OXARUS :: OXARUS AGENT AI

Agentic hyperautomation · Oxarus Agent AI

AI agents that run multi-step processes across email, ERP, CRM, e-commerce and support channels, with approval steps, permissions and an audit trail.

page

/en/automation

unit

Agent AI

path

Oxarus home › Oxarus Agent AI

markdown

/en/automation.md

Interface data marked as representative is illustrative; it is not a real client, result or measurement.

AGENTS THAT RUN PROCESSES END TO END.

OXARUS AGENT AI

Our AI agents read the email, find the record in your ERP, catch the mismatch, ask for approval where it's needed and close out the work. You step in only where a decision is required.

  • Runs on top of the tools you already use
  • Human approval at critical steps
  • Every action in the audit trail
  1. Incoming email
  2. Document and ERP
  3. Check
  4. Human approval
  5. Record and notify

REPRESENTATIVE INTERFACE DATA · Representative agent console: a supplier invoice processed from email to ERP entry.

run

run_4821 · supplier invoice

tag

Representative

running

running

waiting

awaiting approval

done

completed

  1. Read the email

    detail

    Found a PDF invoice attached to the message from Northfield Supply.

  2. Parsed the invoice

    detail

    Extracted line items, quantities, tax and the total amount.

  3. Matched it to the PO

    detail

    Pulled PO-7710 from the ERP and compared the line items one by one.

  4. Flagged a price variance

    detail

    The unit price on NS-4410 is higher than on the PO; the variance breaks the tolerance rule.

  5. Requested approval

    detail

    Sent the finance lead an approval request with a summary of the variance, then waited.

  6. Posted the entry

    detail

    After approval, posted entry AP-5518 to the ERP.

  7. Notified the team

    detail

    Posted a summary and a link to the audit log in #ap-ops.

call

tool call

result

result

reason

Reasoning

pending

pending

back

Back to the live step

hint

Click a completed step to see its call and result.

Incoming invoice

supplier

Supplier

po

PO

lines

Item

qty

Qty

invoicePrice

Invoice

poPrice

PO

mismatch

NS-4410 unit price is {percent} above the PO · tolerance {tolerance}

Approval step

approver

Clara Wendt

role

Finance lead

summary

Variance {total}: {qty} units × {diff}

idle

Queued

waiting

Awaiting decision

approved

Approved

elapsed

elapsed

tools

tool calls

human

incl. human approval

WHEN RULES BREAK, WORK GOES ON.

FROM AUTOMATION TO HYPERAUTOMATION

Classic automation follows pre-written rules; when something unexpected happens, it stops and drops the work into a queue. An agent reads the situation, gathers context and decides what to do within set boundaries.

SAME EXCEPTION

The supplier sent the invoice as a scanned PDF. The invoice has no PO number; it only appears in the body of the email. And one line item's price differs from the PO.

REPRESENTATIVE INTERFACE DATA · Representative

Incoming email

from

Northfield Supply · Accounts

subject

Invoice INV-20931

body

Hello, the attached invoice belongs to your order {po}. Kind regards.

attachment

INV-20931.pdf · scanned

REPRESENTATIVE INTERFACE DATA · Representative

Scanned invoice

po

PO number

missing

none

textLayer

No text layer

REPRESENTATIVE INTERFACE DATA · Representative

play

Play

pause

Pause

replay

Restart

step

Step {n}

REPRESENTATIVE INTERFACE DATA · Representative

running

running

stopped

stopped

queued

waiting in queue

waiting

awaiting approval

done

completed

  • Rule-based flow (RPA)

    1. Attachment downloaded

      PDF saved to a folder.

      ok

    2. PO number field unreadable

      The template field is empty; the scanned document has no text layer.

      fail

    3. Matching skipped

      The rule can't continue without a PO number.

      fail

    4. Dropped into the error queue

      Waiting until someone notices.

      warn

    Result: the flow stopped, and the work has to be done by hand.

  • Agent (Agent AI)

    1. Read the document

      Extracted line items and amounts from the scanned PDF.

      ok

    2. Found the missing data

      Pulled the PO number from the email body: PO-7710.

      ok

    3. Spotted the variance

      NS-4410 has a price variance outside tolerance.

      warn

    4. Sent it for approval

      Summarized the variance, checked with the authorized approver and posted the entry once approved.

      ok

    Result: work closed, the decision stayed with a human, every step logged.

