What AI Automation Actually Costs When You Factor in the IT

A Pacific Northwest manufacturing client asked us last month if they should automate their order confirmation emails with an LLM. The emails were already templated. Variables pulled from their ERP, same five sentences every time, just different product codes and ship dates. They wanted AI because their competitor mentioned AI in a webinar. We told them no. The template was fine. What they actually needed: their ERP to stop dropping the ship date field when the warehouse code contained a slash character. That was a thirty-minute fix, no model required.

This keeps happening. The AI pitch arrives polished. The infrastructure underneath is not ready for it.

The Gap Nobody Mentions

You can sign a contract with an automation vendor tomorrow. They'll demo a system that reads your invoices, extracts line items, writes them to your accounting software. It works in the demo. It falls apart in production because your invoices come in as scanned PDFs with coffee stains, the OCR layer chokes on your vendor's dot-matrix printer output, and your accounting API has an undocumented rate limit that gets hit whenever the system processes more than twelve invoices in ten minutes. The vendor didn't test for that. They tested with clean PDFs, typed text, no rate limits.

The work that makes automation reliable is IT work. Boring, specific IT work. Data pipelines that handle malformed inputs. Error queues that route edge cases to a human. Logging that tells you which vendor invoice broke the parser at 2am. Monitoring that catches when your token spend is 4x projections because someone's pasting entire contracts into a summarization tool. None of this shows up in the vendor's ROI calculator. All of it determines whether the system is still running six months from now.

What You Actually Need Before You Automate

Clean data paths. If your source data lives in three systems and two of them require someone to manually export a CSV every Friday, the automation will inherit that manual step. The model doesn't fix your data pipeline. It sits on top of it. If the pipeline is manual, the automation is half-manual. We've seen clients spend $15K on an AI contract and then realize they're still paying someone to prep the data the AI needs. Defined failure modes. The model will return garbage sometimes. Not often, but sometimes. You need a plan for what happens when it does. Does the system queue it for review? Does it fail silently and log the case? Does it retry with a different prompt? The vendor won't design this for you. They'll tell you the model is "highly accurate." Accurate is not the same as reliable. Reliable means you know what happens when it's wrong. Token cost visibility. Most Pacific Northwest clients we talk to don't track API token usage until the bill surprises them. A model that costs $0.002 per request sounds cheap until you're making 50,000 requests a month and half of them are processing redundant inputs because nobody set up deduplication. Track token spend from day one. Set a budget. Alert when you hit 80% of it. This is basic IT hygiene, but it's skipped constantly because the vendor sold the tool as "low-cost." Someone who can read logs. When the automation breaks, you need someone who can look at the logs and figure out why. The vendor's support team will ask you for logs. If your IT setup doesn't generate useful logs, or if nobody on your team knows where the logs live, you're stuck. Managed IT should own this. If your provider can't explain what failed and why, the automation is going to stay broken.

The Part We Keep Saying Out Loud

Most businesses in the Pacific Northwest don't need AI automation yet. They need their existing systems to work predictably. Automate the thing that's already working manually. Don't automate the thing that's broken manually and hope the AI fixes it. The AI will inherit the breakage, add latency, and cost you tokens.

If your current IT provider can't answer questions about API rate limits, log retention, or how your data moves between systems, they're not going to be able to support the automation you're thinking about buying. The gap between a working demo and a working system is filled with IT work. Someone has to do it.

If you're evaluating AI tools and you're not sure your infrastructure is ready: schedule a consultation with us at craftworkgrp.com. We'll review your setup, tell you what's ready and what's not, and save you from buying something your systems can't support yet.