Insights

Notes from actual builds

Working notes, research and playbooks on the parts of this work that are genuinely hard. No thought leadership, no predictions about the year ahead.

This section is being written, and we are not going to pretend otherwise

Below is the actual queue rather than five hundred words of filler with a stock photograph on top. Each piece comes out of work we have done and gets published when it is worth reading, not on a content calendar.

Want one of these sooner? Say which and we will prioritise it.

In the queue

What we are writing

Why your deflection rate will not be 40 percent

Vendor decks imply thirty to fifty percent. Most teams see ten to fifteen in year one. The variable that closes the gap is not the model.

Research · in progress

When not to fine tune

Four questions that end most fine tuning projects before they start, and why ending them is the right outcome.

Note · in progress

Build the evaluation set before you build anything else

Without a fixed set of questions with known answers, every release is judged by whoever tried it that afternoon.

Playbook · in progress

Your AI bill is one number and that is the problem

Attribution before optimisation. How to instrument spend per feature, per tenant and per workflow, and what usually falls out when you do.

Playbook · in progress

Filtering results is not access control

The moment you copy content into a search index you have created a second access control surface. Most teams do not notice.

Note · in progress

Ask us the hard question

Most of these started as a question a client asked that we could not answer in one sentence. Send us yours.