  • Ambiguous input

    RULE-BASED

    Stops on documents outside the template

    AGENT

    Reads the document and fills gaps from context

  • Exceptions

    RULE-BASED

    Every exception needs a new rule

    AGENT

    Decides within policy, and asks when outside it

  • Cross-tool work

    RULE-BASED

    One trigger, one action

    AGENT

    Works step by step across multiple systems

  • Change

    RULE-BASED

    Breaks when the screen changes

    AGENT

    Works through APIs and authorized tools

  • Traceability

    RULE-BASED

    Often just an error log

    AGENT

    Every decision logged with its reasoning

PICK A PROCESS, WATCH THE AGENT.

PROCESSES

The flows below are examples we build often. In each one the agent moves step by step across the systems your company uses, stops at the step that needs approval and waits for your decision.

human tag

human approval

REPRESENTATIVE INTERFACE DATA · Representative

calls

calls: {count}

REPRESENTATIVE INTERFACE DATA · Representative

skipped

skipped

REPRESENTATIVE INTERFACE DATA · Representative

summary

tool calls: {calls} · human decisions: {decisions} · all in the audit log

REPRESENTATIVE INTERFACE DATA · Representative

running

running

waiting

awaiting approval

done

completed

rejected

rejected

REPRESENTATIVE INTERFACE DATA · Representative

Your approval is needed

hint

The agent stopped at this step. It continues based on your decision.

approve

Approve

reject

Reject

approved

Approved by {role}

rejected

Rejected by {role}

  • Invoice matching

    approver

    Finance lead

    A supplier invoice arrived in the accounting inbox.

    Invoice matched to the PO, entry posted, team notified.

    No entry posted. The agent passed the variance to purchasing to raise with the supplier.

    1. 0

      Recognized the invoice attachment and the sending supplier.

    2. 1

      Extracted line items, tax and totals from the PDF.

    3. 2

      Ran a three-way match against the PO and delivery note.

    4. 3

      Sent the out-of-tolerance price variance for approval.

    5. 4

      Posted the approved invoice as an accounting entry.

    6. 5

      Posted a summary and the entry link in the finance channel.

  • Orders and returns

    approver

    Customer service manager

    A customer opened a return request in the store.

    Return eligibility checked, warehouse and customer informed.

    No refund issued. The request went to customer service with the reasoning attached.

    1. 0

      Read the order, delivery date and return reason.

    2. 1

      Checked eligibility against the return policy.

    3. 2

      Created a return receipt and shipping code in the warehouse.

    4. 3

      Sent the high-value return for refund approval.

    5. 4

      After approval, posted the credit note and stock movement.

    6. 5

      Sent the customer the shipping code and next steps on WhatsApp.

  • Support triage

    approver

    Support team lead

    A new request arrived in the support channel.

    Request classified, simple ones answered, critical ones with the right person.

    No assignment made. The request waits in the team lead's queue with its summary.

    1. 0

      Read the request; identified topic, language and urgency.

    2. 1

      Pulled the customer record and past requests.

    3. 2

      Checked the status of the related order.

    4. 3

      Sent the ready reply to the customer asking about order status.

    5. 4

      Asked for approval to route the complaint to the priority queue.

    6. 5

      Added a conversation summary to the customer record.

  • Lead qualification

    approver

    Sales rep

    A quote request form came in from the website.

    Lead scored, ready in the CRM, meetings offered to qualified leads.

    No reply sent. The draft went back to the rep with a note to revise.

    1. 0

      Read the form answers and company details.

    2. 1

      Searched the CRM for an existing record and past conversations.

    3. 2

      Scored the lead against the agreed criteria.

    4. 3

      Submitted a personalized first reply draft for approval.

    5. 4

      Offered open slots from the sales rep's calendar.

    6. 5

      Logged the opportunity at the right stage with its owner.

  • Hiring and onboarding

    approver

    IT manager

    A new hire's contract was signed.

    Accounts ready, equipment request open, first-week plan shared.

    Admin access not granted. The request went back to the IT manager with its reasoning.

    1. 0

      Pulled the employee record, department and start date.

    2. 1

      Prepared the role-based list of accounts and permissions.

    3. 2

      Submitted access that needs admin rights for approval.

    4. 3

      Opened the approved accounts and raised the equipment request.

    5. 4

      Emailed the new hire a welcome note and first-week plan.

    6. 5

      Announced the new hire and start date to the team.

THE AGENT WORKS, YOU STAY IN CONTROL.

CONTROL AND SECURITY

Every agent runs with defined permissions, limits and approval steps. What it did, why it did it and who approved it are all on record, and a single click stops it when needed.

