Automation
AI automation for documents, triage, and the steps between systems
AI automation applies a model to a step you can check: classifying a request, reading a document into fields, or routing work. The surrounding workflow is ordinary software, with a queue for the cases the model should not finish alone.
- Document intake
- Triage
- Draft-and-wait
- Monitoring
Automate the step, not the department
Good candidates are high volume and easy to sample: inbound emails, supplier invoices, application forms, listing descriptions, support tags. Poor candidates are rare, high-stakes decisions with no historical examples. We measure accuracy on a held-out sample before anyone turns off the old process.
Straight-through only when the error is cheap
If a wrong field creates a wrong payment, a person confirms it. If a wrong tag only mis-files a newsletter, straight-through processing may be fine. That threshold is a business decision we write down, not a default of “fully automated”.
A routine step with a review
Automated
- Intake of one document type
- A suggested class or route
- A queue when confidence is low
Still reviewed
- Payments and legal commitments
- Exceptions you have not listed
- A silent write into finance
What changes an automation
The trigger
An inbox, an upload, and a status change are different starts.
The write
Suggesting a field is smaller than saving it. Say which one this release does.
The exception
The interesting work is the case the rule does not cover. Give it an owner.
What an automation brief needs
- 01
The trigger
An inbox, an upload, or a status change.
- 02
The document
One type, with two real examples.
- 03
The write
Suggest a field, or save it. They are different.
- 04
The exception
Who owns the case the rule does not cover.
Marks an automation can be trusted
- 01
The trigger is real
It starts from an event in a system, not from a person remembering to run it.
- 02
A failure shows
A step that did not complete is visible, with the record it touched.
- 03
A person can pause
Someone can stop the flow without taking the system down.
- 04
The audit is there
Who, when, and which record are written for each run.
What we build
Document intake
Fields extracted from PDFs and scans, with a confidence score and a correction screen.
Triage
Routing to the right queue, using the text and the customer record.
Draft-and-wait
A prepared update that a person sends, for the actions you will not fully automate.
Monitoring
A view of volume, corrections, and the categories that keep failing.
How an engagement runs
- 01
Sample the pile
A few hundred real items, including the ugly ones.
- 02
Shadow the current team
The model suggests. People still act. You compare.
- 03
Automate the confident slice
The rest stay in the queue, and the threshold is adjustable.
Related reading
Questions we hear
Is this RPA?
Screen-clicking robots are a last resort. We prefer APIs and documents. Where a legacy system has no interface, a careful automation may still be justified, and we will say what it will cost to keep alive.
Can you automate sales follow-up?
Drafting and reminders, yes. Sending unsupervised commercial claims, no. See AI agents for the approval pattern.
How does this relate to business automation software?
The solution page covers workflow products. This page covers the AI step inside them. Many projects need both.