What Pacific Northwest Businesses Actually Need Before They Automate Anything

A prospect called last month. SaaS company, 40 employees, Portland. They wanted to automate support ticket classification using an LLM. First question I asked: "How many tickets do you get per week?"

Twelve.

Twelve tickets. They wanted to spend engineering time building a model pipeline for twelve tickets a week that a human resolves in under five minutes each. The automation would have cost more to maintain than the problem cost to ignore.

This keeps happening. A company reads that AI is transforming operations, decides they need to automate something, and picks a process that doesn't hurt enough to justify the complexity. Or they pick one that hurts because the underlying workflow is broken, and the model just inherits the broken workflow at scale.

If you're in Seattle, Eugene, Boise—anywhere in the region—and you're thinking about automation, here's what actually matters before you spend a dollar on it.

Start With the Thing That Costs You Sleep

Automation makes sense when a process is: repetitive, high-volume, and currently handled by someone who could be doing something harder. If it's not all three, you're solving the wrong problem.

Good candidates for automation in a 50-person company:

• Categorizing 200+ support emails a day so they hit the right queue
• Pulling structured data from invoices sent by 40+ vendors in different formats
• Generating first-draft replies to common questions where the answer is already documented

Bad candidates:

• Anything you do fewer than 20 times a week
• Anything where the edge cases are more common than the base case
• Anything where "it depends" is the most frequent answer

One of our clients wanted to automate their onboarding checklist. Sounded reasonable until we looked at the checklist. It had 18 steps. Eleven of them required judgment calls based on the client's industry, size, and compliance needs. We automated three steps—the ones that were pure data entry—and left the rest alone. Took two days. Saved four hours a week. No one has to babysit a model that hallucinates onboarding tasks.

Your Data Probably Isn't Ready

The model is not the hard part. The hard part is getting clean, consistent input data and handling the cases where the model returns garbage.

Before you automate anything, answer these:

• Is the data structured, or is it 40 different formats depending on who submitted it?
• Do you have examples of what "correct" looks like, or are you guessing?
• When the model gets it wrong, does someone notice before it compounds?

We worked with a property management company that wanted to auto-generate lease summaries from signed PDFs. Great idea. Problem: the PDFs came from eight different jurisdictions with eight different lease templates, half of them scanned images, not text. We spent more time building the OCR + cleanup pipeline than we did on the summarization model. That's normal. The demo shows you the model. Production is everything before and after the model.

Rate Limits and Failures Happen at the Worst Time

You will hit a rate limit. You will get a 500 error from the API. You will get a response that's an empty string or a JSON object missing the field you expected.

If your automation doesn't have a fallback—something that routes the task to a human queue when the model fails—you've built a system that breaks silently and no one notices until a client emails asking why their request has been sitting for three days.

Basic things every automation pipeline needs:

• Input truncation so a 60-page PDF doesn't blow your token budget
• Retry logic with exponential backoff so a temporary API blip doesn't kill the job
• Confidence thresholds so low-confidence outputs go to human review
• Logging that actually tells you what failed and when

None of this is exciting. All of it is the difference between a system that runs and a system that runs until it doesn't.

What This Looks Like in Practice

A client in Tacoma automates vendor invoice processing. The model extracts vendor name, invoice number, amount, due date. If all four fields parse cleanly and the amount is under $5,000, it goes straight to approval queue. If any field is missing or the amount is over $5,000, human reviews it. If the model confidence score is below 0.85 on any field, human reviews it.

They process 150 invoices a week. The model handles about 90 of them end-to-end. The other 60 get flagged. That's a 60% automation rate, which sounds unimpressive until you realize it saves their AP person 8 hours a week and the error rate is near zero because a human still touches every edge case.

That's what good automation looks like. Boring. Reliable. Slightly faster than before.

Here's What You Should Do Next

Pick one process that's actually costing you time. Not the flashy one. The repetitive one. Map out the current workflow—inputs, outputs, edge cases. If the edge cases are more than 20% of total volume, you're not ready to automate it yet. If they're less than 20%, call us. We'll tell you if a model is the right tool or if you just need a better form.

Visit craftwork