AI workflows in production
An AI workflow is not a chatbot. It is a piece of your operation that now runs itself, with a person where judgment matters, and someone accountable when it fails.
What is a production AI workflow?
A process that used to run on people and spreadsheets, now run by an agent with access to your systems: it is triggered by a real event, reads the business context, decides, asks a person when the stakes call for it, acts through your tools, and leaves a trace you can evaluate. Production means it runs every day, on real volume, with known failure modes. A demo does none of that.
How we work: what a forward deployed engineer (FDE) is and why it matters →
The anatomy of a workflow
- 01
Trigger
An email arrives, an order changes, a file lands, a threshold is crossed.
- 02
Business context
The rules, the history and the data the decision needs, pulled from where they live.
- 03
Agent
A model with a narrow job and the right tools, not a general chat.
- 04
Tools, APIs, data
Your ERP, CRM, inbox, documents and providers, connected with the least permissions that work.
- 05
Decision
Structured output you can check, not free text.
- 06
Human approval
When the cost of a mistake is high, a person confirms. Everything else flows.
- 07
Action
The record is updated, the reply is sent, the booking is changed.
- 08
Evaluation and observability
Every run is logged and scored, so you know it works and see where it fails.
Four patterns we deploy
Operations
- incoming email
- classify
- extract
- validate
- ERP
- exceptions to a person
The inbox that three people triage becomes a queue that triages itself, with the odd cases routed to the one who knows.
Engineering
- issue
- context
- implementation
- tests
- pull request
- review
Agents write, test and review inside your repository and your rules; engineers decide what ships.
Finance
- invoice
- extract
- validate
- reconcile
- approval
- posting
Documents become structured, matched and approved, with a person on the exceptions and the amounts that matter.
Travel operations
- flight change
- rules
- alternatives
- customer
- booking action
GDS events, fare rules and customer contact in one flow, from the notification to the rebooked ticket.
Production means
Evaluation
We measure whether the workflow actually does the job, on your data, before and after every change.
Observability
You can see what the agent did, why, and where it failed. No black box in the middle of your operation.
Human control
High-impact decisions stay with people. The workflow knows which ones those are.
Security
Minimum necessary permissions, your infrastructure, your accounts. Data stays where it is.
Failure handling
Agents fail. A production system knows what happens next: retry, fall back, or hand over.
Cost
The model follows the economics of the task, not the hype. Small where small works.
Common questions
How is an AI workflow different from a chatbot?
A chatbot waits for a question. A workflow is triggered by your operation, acts on your systems and is measured on the outcome. One is an interface; the other is work getting done.
Do we need to change our tools?
No. We connect to what you have: ERP, CRM, inbox, documents, providers. If a tool is the bottleneck we say so, but we do not start by replacing anything.
Which model do you use?
The one that fits the task and the budget. Most workflows run on a small model for the routine steps and a larger one only where judgment is needed. We have no platform to sell, so the choice is free.
What if the agent gets it wrong?
It will, sometimes. That is why the workflow has checkpoints, evaluation and a defined fallback. The goal is not zero errors, it is knowing when they happen and what happens next.
How long until the first workflow runs?
Weeks, not months. A few days to map the process and connect the context, then iterations in production with your team.
Which process eats your week?
Tell us. We will say whether it is a workflow for AI, and how we would take it to production.