AI agency - Edinburgh

AI agency in Edinburgh

Edinburgh has more long-dated liability data than almost any city in Europe, and most of it is trapped in systems that predate the people running them.

The Edinburgh market is asset management, life and pensions, and general insurance, sitting next to a university pipeline that produces good machine learning graduates every summer. That combination creates a specific question: build the capability internally or buy it. The honest answer depends on what the first three projects are, not on headcount cost.

The back book is the asset and the obstacle

Life and pensions firms hold decades of policy documents, scheme rules, correspondence and amendments, often across several acquired books with incompatible conventions. Answering an apparently simple customer question can require reading four documents written under three regimes. Retrieval over that estate is the highest-value AI work available in the city, and it is also the hardest, because the failure mode is a confident answer drawn from a superseded document.

The fix is structural rather than clever. Version and date every source, make recency and applicability part of retrieval rather than an afterthought, and require the system to name the document and clause behind any answer. If it cannot cite, it should decline.

Model documentation is the deliverable regulators read

Firms here already run model risk governance for pricing and capital. Language models land in the same committee, and the committee asks familiar questions: what does it do, what data trained or grounds it, how was it tested, what are the known failure modes, who monitors it, and what triggers a rebuild. A build that produces this documentation as a by-product moves through governance. A build that treats documentation as a final task stalls there.

Hire or buy

• Buy when the work is one-off engineering: retrieval architecture, evaluation harnesses, production plumbing and integration. It is slow to hire for and you need it once.
• Hire when the work is continuous and domain-specific: evaluation set curation, monitoring, and deciding what the system should refuse to answer. Those judgements belong inside the firm.
• The bad outcome is hiring two engineers to build infrastructure that already exists, then losing them because the interesting problem finished in month four.

Digiton works as the first category and hands over the second, with production deployments across 8 countries. A short AI audit maps which of your first projects are one-off and which are permanent.

Frequently asked questions

Related