LLM integration for systems that have to work on Tuesday morning

Who this is for

This is for businesses with real systems — a product, an internal platform, a heavy document workflow — where AI capability would change the economics, but only if it runs reliably, safely, and at a predictable cost.

• —Document-heavy workflows: extracting details from invoices, applications, contracts, or reports.
• —Customer communication that needs drafting or triage at a volume people cannot sustain.
• —A software product that needs an AI feature without betting the roadmap on a prototype.
• —Search or Q&A over your own knowledge — manuals, policies, job history — that staff currently dig for by hand.
• —An existing AI prototype that works in a demo and falls over on real data.

Problems this work typically solves

1. 01

   From demo to dependable

   Prompts become versioned, tested components with fallbacks and human review where it matters — so output quality stops being a daily surprise.

2. 02

   Data boundaries you can defend

   Explicit control over what the model sees, what is logged, what is retained, and which provider processes it — written down, configured, and aligned with the Privacy Act 2020.

3. 03

   Costs that stay boring

   Model selection, caching, and routing tuned so the bill scales with value rather than with traffic. The cheapest capable model wins; the expensive one is reserved for the work that needs it.

If your prototype impressed the room and now scares the engineers, that is the normal next step — and it is fixable.

What a first engagement looks like

Honest scope: one integration, end to end, over a few weeks — proven on your real data before anything is promised about the rest of the roadmap.

1. 01

   Scope the one integration that pays

   A working session on the candidate workflows, the systems involved, and the failure modes that matter. Output: one integration worth building first, with success criteria you can measure.

2. 02

   Build it production-shaped from day one

   Evaluation on your real data, structured outputs, retries and fallbacks, logging, and cost ceilings — the unglamorous parts that make it dependable, built in rather than bolted on.

3. 03

   Ship, measure, hand over

   Live on real volume with monitoring your team can read, documentation, and a named owner. If the measured quality does not justify expanding, that recommendation is part of the deliverable.

When another path is better

• If there is no existing system to integrate into, custom web apps builds the home first.
• If the work is between off-the-shelf tools rather than inside your software, AI automation gets there with far less engineering.
• If you want an autonomous agent doing multi-step work rather than a capability inside a system, see AI agent implementation .

Frequently asked questions

Which LLM provider should we use?

The one your task, data constraints, and budget point to — often a mix, with a cheap model for volume work and a stronger one for the hard cases. Nelson AI has no vendor stake; the recommendation comes from evaluation on your data.

Can we keep our data in New Zealand?

Most major LLM providers process offshore, so the honest framing is control rather than geography: strict minimisation of what the model sees, contractual terms on retention and training, and disclosure that matches the Privacy Act. Where local or self-hosted options genuinely fit, they are on the table.

How do you stop it making things up?

By design rather than hope: grounding answers in your documents, constraining outputs to structured formats, scoring quality against an evaluation set, and routing low-confidence cases to a person. Hallucination is managed the way any defect rate is — measured, then engineered down.

What does an integration cost to run?

Usually less than people fear once routing and caching are in place — token costs are falling, and most business workflows need far fewer tokens than a chat transcript suggests. The build includes a cost model so there are no surprises.

Bring the workflow and the worry.

A short call is enough to tell whether an LLM integration is justified, what it would take to run dependably, and what it should cost — including when the answer is 'not yet'.