, such as changing a price, closing a record or moving money.

Lower-risk steps, such as filling a CRM field or sorting an inbox, can run without approval if you choose, and they stay logged and reversible. Where the system is unsure, it hands over to your team instead of guessing. You decide where each of these lines sits, and you can move them.

Data and security

Before anything is connected, we write down which data each workflow reads, where it goes and how long it stays. Your customers' messages are personal data, so that record names every outside model provider that sees them, and you approve the list.

Each tool connection has its own credential with the narrowest permission that does the job, kept outside the repository. An agent can only take the actions it was granted, so a message crafted to mislead it cannot reach anything else, and what goes out still passes a person. WhatsApp's business platform has its own rules on opt-in and message templates, and we design inside them.

When automation is the wrong choice

If the process is not settled, automating it makes the confusion faster. If the volume is a handful of messages a day, a person is cheaper and better. If a mistake could harm someone, as in medical, legal or financial advice, the model should not be drafting the answer at all. And if a fixed rule or an ordinary web application can do the job, that is what we will build, because it behaves the same way every time.

What you own at the end

The prompts, the test set, the integration code and the logs live in your repository and your accounts, on infrastructure you control. If the model you started with stops being the right one, the test set shows how a replacement performs before you switch.