01
The failure we design against
The characteristic failure of current AI systems in professional work is not a wrong answer. It is a fluent answer whose basis cannot be recovered. A reviewer who cannot see which source was used, what the system did not look at and how confident it had any right to be is not reviewing anything — they are deciding whether to trust a sentence.
02
What we build instead
Our systems are built so that the source, its version, its effective date, the coverage of the search and the uncertainty of the result travel with the output. A refusal is a legitimate result: a system that declines to answer outside its corpus is more useful than one that improvises, because the first can be relied on and the second has to be checked every time.
03
Why affordability is a design constraint
The same capability priced for a large bank is simply unavailable to the firms that need it most — the practices, chambers and mid-size institutions doing substantially the same regulated work with a fraction of the budget. Bounded resources, explicit refusals and cheap failure modes are what let one system serve both, which is why we treat them as design constraints rather than as cost optimisations to be applied later.
04
From our directors
Our ambition is narrow on purpose. We are not trying to build a general assistant; we are trying to make the difficult, consequential parts of professional work — finding the right source, knowing what changed, understanding what it affects and being able to show your working — genuinely dependable, at a cost that a firm of any size can carry. Bhavika Mukherjee and Aditi Chadha, directors.
Disagree with us? Tell us why.
We publish the measurement, its date and its scope so you can check our reasoning. If you think we got something wrong, write to us.