Your Timesheet Is Not Project Accounting
An engineer finishes a site visit on Tuesday and enters time on Friday. The project code is wrong. The receipt is in email. Finance asks the project manager what happened. Everyone performs memory work, which is expensive and strangely absent from the invoice.
For a small engineering or consulting firm, this is a practical forward deployed AI problem. The engineer maps time and expense evidence into project review and ERP billing. The goal is one accountable handoff, not automated judgment.
Build One Work Record
The agent reads an activity note and suggests the client, phase, task, and billing treatment. Rules verify the worker, date, approved rate, accounting period, and project status. A manager approves exceptions before invoicing.
Each work record needs a durable identity. RFC 9562 defines the current UUID format without central registration. That matters when a request retries and the ERP must not create the same six hours twice.
For invoice exchange, use a known document model when it fits. The OASIS Universal Business Language standard includes reusable invoice schemas. Standards will not fix bad approvals, but they reduce nonsense at the integration boundary.
Reuse the Boring Machinery
The reusable asset is the connector, schema, coding prompt, rules, approval queue, audit log, and evaluation set. Over the past eight months, these pieces have become easier to package and test. The firm still owns its billing rules and exceptions. Generic magic makes specific mistakes scalable.
Keep the ERP as the financial record. Let AI propose classifications. Let code enforce required fields and totals. Let people approve unusual work. A polished dashboard over broken time data is still bullshit with charts.
Count the Rework
Illustrative calculation: 120 weekly entries at six minutes of correction require 12 hours. At two minutes each, they require four hours. At an assumed loaded labor cost of $50 per hour, that is $400 of weekly capacity, not guaranteed profit or payroll savings.
Track invoice lag, recoded entries, duplicate records, reopened periods, and approved hours not billed. Any revenue improvement must be proven through actual invoices and collections.
Finance should review exceptions. It should not spend Friday doing archaeology in a timesheet export.