Sample use case · Back office
Agents inside your workflows
AI that drafts the order entry, the follow-up, the exception report — inside the tools you already use.
Inbox
From purchasing@northlake-ind.ca
PO #4471 — 40× AC compressor kits
Draft ERP entry
96% confident- Customer
- Northlake Industries ✓
- SKU
- HV-2740 · AC Compressor Kit ✓
- Qty
- 40 ✓
- Unit price
- CA$899 · from the PO ✓
- Requested date
- “ASAP” → flagged for review
Human sign-off
Nothing posts itself — a person approves every entry.
Emailed PO arrives
any format, any attachment
Intake agent parses & validates
LLM with tool calls — SKU lookup, price check
Draft entry queued
reasoning attached
Human approves
one click, full trail
ERP posted, reply drafted
in the company's own tone
Ambiguous? → exception board
agents never guess — humans see only the hard cases
Agents are LLMs with hands — tool calls into the ERP, the inbox and the document store — but no autonomy over outcomes: every action is drafted, reviewed and approved by a person until the approval rate earns wider scope.
The problem
An industrial supplier's back office retyped emailed purchase orders into the ERP, chased missing details by hand, and wrote the same status updates all day. Growth meant hiring more people to do more retyping.
What we build
- An intake agent that reads emailed POs and attachments, drafts the ERP entry, and flags anything ambiguous instead of guessing.
- Follow-up agents that chase missing information and confirm order status in the company's own tone — every message reviewed before it sends.
- An exception board where humans handle only the cases the agents couldn't — with the full trail of what was done and why.
Proven before you commit
- Started with one inbox and one order type; a person approved every agent action for the first month.
- Widened scope only as the approval rate proved the agents out — measured weekly, visible to the whole team.
The outcome
80%
of orders entered without retyping
-11 hrs
per person per week on routine follow-ups
100%
of agent actions human-reviewed at start
A representative engagement — numbers illustrate the shape of the result, not a named client reference.