REPRESENTATIVE INTERFACE DATA · Representative

panelTitle

Agent control panel · Northfield Supply

REPRESENTATIVE INTERFACE DATA · Representative

hint

Change a setting: the verdicts on pending actions and the audit log update right away.

REPRESENTATIVE INTERFACE DATA · Representative

Approval steps

  • Out-of-tolerance invoice

    detail

    Price variance above {tolerance}

    approver

    Finance lead

  • Amount above limit

    detail

    Payments and refunds over the per-transaction limit

    approver

    Manager

  • First customer reply

    detail

    First message sent with a new template

    approver

    Support lead

REPRESENTATIVE INTERFACE DATA · Representative

Role permissions

  • read
  • write
  • pay

  • 0

    role

    invoice-agent

    [✓] [✓] [✗]

  • 1

    role

    support-agent

    [✓] [✓] [✓]

  • 2

    role

    sales-agent

    [✓] [✗] [✗]

hint

Click a box: the permission changes right away and is logged.

REPRESENTATIVE INTERFACE DATA · Representative

Action limits

  • Per-transaction amount

    currency

    EUR

  • Actions per hour (actions)

  • Retries on the same job (times)

usage

{used} actions in the last hour

REPRESENTATIVE INTERFACE DATA · Representative

Pending actions

  • 0

    action

    Post invoice INV-20944 to the ERP

    detail

    Baltic Freight · price variance {variance}

  • 1

    action

    Refund return R-1182

    detail

    Order SO-10482

  • 2

    action

    Refund return R-1190

    detail

    Bulk order SO-10477

  • 3

    action

    Update the opportunity stage in the CRM

    detail

    Kestrel Labs · quote sent

  • 4

    action

    Send the first reply to the customer

    detail

    Request T-3310 · new template

  • 5

    action

    Retry entry AP-5521

    detail

    ERP timeout · attempt {retry}

auto

Automatic

approval

Goes to approval · {approver}

permission

Blocked · no permission

limit

Blocked · amount limit

retry

Blocked · retry limit

throttled

Queued · hourly limit

paused

Stopped

auto

{count} automatic

approval

{count} in approval

blocked

{count} blocked

REPRESENTATIVE INTERFACE DATA · Representative

Audit log

[All] [Agent] [Human]

empty

No entries for this filter.

  • 0

    actor

    invoice-agent

    action

    Entry AP-5518 posted to the ERP · approved by Clara Wendt

  • 1

    actor

    clara.wendt

    action

    INV-20931 price variance approved

  • 2

    actor

    support-agent

    action

    Request #3304 routed to the priority queue · reason: complaint

  • 3

    actor

    sales-agent

    action

    No write permission; CRM update suggested to the rep

gateOn

Approval step turned on: {rule}

gateOff

Approval step turned off: {rule}

grant

“{action}” permission granted to {role}

revoke

“{action}” permission removed from {role}

limit

{label}: {from} → {to}

stop

All agents stopped

resume

Agents restarted

REPRESENTATIVE INTERFACE DATA · Representative

Emergency stop

body

Stops all agents immediately. Unfinished work is saved and waits until it's resumed.

stop

Stop all agents

resume

Restart agents

stopped

All agents stopped

{count} agents running

held

{count} jobs waiting in the queue

running

running

paused

stopped

  • Human approval

    You decide which steps are critical; the agent stops there and asks.

  • Least privilege

    Each agent accesses only the systems and actions its job requires.

  • Audit log

    Every tool call, input, decision and approval is stored with a timestamp.

  • Limits

    When amount, volume or retry limits are hit, the agent stops.

  • Emergency stop

    One button stops every agent, and no work is lost.

WE MEASURE FIRST, THEN WE SCALE.

HOW WE WORK

We start with a single process. We measure the result in a pilot, then widen the scope and the agent's permissions step by step once your team is ready.

weeks

Week

  1. Discovery (1–2 weeks)

    We talk with your teams and map the repetitive processes that involve rules and exceptions. Together we choose which ones suit an agent.

    DELIVERED IN THIS PHASE

    • Process inventory and prioritization
    • Systems and data access list
    • Success criteria

    REPRESENTATIVE INTERFACE DATA · Representative

    Process inventory

    • Process
    • Frequency
    • Exceptions
    • Assessment

    • Invoice matching

      frequency

      Daily

      exceptions

      Frequent

      fit

      Pilot candidate

    • Return approval

      frequency

      Daily

      exceptions

      Moderate

      fit

      Next step

    • Lead qualification

      frequency

      Weekly

      exceptions

      Few

      fit

      Next step

    • Budget planning

      frequency

      Yearly

      exceptions

      Many

      fit

      Not a fit

    picked

    Chosen for the pilot: {name}

  2. Process map (1 week)

    We map the chosen process step by step: which system, which decision, which exception, and where human approval sits.

