👋 Hey, this is Artem and this is the first issue of The Blueprint. At WIQ, we spend each day forward deployed at Fortune 500 companies working through the hairy journey of AI transformation.
This newsletter is our way of sharing the lessons we learn with our friends on the same journey.
This week we are releasing an experimental app to record a task, then output a skill for Claude 🧪
At Bardeen (my last startup), we automated 30 million tasks with AI (we started before ChatGPT by the way). It taught us many things, but the one that spawned WIQ (my current company) was this:
You need to understand the work deeply to automate it (duh!).
Some workflows look the same every time - how you answer a simple support ticket, reviewing an NDA against your standard terms, and scoring a lead with their LinkedIn activity. It’s why you see vertical agents dominate these use cases (think Sierra, Legora, Clay).
Operations workflows are different. Take paying an invoice. There’s a high degree of repetition, but each run has judgement and 70% of cases are edge cases (which contradicts the definition of "edge case" 😅, but hey, that's the point, enterprises are complex!).
So it is never a straight line, it's more like a tree, or a forest, or really a dark thick jungle. To automate it correctly requires knowing all directions, understanding every conceivable path, and knowing how to escalate to Nancy in Accounting if all else fails.
Why capable agents fail at ops work
Last year a payments team we worked with handed one of its daily jobs to an AI agent: clearing the exceptions when a payment fails or gets returned.
They gave it access to the payment system, accounting system of record, and wrote the skill based on an SOP the team was supposed to follow. The agent still cleared the wrong ones.
Not knowing how the work is actually done is the #1 time suck and road block in every enterprise AI deployment today. And the problem is, the answer is not written anywhere. Often it's in the head of literally 1 person.
I've seen this more times than I'm willing to admit, where we walk into a review, a manager points at a mistake ("we don't do it like that"), we investigate ("let's ask William from Procurement"), and lo and behold the data is right, the manager is bewildered, and William saves the day (again). And this is not a scrappy startup, this is inside a $1B P&L business unit at a global multinational.
Layer on top of that the need to adapt and change the work all the time (tariffs and Straight of Hormuz ring a bell?), adjust policies and update tools, and we end up in a situation where continuous and automated learning by actual data capturing becomes a must-have.
Blueprints are the missing piece of AI for enterprise
We call it a blueprint: the exact way one job is done at your company, captured from watching the work rather than copied from a policy doc. Automating this used to mean a brittle script that ran the happy path and broke the moment reality diverged.
What makes ops hard is everything that script left out: the exceptions, the judgment, and the dozens of unwritten rules an operator applies without thinking. A blueprint is built to hold all of it.
Take something ordinary but unforgiving: creating a customer invoice in NetSuite, Oracle’s ERP. Every screen carries dozens of fields, and which ones matter depends on how your company set it up.
Filling out a customer invoice in NetSuite.
We built the one below from a public NetSuite tutorial on YouTube. It could as easily be your own SOP, an internal Loom, or a screen recording. Whatever the source, our goal is to create these components of a blueprint:
Two of them are commonly seen in any LLM skill:
The process: the record it creates and its real steps in order, pulling up the customer, setting the posting period, adding each line item by its code and location, calculating shipping, saving.
The connectors: the exact tool each step runs through, here NetSuite, reached through a connector or driven in the browser.
We add 2 new elements enterprise automation needs to work:
The escalation paths: where the agent stops and hands to a person, a customer it can’t find, a missing item code, an invoice that won’t save.
The guardrails: the company’s own rules, never save an invoice with no customer, always set a posting period, never touch a project already billed.
It seems simple, but doing this at scale, continuously, across any agent platform, is anything but.
Go and see
When writing this first issue, I looked back to Digital Transformations of the early 2000s for a good story. Instead I found something older.
Sixty years ago, Taiichi Ohno, the engineer who built the Toyota Production System, trained his people with a piece of chalk. He'd draw a circle on the factory floor, tell a young engineer to stand inside it, and watch. He'd come back hours later, and if they hadn't seen enough, he told them to keep watching.
The lesson: you cannot understand how a job is really done from a desk, a report, or an SOP. You have to go and see. Toyota still calls it "go and see."
So, in an effort to help you do this (with AI obviously), here’s an app that gives you a taste of how WIQ does this at scale.
Create a blueprint for yourself
We’re making a key part of the WIQ discovery engine available for experimental preview. You can record a blueprint here or book some time for us to give you a demo of this feature and the broader platform.
