The Agent Layer
Your systems stay. A layer of agents does the work between them.
The Agent Layer is how we build: agents with their own limited logins, your rules written down and enforced, a person in the loop wherever you want one, and a record of everything they do.
Where the agents sit
Your team
Decides, approves the exceptions, handles the relationships. The judgement stays with your people.
The Agent Layer
Reads documents and messages, applies your rules, acts in your systems, asks when it is unsure, and logs everything it does.
Your systems
Stay the record. ERP, accounting, commerce, your CRM, EDI, spreadsheets — nothing is replaced.
How an agent earns its keep
Watch
The agent sees the real work: inboxes, documents, the orders and bills your team handles. It does not act.
Draft
It drafts what it would do. Your team keeps working normally and compares.
Approve
Your team approves, edits or rejects each draft. Every correction teaches it your rules.
Act with limits
On the kinds of work it gets right, you let it act. Anything above your limits still comes to a person.
Report
You see what it did, what it escalated and why — every day.
Autonomy is granted per type of work, never all at once. You can move any agent back a stage at any time.
What you control
| Set by you, changed any time | |
|---|---|
| Approval limits | Dollar amounts, new customers, price changes. Above your line, it goes to a person. |
| System access | Which systems each agent can touch, and whether it can write or only read. |
| Working hours and send limits | When an agent runs, and how much it is allowed to send out. |
| Who gets the escalations | Which person sees which exception, and where it lands. |
| Switching an agent off | One switch per agent. The work goes back to your team as it was. |
The second agent is faster than the first
Every agent shares the same connections to your systems, the same map of your customers, items and part numbers, and the same rules.
Build Order Entry first and Customer Service already knows your orders. Build AR next and it already knows what shipped and what was disputed.
- Order Entry Agent
- Customer Service Agent
- AR & AP Agent
One shared base
- Connections
- Customer & item map
- Your rules
- Action log
Agents vs. what you’ve tried
| Rule-based automation (RPA, Zapier) | Chatbots & copilots | Agents on The Agent Layer | |
|---|---|---|---|
| Handles messy inputs (PDFs, free text) | Breaks when the format changes | Reads them, but can't act in your ERP | Reads them and acts, inside your rules |
| Takes action in your systems | Yes, on fixed paths | Rarely | Yes, with its own scoped login |
| Knows your customers and rules | Only what's hard-coded | Only what's in the prompt | Learns and keeps them; corrections stick |
| Asks a person when unsure | No — it fails or does the wrong thing | Depends on the user noticing | Built in, with limits you set |
| Record of what it did | Partial | Chat history | Every action with its reason |
What you are actually agreeing to
Least-privilege access
Each agent gets its own scoped login with only the permissions its job needs. No shared admin accounts.
Every action logged, your limits enforced
Each action is recorded with what the agent saw and why it acted. You set the approval thresholds — dollar amounts, new customers, price changes.
Your data stays yours
We never use your data to train models. NDAs are standard on every engagement.
How it's built (for your IT lead)
Scoped service users, one per agent
Each agent authenticates as its own service user with least-privilege roles in your systems. No shared admin account, and access is revoked in one place.
An action log you can audit
Every action stores the input, the decision, the rule applied, the approver and the timestamp. The log is exportable.
Kill switch and dry runs
A kill switch per agent, and a dry-run mode for any new write action so you can see the effect before it happens for real.
Built for your process, not ours
The scaffolding underneath is shared — the connections, the customer and item map, the log. The agent itself is written around how your team actually works: your approval points, your part numbers, your exceptions. It is not an out-of-the-box agent you have to bend your process around.
Questions & Answers
FAQs
Bring one process and we'll tell you straight
A working session on your real documents. You leave knowing whether an agent can do the job, and what counts as done.