AI in the operation

Decisions made against your rules and not against a model’s intuition. Agents wired into a specific process, with a person accountable for the outcome and a record of every decision.

Tell us about your project

The problem

An agent can only decide by comparing a new case against explicit rules. If your rules live scattered across spreadsheets, chat threads and the memory of whoever has been there longest, the agent validates nothing: it writes text that sounds right.

How you notice it

If someone asks where your business rules live, the answer starts with "it depends".
A new hire cannot look them up without asking a person.
Your data lives in systems that do not talk to each other.
You already bought an AI tool and cannot say what it decided, or why.

The delivery

What we build

First the data model there is something to decide against: what exists, under which rules, meaning the same thing everywhere. Then the agents on top, wired into a specific process, with a person accountable for the outcome and a record of every decision so it can be reviewed. When the process runs through a conversation, the agent answers on WhatsApp or chat: it replies from a knowledge base of yours, asks what is missing, books the meeting, and at any moment a person takes over without the other side noticing. It serves to screen whoever answers a job ad just as well as to support a customer: the script changes, the system does not.

What you get

The catalogue or data model that was missing, useful even if you remove the AI.
Agents operating inside a process, not in a separate window.
Traceability: what it decided, on which data, and when.

What we have written about this

Where we have done this

Services · AI in the operation

Recruiting · WhatsApp

The agent screens, the person decides

Whoever answers a job ad writes on WhatsApp and an agent replies straight away: it resolves questions with the company’s own information, asks what the role requires, scores and books the interview. The recruiter watches every conversation live and can take any of them without the candidate noticing the change. Underneath there are channel-approved templates, flows per funnel stage, a knowledge base the agent draws its answers from, and candidate search by meaning rather than by exact word. The same machinery serves customer support: the script changes, the system does not.

Scope: Agent, console and knowledge base
Channel: WhatsApp
Status: Online

Editorial content · Agents

From the topic to the published piece

A queue of topics, agents that research and write, images generated for each piece, and a console where a person approves before anything goes out. The site builds itself from what was approved. The same machinery feeds an email newsletter that assembles itself from the business data and goes to its distribution list.

Scope: Agents, site and console
Control: Human approval
Status: In operation

Marketing operations · Catalogue and agents

A single catalogue, then the agents

A company running campaigns with partner brands kept its creative assets, eligibility rules and campaign briefs spread across spreadsheets and chat threads. We built the catalogue that holds brands, assets, campaigns and rules in one place, and on top of it the chain of agents that takes a brief in by chat, checks it is complete, validates every piece against the rules and turns it into tickets the team can execute.

Scope: Data model + agent chain
Intake: Chat to tickets
Status: Delivered

How we work

01 First contact
02 Diagnosis
03 Proposal
04 Sprint build
05 Ongoing support

See the full process

Ways to work together

Diagnosis sprint 2 weeks
Custom build From 6 weeks
Monthly team Ongoing

See the models

45 minutes to find out whether this applies to you

Four questions first, so the call starts where it matters instead of at the beginning. If in those 45 minutes we see this is not for you, we say so right there.

Answer four questions