Bookkeeper AI - AI Practice Assistant for Independent Bookkeepers - Vertical Agent Spec

One-line definition

An agent that drafts, routes, and optionally sends client-facing communications for independent bookkeepers based on live client data and engagement context.

The workflow it owns end-to-end

Trigger: Bookkeeper flags a situation (late invoice, scope creep, year-end reminder, price increase, client question about a transaction) either manually or via a rule-based watch on AR aging or inbox activity.
Context pull: Agent retrieves the relevant client record, including engagement letter terms, outstanding balance, last communication date, service tier, and prior message history.
Draft generation: Agent produces a draft calibrated to communication type (collections vs. onboarding vs. scope renegotiation vs. educational explainer) in the bookkeeper's established voice and tone.
Review gate: Draft surfaces in a simple approve/edit/discard interface. Bookkeeper approves in one click or edits before send.
Send and log: Agent sends via connected email, logs the communication against the client record, and sets a follow-up reminder if a response is expected within a defined window.

What it knows that a generic LLM doesn't

Industry vocabulary in context: "catch-up bookkeeping," "cleanup engagement," "bank feed disconnect," "chart of accounts restructure," and "reconciliation discrepancy" used correctly and not interchangeably.
Scope creep framing: The difference between a polite heads-up and a formal scope-change request, and which tone fits which client history.
QBO and Xero workflow rhythms: Knows that January and April are high-noise months, that year-end requests spike in Q1, and that client urgency during tax season rarely matches the actual deadline.
Collections escalation norms: Bookkeepers rarely want to sue a client. The agent knows the three-step soft escalation pattern the industry uses before hard collections: reminder, pause-of-services warning, final invoice.
Fee increase communication: The specific framing that retains clients (lead with value delivered, give 30 days notice, offer a call) versus the framing that triggers churn.
Engagement letter anchoring: When a client asks for something outside scope, the agent references the actual engagement terms rather than writing a generic brush-off.

What it explicitly declines

Tax advice: Any question requiring a CPA or EA license to answer. The agent redirects these to the bookkeeper's referral list or declines to draft a response at all.
Legal commitments: Nothing involving contract modification, liability waiver, or dispute resolution without explicit bookkeeper approval and a human-written sign-off.
Collections escalation past the soft tier: The agent does not draft demand letters, small claims notices, or anything a lawyer would need to review.
Auto-send on new client threads: For clients in the first 90 days of an engagement, the agent always requires manual approval before sending, regardless of auto-send settings configured elsewhere.

Tools and integrations required

QuickBooks Online API: AR aging reports, client billing status, transaction-level context for questions like "why did you charge me this."
Gmail or Outlook via OAuth: To send from the bookkeeper's actual address, not a relay domain that confuses or alarms clients.
Ignition or HoneyBook: Engagement letter terms, service tier, active vs. paused status per client.
Jetpack Workflow or Karbon: Task context so the agent knows where a client sits in a workflow before drafting (do not send a "your books are ready" email if the reconciliation task is still open).
Twilio (optional): SMS follow-ups for clients who do not respond to email, gated behind bookkeeper approval before any use.

Trust escalation: when it pings a human

Churn signal detected: Client mentions canceling, asks about their data export, or references a competitor. The agent stops, flags the thread, and waits for bookkeeper guidance before drafting anything.
Disputed charge: Any communication where the client is questioning a specific invoice or fee. The agent surfaces the engagement letter and the relevant QBO line items for the bookkeeper to review before responding.
Draft confidence below threshold: If the agent cannot locate a matching engagement record, cannot identify the communication type, or detects a tone mismatch (client is angry and the only calibrated template is neutral), it flags for human authorship rather than guessing.
Regulatory or legal mention: Client references an audit, an IRS notice, or a complaint of any kind. Agent stops and marks the thread as requiring human review before any response goes out.

Pricing model

A per-workflow model at $3 per completed communication (draft approved and sent) is more defensible than $39/mo flat, but harder to sell to a price-sensitive solo operator who will immediately calculate their monthly email volume. A more realistic structure is a base of $19/mo for up to 30 sent communications, then $1 per additional send, with a hard ceiling at $49/mo. That keeps cost predictable while aligning it to usage. The honest problem: a bookkeeper sending 15 client emails a month might pay $19, compare it to Claude at $20/mo, realize they can get equivalent output with a saved prompt, and churn within 60 days. The pricing only holds if the integration value (one-click send from inside QBO context, no copy-paste, no tab-switching) is genuinely and measurably faster than the free alternative.

Differentiation from a generic LLM wrapper

The case for this agent over "I'll just paste my workflow into Claude" rests entirely on two things: live data access and zero context-switching. If the agent pulls AR aging from QBO, matches it to the client's engagement tier in Ignition, and produces a calibrated draft without the bookkeeper leaving their browser tab, that is a real workflow improvement. If it requires the bookkeeper to copy-paste the client's balance, their contract terms, and their communication history into a chat window first, the agent adds no net value over the free tool they already have open. This product only survives as a standalone if it embeds directly inside QBO or Karbon as a sidebar or Chrome extension. Otherwise it is a prompt template with a subscription fee, and the 18-month commoditization risk identified in the GTM analysis is not a risk at all. It is a timeline.