---
title: "Where to Start: The Process Diagnosis That Defines Your First AI Agent"
description: "A process diagnosis for AI shows you where to start: map, score by volume, risk and value, then pick your first agent with clear metrics and human gates."
url: https://addabra.com/en/blog/process-diagnosis-for-ai
lang: en
date: 2026-10-29
category: "Strategy"
keyword: "process diagnosis for AI"
publisher: Addabra
alternate: https://addabra.com/blog/diagnostico-de-procesos-ia
---

# Where to Start: The Process Diagnosis That Defines Your First AI Agent

> A process diagnosis for AI shows you where to start: map, score by volume, risk and value, then pick your first agent with clear metrics and human gates.

Friday, 7:15 p.m. A COO has a list open with thirty AI ideas sent in by her managers: a sales assistant, another for procurement, an HR bot, meeting summaries. They all sound good and none of them come with numbers. Without a **process diagnosis for AI**, the decision will go to whoever pushed hardest, not to the idea that creates the most value.

Picking the wrong first agent is expensive: not just the project budget, but the team's trust in everything that comes after.

## Why starting with the tool is a mistake

Most companies start by asking, "What can we do with AI?" That's the wrong question. The right one is, "Where are we leaking time and margin?" AI is the means; the process is the starting point.

What's more, a useful agent doesn't live in isolation. It needs business memory, clear rules, a human approval point and a way to measure whether it's doing the job well. If you pick a process without thinking about those pieces, you end up with a cute experiment nobody uses.

## Step 1: map processes as they actually happen

Bring together the people who do the work and sketch each process end to end: what triggers it, what steps it has, which systems it touches, who decides and where it waits. Don't use the manual; use what happens on an ordinary Tuesday.

Three things usually show up on this map:

- **Invisible steps**: copying data from one system to another, hunting for a document, asking a coworker.
- **Wait times**: the process stalls because someone hasn't responded or doesn't have the information.
- **Knowledge in people's heads**: decisions only one person knows how to make. This is what we call [tribal knowledge](https://addabra.com/en/blog/tribal-knowledge), and it's one of the biggest risks any company faces.

## Step 2: find where time and margin are leaking

For each process, estimate four basic data points: how many times it happens per month, how many minutes of staff time it consumes, how often it goes wrong and what happens when it runs late. You don't need perfect precision; you need honest orders of magnitude.

Typical leaks are in processes that combine high volume with manual work: responding to requests, preparing documents, reconciling information, following up with customers or suppliers.

## Step 3: score each process

With the map and the data, score each process from 1 to 5 on five criteria. That turns a debate of opinions into a conversation with numbers.

| Criterion | What to ask | Scores high if... |
|---|---|---|
| Volume | How many times a month does it happen? | It happens hundreds or thousands of times |
| Repetition | Does it follow clear rules or patterns? | Most cases look alike |
| Risk | What happens if it goes wrong? | An error can be fixed before it reaches the customer |
| Value | How much time, money or speed does it free up? | It frees hours from expensive roles or speeds up revenue |
| Data availability | Is the information in accessible systems? | It's digital and searchable |

Note that risk is scored in reverse: a process where an error is serious and hard to detect scores low as a starting point. That doesn't mean it can never be automated; it means it isn't a good place to begin.

## Step 4: choose your first agent

The ideal first agent usually sits at the intersection: high volume, high repetition, manageable risk and value leadership can see. Some common candidates:

- Preparing sales proposals from a template and the customer profile.
- Triaging and answering support requests that follow known patterns.
- Reconciling payments and preparing collection reminders.
- Gathering and verifying documentation for a case file.

Pick one. Not three. One agent in production teaches you more than ten in a slide deck.

## Step 5: define the human gates

A gate is a point in the workflow where a person in a specific role reviews and approves before things move forward. Define them from the design stage, not as a patch:

1. **What gets approved**: the final draft, a decision, an amount.
2. **Who approves it**: the role, not a specific person.
3. **When it can be relaxed**: after a certain number of error-free cases, some gates can shift to sample-based review.

The details of how to design them are in [human approval by role](https://addabra.com/en/blog/human-approval-by-role).

## Step 6: define success metrics

Before you build, write down how you'll know it worked. Three to five metrics, each with today's baseline:

- Time per task.
- Volume handled per person.
- Error or rework rate.
- Customer response time.
- Percentage of cases the agent resolves without escalating.

With those metrics, you can estimate the return before you invest, using the method in [how to calculate the ROI of an agentic system](https://addabra.com/en/blog/agentic-system-roi).

## Illustrative example: an industrial maintenance company

Imagine a company that services industrial equipment. Its initial list included an assistant for the field technicians, an HR bot and a report generator.

After mapping, something shows up that nobody had put on the list: every service order requires a coordinator to pull together the equipment history, review the customer's contract, prepare the quote and send it. It happens hundreds of times a month, follows clear rules and errors are caught before the quote goes out.

The process scores high on volume, repetition and value, and its risk is manageable. It's chosen as the first agent: the agent prepares the quote using the history and the contract, and the coordinator approves it. The main metric is the time between the request and the quote being sent. The technician assistant, flashier but riskier, is saved for phase two.

> **The lesson** The best first agent is almost never the flashiest idea on the list. It's the process that repeats the most and that everyone takes for granted.

## Common mistakes in the diagnosis

- **Doing it without the people who run the process.** The map comes out pretty and wrong.
- **Choosing on enthusiasm.** The most talked-about idea isn't necessarily the most valuable.
- **Starting with the most critical process.** The first agent should build trust, not bet the operation.
- **Not defining metrics up front.** Without a baseline, any result is debatable.
- **Forgetting the data.** If the information isn't in accessible systems, the agent won't be able to work.

## What to do on Monday

Pick one department and schedule two mapping sessions with the people who do the work. Fill in the scoring table with rough data and put the top three processes in front of leadership. Within a week you'll have a decision based on numbers, not on a wish list.

At Addabra, the diagnosis is the first step of our method: diagnosis, design, build, production and evolution. We do it with your team and finish with your first agent selected, its human gates defined and its success metrics in place.

## Frequently asked questions

### How long does a process diagnosis for AI take?

For one or two departments, a few weeks is enough if you work with the people who run the process and with real volume and time data. What matters isn't the duration but ending with a first agent selected and clear success metrics.

### Why not start with the company's most important process?

Because the most important process is usually also the one with the most risk and the most exceptions. It pays to start with one that has high volume, clear rules and manageable risk, prove results and then move toward critical processes.

### Who should take part in the diagnosis?

The business owner of the process, two or three people who run it every day and someone from leadership who can set priorities. Without the people who do the work, the map reflects what the manual says, not what actually happens.

## About Addabra

Addabra is an AI-software-first software factory that designs and builds custom agentic systems for companies: one agent per process, a shared business memory, orchestration, human approval by role, auditing and cost control. Free diagnosis: https://addabra.com/en/blog/process-diagnosis-for-ai#contacto