Intelligent
automation.

Automate the decision, not the screen

A great deal of what gets sold as automation is a script driving a user interface. It works until the interface changes, and then it fails silently in a way nobody notices for a fortnight. The durable version automates the decision and integrates at the data layer — an API, a queue, a database — so it does not depend on a button staying in the same place.

The decisions worth automating share a shape: high volume, repetitive, currently made by a human applying consistent criteria to available information, and cheap to reverse when wrong. Invoice coding, document classification, ticket routing, eligibility pre-checks, data reconciliation. The value comes from removing the queue rather than from the sophistication of the model.

Equally important is knowing which decisions to leave alone. Low-volume decisions do not repay the engineering. Decisions where the criteria are genuinely contested inside your organisation cannot be automated, because there is no consistent rule to learn — that is a process problem wearing a technology costume, and automating it just makes the inconsistency faster.

Confidence banding is the pattern that works

The failed version of automation is binary: the system handles everything, or a human does. The version that survives contact with reality routes by confidence.

In practice that means three bands. High confidence cases process automatically with a sampled audit. Middle cases go to a human with the model's assessment and reasoning attached, which is faster than an unassisted review because the work is verification rather than analysis. Low confidence and novel cases go to a human with no suggestion at all, because a wrong suggestion is worse than none — it anchors the reviewer.

This structure has a property that matters more than its accuracy: it degrades gracefully. If the model gets worse, more work flows to humans and throughput drops. It does not start making confident errors at scale. And because every human decision in the middle band is a label, the system generates the training data to widen the automated band over time. You start conservative and earn automation with evidence.

Where language models changed what is possible

The genuine shift is in handling unstructured input. Extracting structured facts from an invoice, a contract, a medical report, or an email thread previously needed either rigid templates that broke on format variation, or a bespoke NLP pipeline that took months. A well-specified model with a defined output schema now does that job across format variation with far less engineering.

That said, we are deliberate about model choice rather than reaching for the largest available. For a stable high-volume classification task with labelled examples, a small fine-tuned model is typically faster, an order of magnitude cheaper per call, and — importantly for automation — deterministic, which matters when you need to reason about behaviour. A common pattern is to use a large model to label and bootstrap, verify those labels, then deploy the small model that actually runs in production.

One caution we raise early on any automation touching untrusted input: content the system reads — emails, uploaded documents, web pages — may contain instructions the model follows. Treat it as untrusted data, never let it authorise an action, and keep tool access narrowly scoped. See AI agents for how we contain that.