    DELIVERED IN THIS PHASE

    • End-to-end flow and exception map
    • Approval steps and permission model
    • Baseline measurement

    REPRESENTATIVE INTERFACE DATA · Representative

    Flow and exception map

    • Email arrived
    • Document read
    • Matched to PO
    • Tolerance check
    • Human approval
    • ERP entry
    gate

    approval step

    exception

    Exception: no PO number on the invoice → taken from the email body

    toggle

    Show the exception path

  3. Pilot agent (3–6 weeks)

    We connect the agent to your real systems and run it first in observation mode, then with limited permissions.

    DELIVERED IN THIS PHASE

    • Working pilot agent
    • Control panel and audit log
    • Test and exception scenarios

    REPRESENTATIVE INTERFACE DATA · Representative

    Pilot mode

    • Observation mode
    • Limited permissions
    • Post entry AP-5519 to the ERP
    • Reconciliation email to the supplier
    • Payment instruction for INV-20950
    suggested

    Suggested, not applied

    applied

    Applied

    approval

    Sent for approval

    blocked

    Outside permissions, not done

  4. Measured rollout (ongoing)

    We track results against the criteria set in the pilot and raise volume and permissions gradually as trust builds.

    DELIVERED IN THIS PHASE

    • Measurement dashboard
    • Staged permission plan
    • New process proposals

    REPRESENTATIVE INTERFACE DATA · Representative

    Permission levels

    • Observation

      body

      The agent reads and suggests; it changes no records.

    • Approved actions

      body

      Every action is applied with a person's approval.

    • Limited authority

      body

      Work within limits runs on its own; anything outside goes to approval.

    • Next process

      body

      The next process is added with the same controls.

    current

    now

    raise

    Go up a level

    lower

    Step back

  5. Handover and operations (ongoing)

    We teach your team to run the agent. If you prefer, we keep handling monitoring, updates and new processes.

    DELIVERED IN THIS PHASE

    • Operations guide and training
    • Monitoring and alerting setup
    • Maintenance and development support

    REPRESENTATIVE INTERFACE DATA · Representative

    Monitoring and alerts

    • 0

      when

      An approval waits longer than 4 hours

      then

      The owner gets a reminder

    • 1

      when

      The same error happens 3 times in a row

      then

      The agent stops, the team is alerted

    • 2

      when

      A new type of exception appears

      then

      The job is set aside for review

    report

    Weekly agent report

    reportDetail

    Every Monday in the team channel

    on

    on

    off

    off

QUESTIONS ABOUT AGENTS.

FAQ

How agents are kept in check, which systems they work with and how a pilot starts.

Let's look at your process

Discovery call

If your question isn't answered here, write to us. We can go through the details of your process together.

What happens if the agent makes a wrong decision?

The agent acts only within the permissions and limits you define. It asks for approval at critical steps and stops to ask when a case falls outside the policy. Every decision is logged with its reasoning, and all agents can be stopped with a single button.

Do we need to replace our existing systems?

No. Agents work on top of your email, ERP, CRM, e-commerce, document and support channels, through APIs and authorized tools.

How is this different from classic automation (RPA)?

A rule-based flow stops on a non-template document or an unexpected exception and drops the work into a queue. An agent reads the document, fills gaps from context and decides within policy; when a case is outside the policy, it asks.

Which processes suit an agent?

Repetitive processes that involve rules and exceptions and move across several systems. In discovery we map your processes with your teams and choose the right one together.

How does a pilot start, and how long does it take?

Discovery takes 1–2 weeks, the process map 1 week and the pilot agent 3–6 weeks. The pilot agent runs first in observation mode, then with limited permissions; permissions grow gradually against the criteria set in the pilot.

Can our own team run the agent?

Yes. In the handover phase we teach your team to run the agent. If you prefer, we keep handling monitoring, updates and new processes.

LET'S PICK YOUR FIRST PROCESS.

LET'S START

Tell us about the process your team spends the most time on. We'll assess together whether it's a fit for an agent, and what kind of pilot it could start with.

ACTIONS

DOCUMENTS

llms.txt

/llms.txt

agents.md

/en/agents.md

Every page is also served as Markdown: add .md to the address or request it with Accept: text/markdown.