Drew spent 80-hour weeks for two months building something most companies never build: a single document that fully describes how the books get done at Haven. That document is the knowledge base a new bookkeeper would need. It is also the knowledge base an AI agent would need. Here is the plan.
Your bookkeeping SOP document is unusually agent-friendly. Macro overview to weekly cadence to monthly close to quarterly catch-up. Each leaf node is a single specific task paired with a single specific training video. There is no guessing about what the workflow is — it is already written down.
This is the rare case where the hardest part of AI automation, capturing institutional knowledge, is already done. The remaining work is translating that knowledge into agents that operate the same systems your bookkeeper does.
Most consultants will read the SOP cover-to-cover and then code each task by hand. That's a year of work and it's brittle the moment Buildertrend ships a UI change. The better approach is multi-agent: one reads, one plans, three debate the plan, one builds, and a supervisor watches the whole thing run.
A planning agent reads the SOP doc, every Loom transcript, and the Drive Homebase. It produces a structured system map — every task, input, output, and decision point.
A second agent proposes an architecture. Three independent critic agents debate it — one optimizing for accuracy, one for cost, one for resilience to platform changes.
Pick one workflow (Cost Inbox is the right one). Build it end-to-end with full observability, exception handling, and a daily QC report. Once it runs clean for two weeks, add the next.
This follows the wyth.ai engagement model: one session to find the opportunity, then we build it, then we keep building. The Discovery Session is on the house for Haven.
A focused working session with Drew. We ingest the SOP + 58 videos, map every workflow, rank the automation candidates, and come back in 48 hours with a written action plan and a prioritized build sequence.
Free for Haven (normally $1,500–$2,500)We take the top priority from Discovery — Cost Inbox — and build it end-to-end. Agent harness, MCP connections, Buildertrend browser rig, QBO sync, daily run with human approval gate. Shipped, working, on real data.
Included in the retainerEach month we identify Haven's next highest-value automation and build it: credit cards, vendor credits, accounting inbox, monthly invoice drafting, budget variance reports. A continuous improvement engine at a predictable monthly cost.
$6,000 / month retainerCost Inbox processing is mentioned across at least four Loom videos and is the most-clicked manual workflow in the entire SOP doc. It is daily, high-volume, and not judgment-heavy. If we nail this one first, it likely pays for the retainer by itself.
We pick it, build it well, prove it for two weeks, then expand. The rest of the workflows follow the same harness pattern once it's in place.
Pulled directly from the SOP doc and ordered for sequencing. Time-saved figures are order-of-magnitude, to be refined in the Discovery Session.
| # | Workflow | Frequency | Time saved / mo | Difficulty |
|---|---|---|---|---|
| 1 | Cost Inbox processing BT cost inbox → bill → QBO match | Daily | 20–40 hrs | Medium · UI-heavy |
| 2 | Credit card categorization QBO transaction matching + coding | Daily / Weekly | 15–25 hrs | Low · API |
| 3 | Accounting Email Inbox triage Sub invoices, vendor statements | Daily | 10–20 hrs | Low · API |
| 4 | Monthly Build invoice drafting Deposits, % progress, COs, pay apps | Monthly | 15–25 hrs | High · judgment |
| 5 | Monthly Design invoice drafting Hourly + lump sum contracts | Monthly | 8–12 hrs | Medium |
| 6 | Bank statement reconciliation Monthly close activity | Monthly | 5–10 hrs | Medium |
| 7 | BT vs QBO quality control Bills and expenses comparison | Monthly | 4–8 hrs | Low · both APIs |
| 8 | Scoreboard updates Zoyler + financial performance sheet | Weekly / Monthly | 6–10 hrs | Medium |
| 9 | Utility bill splitting Duke, Water, Meridian → per project | Monthly | 2–4 hrs | Low |
| 10 | Unknown / Uncategorized reports Staff follow-up pings | Weekly | 3–5 hrs | Low |
Data-to-data is always preferred. Computer-use is the fallback when no API exists. This is how each system gets touched:
The NDA is signed and the SOP is in hand. The Discovery Session is the last thing between a plan on paper and a working Cost Inbox demo on your real data. It's free, and the written action plan comes back within 48 hours.
Book the Discovery Session Review the plan