Penny Invoice and Admin Agent Setup

We start Penny with a task you can describe from beginning to end. For example, an approved completion record arrives in a folder, Penny prepares a draft from the matching job details, and the office receives a supporting packet. The business chooses which record proves the job is ready. A confident message from an agent doesn't replace that evidence.

Your first workflow needs a predictable handoff. If every job is handled differently, we'll help identify the part that can be bounded before choosing a plan. A simpler draft from a controlled set of files may be more useful than a wide connection to every office application. We can define more office tasks later. First, the team needs to explain how the original task succeeds, fails, or waits for a decision.

An early estimate might show a proposed amount. A later change order might add work. A crew note might explain what happened without proving that the customer accepted the change. Penny's source rules should distinguish those records, carry their dates and identifiers, and flag disagreements for the reviewer. It should leave an unresolved amount unresolved.

We agree the fields your team needs. These can include the customer, job number, billing address, items, quantities, rates, and instructions from your accounting team. Optional notes remain visibly optional. If the completion evidence is missing, the output can become an exception notice rather than an invoice draft. That makes the next staff action clear and keeps an incomplete job from quietly entering the send queue.

The Stripe invoice API has distinct steps for creating a draft, finalizing it, sending it, and collecting payment. Its official client code also explains automatic advancement and collection controls. That distinction matters: a draft-first workflow needs the relevant automatic behavior reviewed, rather than assuming that the word draft prevents every later action. The official Stripe invoice resource explains those operations.

This is a setup consideration, rather than a claim that Penny already connects to your billing account. On the call, name the product, account type, existing integrations, and any automatic billing rules. We can plan a document-only task outside the accounting system when that is the better first step. Publishing an accounting record, sending it, and collecting funds each need their own agreed authority and tested path.

Microsoft Graph documents creating a message draft and sending an existing draft as separate operations. Draft creation requires permission to write mailbox content. That is still meaningful access, even when the workflow doesn't send email. We check the mailbox and the permission granted to it. Creating a draft still gives the connection access to private records.

A send request being accepted by a provider also doesn't establish that the customer read it or that every attachment arrived as intended. Your approval queue should distinguish prepared, approved, attempted, and confirmed states using the evidence the selected system provides. Delivery monitoring, bounce handling, and follow-up can be separate approved workflows. Those workflows need their own agreement; an email address in the first draft doesn't include them.

We discuss where this rule can be enforced in your selected tools. An application may provide an approval status or a protected send action. Another setup may require staff to send the approved document manually from their own account. The written agreement should explain that boundary in language the office can follow.

Approval availability matters too. If the reviewer is away or the approval system can't confirm a decision, the workflow needs a defined waiting or exception state. It shouldn't infer permission from silence, a confidence score, or a note inside an attached document. The OWASP AI Agent Security Cheat Sheet recommends binding high-impact permission to the exact action and checking it outside the agent's proposal process.

Checking the draft before sending is part of Penny's intended service. Payment initiation, bank changes, refunds, and other financial actions require separate decisions and account controls. We won't turn a general invoice-drafting request into standing permission for those operations.

We can define a task identifier from the authorized job record and relevant revision. Before another draft is created, the process checks its recorded state. An earlier attempt may have created an invoice even if its confirmation was lost. Check that result or ask staff before trying again. Blindly starting again can create a second bill.

Some providers support their own duplicate protection. Others require a small tracking record in the chosen workflow system. We assess what the actual tools support and include that work in the agreement. The acceptance exercise should cover repeated events, two staff members requesting the same task, and a connection failing after an operation was accepted.

Retries count as processed runs under the plan and may generate provider charges. We set a bounded retry policy and a named exception recipient so a fault doesn't become a background loop that keeps consuming usage.

The important starting point is ownership. Someone should decide which files belong to the task, how current versions are identified, and where completed output goes. Names such as final, latest, or new don't reliably resolve conflicting documents. A dated version and a source owner give the assistant and reviewer a firmer basis.

A useful document index can retain the file reference, relevant date, job identifier, and approved category. Proposed renaming or folder moves need the permissions and approval boundary agreed for that workflow. Deleting source records or reorganizing a whole shared drive is wider work than preparing an office summary.

If you're deciding whether to configure the first task yourself, our business AI agent guide walks through defining a narrow job and testing its handoff. Penny is the service path when you'd rather have us scope and set up that work with your team.

For the standard knowledge allowance, each page or file needs ready-to-use text. Its maximum size is five megabytes. Scans needing OCR, transcription, and large document collections require a separate quote. We'll decide whether those inputs belong in the first workflow and what evidence the reviewer needs when extraction is uncertain.

An extraction process should preserve a link or reference to the original page where the reviewer can check it. For a billing amount, a plausible reading isn't a substitute for a clear source. We can configure missing or conflicting fields to stop drafting or appear in an exception package, depending on the task you've approved.

Before onboarding, gather a small set of sample records with customer details removed. Include an ordinary job, an amended job, a document with a missing field, and a source your office would reject. Those examples help us understand the work without asking you to put a live customer record into a public website conversation.

