---
title: "AI in client systems: building with it, not around it."
date: 2025-10-14
url: https://remiam.co.uk/notes/ai-in-client-systems
tags: [AI, Client Systems, Architecture]
read_time_minutes: 7
description: "Where AI actually earns its place in client systems in 2025 — the integrations that survive contact with production, and the ones that don't."
---

# AI in client systems: building with it, not around it.

*Published 2025-10-14 · 7 min read · by Liam (Remiam)*

Clients in 2025 don't ask us 'should we use AI?'. They ask us where it goes, what it costs to run, and what it can't do safely. Here's the answer we've been giving.

Three years ago the AI conversation with a client was about whether to. Now it's about where, how, and what guardrails to ship around it. The technology is no longer the interesting variable. The interesting variable is what you ask it to do, and what you don't.

## Where AI is genuinely earning its keep

- Customer-facing search that understands intent, not just keywords.
- Operational summarisation — turning long email threads, support tickets, or call transcripts into structured records the team can act on.
- Content classification at scale — tagging, routing, deduplicating user submissions.
- Drafting — copy, code, configuration, briefs — for a human to then refine.

## Where we are careful

- Anything load-bearing for compliance or finance. Humans approve, then sign.
- Customer-facing decisions about a real person — credit, eligibility, exclusion. Hard no on full autonomy.
- Anything where 'wrong but confident' is worse than 'I don't know'. Build the 'I don't know' explicitly.

## What the integrations look like

The patterns that survive production share a shape: a tight scope, a fallback path that doesn't depend on the model, observable behaviour, and a human-shaped escape hatch on every action that matters. The flashier 'AI does everything' demos burn out within months. The boring ones quietly become the most-used feature in the product.

## A grounded pattern, in shape

| Layer | Responsibility | Failure mode |
| --- | --- | --- |
| Input validation | Reject inputs that fall outside model scope | Returns "out of scope" to the user |
| Retrieval | Find grounded source documents | Returns explicit "no matching context" |
| Model call | Generate with retrieved context + scope | Times out, retries with backoff |
| Output validation | Schema-check, hallucination flag | Logs + falls back to non-AI path |
| Human review | For anything load-bearing | Default action requires sign-off |

*A production AI-feature shape that survives real users.*

> Pick one workflow that costs the team an hour a day. Wire AI into it as a draft generator. Have a human approve. Iterate from there. It's less exciting than the marketing decks. It's also the version that ships.

If you're a client wondering where to start: pick one workflow that costs the team an hour a day. Wire AI into it as a draft generator. Have a human approve. Iterate from there. It's less exciting than the marketing decks. It's also the version that ships.

## References

1. [OpenAI — production usage guidelines](https://platform.openai.com/docs/guides/safety-best-practices)
2. [Anthropic — Claude API documentation](https://docs.claude.com)