Skip to content
How we work

Six principles. Four stages. Fourteen days.

We’ve run this sequence 60 times across e-commerce, subscription and marketplace businesses. It is deliberately boring, because the projects that ship are the boring ones.

04How we work

Signed spec to live in 14 days.

No six-month discovery. No pilot that never leaves staging. A fixed, four-stage sequence we’ve run 60 times.

  1. 01Day 0–3

    Audit

    We sit in your tools for three days. Ticket volumes, approval loops, where the week actually goes. You get a ranked map of every automatable workflow with hours and euros attached — yours to keep whether or not you hire us.

    Automation map + business case

  2. 02Day 4–7

    Spec

    We pick the top three by leverage, not by novelty, and write the spec: inputs, decision boundaries, escalation rules, what the agent must never do, and the metric it will be judged on.

    Signed agent spec + eval set

  3. 03Day 8–14

    Build

    We build inside your infrastructure — your cloud, your keys, your repo. Retrieval over your real data, evals before launch, a shadow run against live traffic before anything touches a customer.

    Agent live in production

  4. 04Ongoing

    Run

    Agents drift when your business changes. We monitor accuracy, review escalations weekly, retrain against new edge cases, and report the number in the same format every month.

    Monthly performance report

05Principles

The rules we don’t break for anyone

01

You keep the keys

Every agent runs in your cloud, under your API keys, in a repo you own. If you end the engagement tomorrow, everything keeps running and your team can maintain it. No black boxes, no per-seat licence, no hostage data.

02

The metric comes before the build

Every spec names the number it must move and the threshold it must clear. If a launched agent misses at the 30-day review, we fix it at our cost until it clears or we refund the build fee.

03

Delete before automate

The three-day audit usually finds one or two workflows that should be removed rather than accelerated. We tell you that before there's a contract to protect. Automating a process that should have been deleted is the most expensive way to speed it up.

04

Humans keep the exceptions

We build escalation first, not as a fallback. Good handoff — with a written summary, the retrieved context and the actions already taken — is what lets you deploy conservatively. Conservative deployment is what keeps CSAT intact.

05

Model-agnostic by design

We route per step and re-benchmark quarterly. Switching providers should cost you a config change, not a rebuild. You're never locked to one vendor's roadmap.

06

We run our own agency on it

Every system we sell runs inside our own operation first. If it doesn't survive a real Tuesday here, it doesn't reach a client.

06The fine print

Runs inside your stack. You keep the keys.

Most security reviews we go through take about a week, because there isn’t much to review — nothing leaves your infrastructure.

Runtime
Your cloud — AWS, GCP, Azure or on-prem. EU-hosted available.
Models
Routed per step and re-benchmarked quarterly. No provider lock-in.
Data
Never used to train third-party models. DPA in place, minimised at the retrieval layer.
Observability
Every agent action logged and attributable. Full audit trail from day one.
Handover
Repo, infrastructure-as-code, eval sets, runbooks and documentation.
Compliance
ISO 27001 aligned. GDPR. SOC 2 Type II in progress.
06In their words

Operators don’t buy demos. They buy the week back.

We'd been quoted six months and a seven-figure number by a consultancy. AIMAGENTIC had the triage agent answering live tickets in eleven days, and it was better than our macro library on day one.
Head of CX·D2C apparel, €40m GMV
The part I didn't expect: the audit alone was worth the fee. They showed us two workflows we should just delete rather than automate. Nobody selling AI tells you that.
COO·Subscription nutrition
Everything runs in our AWS account under our keys. Our security team signed it off in a week, which has never happened with a vendor before.
CTO·Marketplace, 40k SKUs
Monday used to cost me six hours of pulling numbers. It now costs me the time it takes to read one page. That's my whole review of the engagement.
Founder·Home & living
07Questions

The eight we always get asked

Fourteen days from signed spec to production for a first agent. Days 0–3 are the audit, days 4–7 the spec and eval set, days 8–14 the build, shadow run and launch. Complex retrieval or bespoke integrations can add a week — we tell you that before you sign, not after.

You do. Full ownership of the repository, the infrastructure, the prompts, the eval sets and the documentation. Everything runs in your cloud accounts under your API keys. If you end the engagement tomorrow, the agents keep running and your team can maintain them.

Whichever wins the eval for that specific task. We route per step — a cheap fast model for classification, a frontier model for anything customer-facing or judgement-heavy — and re-benchmark quarterly. You are never locked to a single provider, and switching costs you a config change rather than a rebuild.

Three things. Retrieval grounded in your own systems, so answers are drawn from live order and policy data rather than model memory. Hard decision boundaries, so the agent can only take actions on an allow-list. And a confidence threshold that hands off to a human with a written summary rather than guessing.

It replaces the part of their week that they hate. In every deployment we've run, headcount held and the work changed: support leads move onto retention and product feedback, marketers move from asset QA onto strategy. We'll tell you honestly during the audit if a role is genuinely at risk.

A fixed fee per agent, scoped after the audit, plus an optional monthly run-and-improve retainer. The audit itself is a flat €2,500 and is credited in full against the first build. No per-seat licence, no percentage of savings.

Everything runs inside your infrastructure and your data never trains a third-party model. We work under a DPA, apply data minimisation at the retrieval layer, log every agent action for audit, and can deploy fully within the EU. Our own operation is ISO 27001 aligned and SOC 2 Type II is in progress.

The spec names the metric and the threshold before we build. If a launched agent misses its threshold at the 30-day review, we fix it at our cost until it clears or we refund the build fee. That's why the audit exists — we only take work we're confident we can hit.