CPA Billing AI: From Team Knowledge to Working Agent
An accounting firm does not get a dependable billing agent simply by connecting its invoicing software to AI. The agent needs to understand which work is billable, which client agreement applies, when an exception is authorized, and when a person must decide.
For one accounting-firm deployment, Fixed Labs spent two weeks working with the invoicing team, embedded in its Zoho Cliq channels and day-to-day team discussions. We gathered the data, explored the exceptions and turned the team’s billing knowledge into documented instructions for an agent. Fixed Labs now manages that supervised workflow through Susan.
This is the implementation side of the story: how knowledge held in conversations and employees’ heads becomes a working process. For the operational results and invoice checks, read the companion Susan billing-agent case study. These are our connected delivery offerings, not an independent product recommendation. The customer remains anonymous.
Why we worked inside the invoicing team’s channels
The useful information was not limited to a written procedure. It also lived in the conversations around the work: a client concession, an instruction to hold certain time, a special situation that changed the invoice, or a question about which office’s treatment applied.
Embedding in the team’s Cliq channels put Fixed Labs alongside those decisions. Instead of asking the firm to summarize its entire operation in one kickoff call, we worked with the invoicing team to collect the rules and possible exceptions that would determine what the agent should do.
A billing conversation is not yet an executable rule
Consider a typical statement: “Give this client the adjustment we discussed.” A person familiar with the account may know exactly what that means. An agent needs an attributable instruction: which client, which service, which period, what adjustment, and whose approval governs it.
The following is an illustrative rule-mapping example, not a transcript or a reconstruction of a customer invoice:
| What the team knows | What the documentation must resolve | What to do if it is missing |
|---|---|---|
| This client has a special discount | Approval, reason, eligible service and period | Ask for the authorized treatment |
| Some time should not be billed yet | Which entries are held and when they may be reconsidered | Preserve the hold; do not guess |
| This office bills differently | Office routing, template and applicable rule | Confirm the correct billing path |
| We should discount an inefficiency | Who decides the adjustment and what work it covers | Request a commercial decision |
| We already prepared this invoice | Existing draft reference and current state | Check existing work before creating more |
This is especially important when a firm serves many clients across offices, countries or states. “Our billing process” may actually contain many approved variations. The implementation must preserve those differences, not flatten them into one generic invoice template.
What does Fixed Labs need from your billing team?
A practical starting point is access to the people who prepare and approve invoices, the existing procedures, and representative examples that can be shared through an agreed access process. We need to understand both the normal path and the situations where an experienced employee stops to ask a question.
For another firm, we would begin by mapping the systems, the source of billing instructions, the eligible work queue and the approval boundary. We would agree what information is needed before accessing client records. Embedding with a team does not mean unrestricted access to all of its conversations or data.
If your procedures are incomplete, that is a reason to start with documentation—not to tell the agent to improvise. Our CPA AI readiness assessment scorecard explains how to assess source quality, exception handling and ownership before choosing a pilot.
How we turned billing knowledge into a supervised agent
The two-week collaboration was the preparation behind the autonomy. Fixed Labs worked with the invoicing team to collect the data and possible exceptions, create the documentation, and get the agent working under supervision. The ongoing operation is managed through Susan.
What does “training the agent” mean here?
It means teaching the workflow through documented rules, approved examples, feedback and defined checks. It does not mean we are claiming to have trained a new foundation model or fine-tuned one on the firm’s confidential records.
The agent needs to know how to interpret the firm’s instructions and how to recognize when they are insufficient. A repeatable rule can be executed. An unresolved commercial decision must be escalated.
Follow an exception from conversation to invoice
Imagine a one-time concession for work that took longer than it should have. Simply seeing a lower amount on an earlier invoice does not establish permission to repeat it.
- Capture the decision. Identify the instruction and the person authorized to approve the concession.
- Define its scope. Record the affected client, service and period, and whether the treatment is one-time or recurring.
- Apply it to eligible work. Keep held or unrelated time separate from the current invoice.
