What you will take away
- The best cases are internal, high-volume and verifiable.
- Structured extraction from documents is the most reliable early win.
- Cost and data boundaries need designing before the pilot, not after.
Why marketing copy is the least interesting use
Content generation is the visible use because it is easy to demonstrate. It is also the one with the weakest economics: the output needs heavy editing, quality is subjective, and the volume rarely justifies a programme.
The applications that pay are internal, boring and measurable - work that happens thousands of times a month, where the output can be checked automatically or by the person who already does the task.
Document processing, which is where most of the value is
Invoices, contracts, delivery notes, claim forms and applications arrive as unstructured files and get rekeyed by people. Models now extract structured fields from these reliably enough to change the workflow, especially when the output is validated against business rules and low-confidence extractions are routed to a human.
The measurable effect is straightforward: handling time per document, error rate, and the size of the queue at month end. Those are numbers a finance director already tracks, which makes the business case short.
Automate the ninety per cent of documents that are routine and route the rest to a person. Chasing the last ten per cent is where these projects go to die.
Four more that work in practice
- Internal knowledge retrieval over policies, runbooks and past tickets, which shortens onboarding and reduces interruptions to senior staff.
- Synthetic test data that is realistic in shape but contains no personal data, which unblocks development and testing under privacy rules.
- First-pass code review, catching mechanical issues before a human reviewer spends attention on them.
- Drafting support replies for an agent to edit and send, where the agent stays accountable for what goes out.
The two constraints to design for up front
The first is data boundary. Decide before the pilot which data may leave your environment, what is retained by the provider, and whether a self-hosted model is required. Retrofitting that decision means rebuilding the integration.
The second is unit cost. Per-token pricing behaves very differently from per-seat licensing when volume grows. Model the cost at full rollout volume, not at pilot volume, and design prompts and retrieval for efficiency early - it is much harder later.
How to pick the first project
Choose something high volume, low risk and easy to verify, with a named owner in the business who wants it. Run it against the current process in parallel for a month and compare on the metric that owner already reports.
A pilot that quietly saves a department six hundred hours a year is worth more than an impressive demonstration nobody owns, and it is the one that gets the second project funded.
OrbixEdge Engineering
The engineers who build and run the systems we write about - platform, cloud and product teams shipping for clients in retail, health, logistics and fintech.
