E82 Studios
Design · Build · Evolve
Skip to content
AI3 min readWritten by the studio

Start with the bottleneck, not the model

Technical architectural diagram demonstrating AI workflow integration and business process automation.

The useful question is never "where can we add AI?" It is "where does work pile up, and why?"

Most AI projects that stall were scoped from the technology inwards. A tool looked impressive, a pilot was funded, and a use case was found to justify it. The work gets built, it demos well, and then it sits beside the process it was supposed to replace because the process never actually changed.

Projects that survive contact with a business start from the opposite end. Find the queue: the inbox nobody gets to before lunch, the enquiries that go cold overnight, the report someone rebuilds by hand every Monday, the calls that ring out after five. Those queues are visible without any analysis, and everyone in the company can name them.

Map the work before automating it

Before anything is built, we write out the current path of a single unit of work (one enquiry, one order, one support ticket) from arrival to resolution. Who touches it, what they decide, what they have to look up, and where it waits. It is usually a short document, and it is usually uncomfortable reading, because most delay turns out to be waiting rather than working.

Automating a step that was never the bottleneck just moves the queue somewhere less visible.

Enterprise AI automation pipeline architecture diagram highlighting interconnected operational nodes.
Trigger, understanding, action, outcome: every automation worth building can be drawn as one line.

Judgement versus retrieval

Once the path is on paper, sort each step into one of three buckets. It clarifies almost everything about what to build.

  • Deterministic: rules a computer has always been able to follow. Automate with ordinary software with no model needed, and no model wanted.
  • Language-shaped: reading, summarising, classifying, drafting, extracting. This is where current models genuinely earn their place.
  • Judgement: pricing, exceptions, relationships, anything with consequences. Keep a person here, and use the model to prepare their decision rather than make it.

Teams that skip this sorting tend to point a model at the whole process and hope. Teams that do it end up with something narrower, duller and far more reliable: a system that reads the incoming message, pulls the three facts a human would have looked up, drafts the reply, and hands it over when the answer is obvious, while routing anything unusual to a person with the context already assembled.

Design for the wrong answer

Any system that reads language will sometimes be confidently wrong. That is not a reason to avoid it; it is a specification. Before launch we decide what happens on a bad output, and the answer is never "the user will notice".

  1. 01

    Give it a narrow job

    A small, well-defined task fails in ways you can predict and contain.

  2. 02

    Show its working

    Surface the source, the record, the reasoning: whatever a reviewer needs to check in seconds.

  3. 03

    Put a human at the consequential step

    Approval before money moves, before a commitment is made, before anything leaves the building.

  4. 04

    Log everything and read the logs

    The failures you never see are the ones that quietly train people to distrust the system.

The integration is the project

The model is rarely the hard part. The hard part is the plumbing: getting clean access to the systems where the work actually lives, handling the records that contradict each other, deciding what happens when an API is down, and fitting the new path into how the team already works so nobody has to remember an extra habit.

This is why we treat AI work as engineering with a model inside it rather than a category of its own. The interesting questions (what triggers this, what does it have access to, who checks it, what does it write back, how do we know it worked) are the same questions any good system has always had to answer.

Start with the bottleneck. Build the smallest thing that clears it. The technology stops being the point about a week after it goes live, which is exactly how it should feel.

Working on something this applies to? We're happy to look at it with you.

Have something worth building?

Tell us what you're trying to launch, improve or automate.

We reply within 24 hours.