- Check the saved result. Compare the draft’s customer, descriptions, amounts and selected work with the approved instructions.
- Keep the reason available. Preserve the basis for the decision so the reviewer can understand the treatment and the next period does not inherit it accidentally.
That sequence is a useful design pattern for another firm. It illustrates how to turn an exception into controlled work; it is not a claim that every historical invoice contained every proposed documentation field.
What software did this deployment use?
The workflow connects Zoho Projects and Zoho Books, with team handoff through Zoho Cliq. Susan gathers the relevant project and billing evidence, prepares or resumes an eligible invoice draft, checks the saved result and records the handoff. In this deployment, invoices remain unsent for human review.
If your firm uses different software, compatibility is a scoping question, not something this case proves. Fixed Labs would assess the required access, supported actions and verification steps before agreeing what to implement.
Who supervises the agent after launch?
Fixed Labs continues to manage this implementation through Susan’s managed AI-agent service. The firm’s people retain authority over commercial decisions and invoice approval. Ongoing management is the continuation of the implementation—not a handoff of an unexplained automation and a goodbye.
In the featured workflow, at least 10% of weekly invoices still involve human intervention or feedback for exceptions outside the documented rules. Examples include special discounts, unusual client situations and adjustments for inefficiencies. This is an intervention share, not an error rate.
When a new exception arises, the important question is whether it is a one-off decision or an approved change to future behavior. A useful operating discipline is to document that distinction before updating the instructions. A casual conversation should not silently rewrite billing policy.
What this means for another accounting firm
The main opportunity is not merely faster invoice drafting. It is giving a billing employee a process that can gather evidence, apply documented rules and verify the draft while the employee focuses on review and exceptions. Depending on scope and complexity, one person supported by an agent can handle preparation work that previously required several people.
What results did the billing workflow produce?
The published Susan case records 513 completed invoice records across three observed weekly windows ending September 5, 2026. Susan measured approximately six minutes of agent processing per invoice, compared with the team’s approximately 15-minute manual preparation estimate.
At that manual baseline, 513 records represent 128.25 hours of equivalent preparation work. That is workload context, not a claim that an entire position disappeared or that every part of billing and collection was automated. The full weekly breakdown belongs in the linked operating case; this article explains the implementation work behind it.
Will our firm also be ready in two weeks?
Two weeks describes the embedded preparation for this firm, not a universal implementation promise. A firm with documented rules, accessible source records and an available reviewer starts in a different position from one with conflicting agreements and knowledge held by a single employee.
For a new engagement, the useful sequence is to map one workflow, document its variations, agree approval boundaries, review representative outputs and then expand the supported scope. The schedule follows those requirements.
Should we hire a billing employee or implement an agent?
Look at the work you need covered. If the gap is recurring preparation—finding instructions, selecting time, applying known rules and checking drafts—an agent may increase the capacity of your existing employee. If the gap is client relationships, commercial judgment or approval ownership, you still need an accountable person.
The assessment should separate those responsibilities before making a staffing decision. Counting invoices alone misses the time spent resolving unclear instructions and unusual cases.
What should we prepare before speaking with Fixed Labs?
- The systems used for projects, time tracking, invoicing and team communication.
- Approximate weekly invoice volume and preparation effort.
- The exceptions that repeatedly interrupt the team.
- Where client terms, discounts and office-specific rules are documented today.
- Who can explain the workflow and who approves changes to billing treatment.
You do not need to send client records for the initial conversation. Start with a description of the work. Fixed Labs can then discuss what needs to be assessed, what needs documentation, and whether a supervised implementation is a sensible next step.
Want us to map the billing work inside your firm? Discuss a billing workflow implementation with Fixed Labs. We will start with your systems, recurring workload and exceptions—not a promise that every process can be automated.
About this implementation account
Manuel Castillo, AI Lead Engineer, supplied the firsthand implementation account and the approved operating context used here. First draft assembled with AI from that account; the illustrative examples and proposed rollout steps are distinguished from observed customer work. Customer names, private channel content, billing amounts and confidential rules are omitted.