MORE DONE WITH THE PEOPLE YOU HAVE
AI Automation & Agents
Your team gets more done without getting bigger: automation that holds in operation, and agents only where they have earned their place, bounded and evaluated. You own what we build, and we run our own company on it. Autonomy is earned here, not assumed.
THE THESIS
Agents are the new pilots.
The pattern is familiar. The demo impresses. The initiative stalls before production. Most agent projects fail for the same old reasons: integration, permissions, evaluation, and no one who owns the result. The technology moved fast. The last mile didn't move at all.
Our position is calm on purpose. The capability is real, the timing of the breakthrough is unknowable, and the winners won't be the organisations that believed hardest. They will be the ones whose automations actually run. We build with this technology every day. We just refuse to sell it as magic.
THE LADDER
Automation first. Agents where they earn it.
Three rungs, in rising order of freedom and falling order of predictability.
An agent earns its autonomy the way a new employee does: small scope, supervised, expanded as it proves reliable. Anyone who deploys the other way round is running an experiment on their own operation.
The most underrated AI decision is choosing not to use AI for it. What you should get from AI automation consulting is the rung your problem actually needs, and the honest answer is often the cheapest one.
- 01
Deterministic automation
Rules, triggers and integrations. Cheap, reliable, boringly effective, and more often the right answer than the market admits.
- 02
AI inside a defined step
A model doing one thing well, drafting, extracting, classifying, with a rule or a human on both sides of it.
- 03
Agents
Goal-directed, multi-step, tool-using. For work where variability defeats rules, and only there.
WHAT WE BUILD
Built to be owned. Agents especially.
Triage and routing of incoming matters. Drafting with approval steps: replies, summaries, documents. Extraction and structuring from documents into the systems you already run. Recurring reporting and follow-up across systems. Multi-step workflows that span your existing tools instead of replacing them.
All of it delivered the way we deliver everything: built to production criteria, run until it holds, and transferred with the evaluation framework, the logging and the documentation your team needs to own it.
That matters twice as much here, because agents drift. Models change, prompts age, integrations move. An agent that only its builder can keep honest is a dependency, not an asset. What we hand over includes the migration readiness to keep yours honest without us.
WHAT WE RUN OURSELVES
We run our own company on this.
Our sales pipeline, planning, marketing production, finance administration and client-delivery tooling run on automation and agents we built ourselves, on the same principles we sell: bounded autonomy, approval where actions touch the real world, evaluation before expansion.
It is our own operation, so we feel every failure mode first: the drift, the model swaps, the integration that breaks on a Tuesday. That is the point. Ask us to show you.
AUTONOMY AND GOVERNANCE
Freedom needs boundaries. By design.
An agent that acts, sends, books, writes to records, needs its boundaries designed in, not bolted on: approval points where actions touch the real world, logging that survives an audit, an evaluation framework, and clarity on what the system may decide alone.
Where it interacts with people, transparency duties may apply, and our EU AI Act page carries the details. All of it is built during Build and transferred with the system.
QUESTIONS
Before you let an agent into live systems.
Automation follows rules you set, while an AI agent pursues a goal across several steps and picks its own tools along the way. We work from a ladder of three rungs: deterministic automation built from rules, triggers and integrations; AI inside one defined step, such as drafting or extraction, with a rule or a person on both sides of it; and agents, reserved for work where variability defeats rules. Most problems belong on the first or second rung.
Automation and AI agents suit repetitive work that crosses system boundaries. We build five kinds: triage and routing of incoming matters, drafting with approval steps for replies, summaries and documents, extraction and structuring of data from documents into the systems you already run, recurring reporting and follow-up across systems, and multi-step workflows that span your existing tools. Nothing is ripped out to make room for them.
An AI agent earns its autonomy in stages, the way a new employee does. It starts with a small scope, approval points wherever an action touches the real world, logging that survives an audit, and an evaluation framework that runs before the scope grows. Freedom expands once reliability is proven, never before. We hold the same boundaries for the agents that run our own sales pipeline, planning and finance administration.
Yes: an AI agent that interacts with people falls under the transparency duties in the EU AI Act, which apply from 2 August 2026. People have to be told when they are dealing with an AI system, and certain generated content has to be marked. We design that disclosure into the agent from the start rather than bolting it on afterwards, and the duty transfers with the system alongside the logging.
An AI agent feels a model change more than anything else we build: prompts that worked start to drift, and evaluations have to be rerun. That is why the handover includes the evaluation framework and a migration plan, alongside the code, the prompts, the logging and the documentation. Your team, or any competent partner, can move the agent when the market moves. An agent only its builder can keep honest is a dependency, not an asset.
The cost of automation or an AI agent is set by which rung of the ladder the work actually needs, not by how advanced it sounds. Deterministic automation built from rules, triggers and integrations is the cheapest and lasts longest. AI inside a defined step costs more, and a multi-step agent most, because the evaluation framework and the approval points are part of the build. Scope and price are agreed before we start, and if a rule beats an agent we say so. The honest answer is often the cheapest one.
You own the AI agent outright: the code, the prompts, the evaluations, the logging and the documentation. There is no dependency on us and no retainer needed to keep it running. Transfer is part of the assignment rather than an extra, and it counts as finished only once someone on your side has run the agent alone, without us in the room.
INSIGHTS
Recent analysis on this work.
Why last year's benchmark figure cannot be compared with this year's
The model releases got the attention. Meanwhile almost every serious measurer published a correction to its own instrument, and one withdrew its measure entirely. That decides how a vendor's figure should be read.
Read the analysis
When an AI agent is ready for production, and when a rule is enough
A benchmark saying an AI agent can do the task answers the wrong question. The question that decides the cost is how often it does, and there is a measure for that.
Read the analysis