A connector's permissions can be broader than the job description. For example, a mailbox permission used to create drafts can also allow other mailbox changes. That difference needs account-level and application controls, plus a working process for staff. An instruction that says prepare invoices carefully doesn't narrow a token's authority.

We prefer client-owned provider accounts with clear administrative ownership. Staff should use supported sign-in methods, without sending passwords through our website chat. Secrets belong in the selected system's protected configuration, and the handoff should identify how they are rotated or revoked.

The first role also needs a defined business boundary. If your office bills for several companies, distinguish their records and accounts before connecting them. Shared folders can hold records for several companies. The draft must use the right entity's address, bank details, and customer details.

Use the smallest record set that supports the task. A draft description often doesn't need a complete customer history. A testing sample usually doesn't need real contact details. Reducing the details passed between systems can make the workflow easier to inspect and limit unnecessary exposure.

Retention is an operating decision too. Decide which draft versions and task logs need to be retained, where they are held, and who can remove them under your policy. Your accounting records may have separate retention obligations. We can configure supported controls within the agreed work; your business supplies the relevant legal and accounting requirements.

If local or private processing is an important requirement, explore Atlas's local agent setup. A private model can change where inference happens, while mail, billing, backups, and other connections still need their own data-flow assessment. We'll quote that wider deployment separately.

We can test a sample input with a misleading instruction and check that the output stays within the agreed task. The agent should continue to use the business's authorized rules and account permissions. Suspicious content can be flagged for the reviewer without acting on it.

This protection needs more than a warning in the prompt. Tool availability, account permission, allowed destinations, and the final approval path each help define what can actually happen. If the first workflow only prepares a document, sending and payment tools can remain unavailable to that role.

A sensitive workflow also needs change control. Adding a new tool, broadening folder access, or changing the model can affect the boundary tested during setup. We review those changes within your support allowance where practical or quote the additional work before widening the approved work.

The queue should distinguish an incomplete source, a prepared draft, a reviewer request for corrections, and a decision ready for the next approved step. Those states make daily work easier to understand than one broad success label. They also help staff avoid sending a draft while another person is still changing it.

You may prefer a workflow that runs only when a staff member requests it. A scheduled morning package can be separately scoped if the source records and team availability support that approach. The schedule needs an agreed timezone, a defined input cutoff, and a stop method. An office computer that is asleep or a service that isn't running can affect the job.

We don't promise instant drafting or a fixed rollout date before reviewing your systems. The quote depends on the records, supported connections, approval process, and testing needed for your business.

Start with your own ordinary jobs. Record how staff currently prepare them and where they have to ask someone else for details. After the agreed pilot is running, compare similar work with the same review standard. A faster first draft can still require extra checking, so the total process matters.

Useful measures include drafts accepted without a material correction, tasks stopped for missing details, and duplicated requests recognized. Track reviewer time per comparable job too. A reduction in repetitive work can free the team for customers or more demanding tasks. It doesn't on its own mean a salary saving or a reason to remove a staff role.

Judge this task by the effort your office actually spends on it. Your operating check should use observed outcomes from your configured workflow, with the source period and limitations kept clear. If the work isn't useful, we can discuss narrowing it or stopping the optional process.

We prefer providers billed directly to the client. If Trojan pays an approved provider on your behalf, documented charges are passed through at cost. Before launch, agree the accounts, currency, usage allowance, and who can authorize an increase. A run allowance doesn't remove the provider's own billing rules.

The setup can include supported usage controls and an alert at eighty percent of the agreed budget. At the ceiling, optional outbound work should pause or go to the named team contact according to the written terms. Provider reporting can arrive late, so a safety buffer is more honest than a universal promise that every account stops at an exact dollar.

For a draft-only office task, don't buy voice or text capacity unless you need it. Optional channels add their own limits and charges. A new workflow that chases payment, phones customers, or sends bulk reminders needs its own work agreement and approval design.

For a home-service business, Hailey can support weather-related notices and response drafts within the chosen territory. Penny may later prepare documents from jobs the business has actually approved. A weather event alone shouldn't create a job, promise a crew, or establish a billable charge. The contractor weather-agent guide explains those operating boundaries.

Mercury addresses paid-campaign planning and approved account work. Advertising spend and customer billing have different owners and controls. When two roles share details, we define the handoff and permissions rather than granting every role access to every account. Team can include up to two roles within its agreed workflows; extra roles and wider work require the appropriate scope.

Also name the billing app, document location, reviewer, and person who owns account permissions. Tell us whether your priority is preparing the draft, collecting the supporting files, or making exceptions easier to handle. We'll choose the plan around that first useful result and identify any extra integration or data work before payment.

From our Indianapolis, Indiana base, Trojan works with U.S. businesses. Our first strategy call is free. During business hours, we aim to reply within five minutes. Call (317) 981-2386 to discuss a Penny setup, or send a callback request. Staff confirms the approved work and payment before we begin onboarding.