1
AI drafts, a person decides
Every change is reviewed by a senior developer who understands it well enough to explain it. Nothing merges on a green tick alone.
How we use AI
The honest version
Most agencies now say they use AI. Very few will tell you how, which model touched your code, what it cost, or what happens when it gets something wrong. This page is our attempt to answer all of that in one place, in plain English, so your technical lead and your procurement team can read the same thing.
The short version: we work AI-first because it makes senior developers dramatically more productive, and we treat it like any other powerful tool, with rules, cost discipline and a human who is accountable for the result.
The rules
Six rules we work to on every project
They are written into how we onboard developers, not pinned to a wall.
2
Tests come first
We write the test, then ask for the code. A model that has to satisfy a real test is a lot less creative with the truth.
3
Flat tooling, no meter
Developer AI tooling is a fixed subscription per person. It is included in your plan and never becomes a metered line on an invoice.
4
Product AI built to be cheap
When your product calls a model, we decide what actually needs a call, what to work out once and store, and what to cache. Choosing well is most of the cost saving.
5
People own the architecture
Models are excellent at filling in a shape. Deciding the shape, the data model and the boundaries stays with the humans who will maintain it.
6
Your code is never training data
We use providers whose terms say so, we name them in your agreement, and we do not send more than the task needs.
Where the line is
What AI never does here
Being good with AI is mostly knowing when not to use it. These are the things a person always does, because the cost of getting them wrong is not worth the minutes saved.
- Deciding architecture and data models
- Approving anything that touches money, auth or personal data
- Final copy and final artwork for a brand
- Talking to you. Your handler is a human, always
lines we don't cross
Cost
The bit nobody talks about
There are two AI bills, and it helps to keep them apart.
The tooling we build with. Every developer here works on a flat subscription we pay per person. It does not scale with how much we use it, it is included in every plan, and it never turns up as a metered line on your invoice. That one is boring on purpose.
The AI inside what we build for you. This is the one that quietly becomes a big number if nobody designs for it. A feature that asks a model the same question on every page load, when it could have asked once and stored the answer. A frontier model doing a job a small one could do. Context stuffed with things the model does not need. We design against all of that from the first conversation: what genuinely needs a model call, what can be computed once and kept, what can be cached and reused, and which model each job actually warrants. It is the difference between an AI feature that costs pennies a month and one that eats the margin.
two bills, one of them boring
Security and data
What your security team will ask
Which providers see our code? Only the ones named in your agreement. Today that is primarily Anthropic, via tooling we run inside our own review and test gates. If that changes, your agreement changes first.
Is it used for training? No. We use commercial terms that exclude training on customer data, and we can share the relevant clauses.
What gets sent? The minimum a task needs. Secrets, credentials and production data never go into a prompt. We scrub and we check.
Where is the audit trail? In your repository. Every change is a reviewed commit with a named human approver, which is more traceable than most human-only teams manage.
Can you sign our DPA? Usually yes, and we have our own with a sub-processor list ready if you would rather start from ours.
Good questions
Things people ask about AI at Ikigai
If yours is not here, ask it on a call and we will add it.
Want the long version?