Client Profitability: Reports vs Excel vs AI Agents

By Jude Lee · · Comparison

Accounting firm partners and an operations lead reviewing client profitability and WIP reports on a monitor

Why firm profitability is the last report nobody automates

Ask an operations lead which clients lost money last quarter and you’ll usually get a confident answer about the two obvious problem accounts and a shrug about the other forty. The data exists — time entries, invoices, write-offs, WIP, recurring fee schedules, subscription costs passed through — but it sits in three or four systems that don’t join cleanly, and nobody owns the join.

The result is that firm profitability gets measured at the entity level — revenue minus expenses — and almost never per engagement, which is where pricing decisions actually get made. That gap is the whole opportunity here.

What automation really means when the client is your own firm

Automation in accounting usually means removing manual keystrokes from a repeating process: bank rules that code recurring transactions, scheduled invoice runs, document OCR, approval routing. All rule-based, all deterministic.

Profitability analysis is a different animal. The mechanical part (pull hours, pull billed amounts, compute realization) is deterministic. The valuable part — why a given engagement slipped from, say, 92% realization to 68% (illustrative numbers, not a benchmark), and whether that’s scope creep, a staffing mismatch, or a one-time cleanup — is interpretive. That split is why this workflow sits in the middle of the AI readiness ranking for accounting automation: the math belongs in software, the narrative belongs to a drafting agent with a reviewer.

Four ways to get a monthly client profitability read

Option 1: built-in practice management reporting. Karbon, TaxDome, Canopy, Financial Cents and similar platforms ship realization and WIP views. Zero build cost, vendor-maintained, numbers tie to the system of record. The ceilings are specific: the platform only knows what’s inside it (not your software cost per client, not offshore contractor hours billed elsewhere, not the GL); realization formulas are generally fixed, so you can’t override how write-offs or fixed-fee progress are treated to match your partners’ definition; there’s no cost-to-serve allocation for subscriptions, software, or overhead; and drill-down frequently ends at an export — the moment you need the detail behind a number, you’re in Excel anyway.

Option 2: a spreadsheet model you maintain. Export time and billing, drop it into a model, add a pivot, add your allocation assumptions. Infinitely flexible, cheap, auditable by anyone who can read a formula. The failure mode is human: it gets refreshed when someone has a spare Friday, and the person who built it eventually leaves.

Option 3: a BI dashboard over the same exports. Power BI, Looker Studio, or your platform’s warehouse connector. Scheduled refresh, deterministic math, versioned logic, and no review gate — nothing is generated, so a number is either right or it isn’t. Cheaper to run than an agent and easier to defend to a partner who wants to trace a figure to source. What it won’t do is write anything: you get a chart showing realization fell, not a paragraph on which engagements moved and what plausibly caused it. If your partners actually read dashboards, stop at this option.

Option 4: an AI agent with governed, read-only access. The agent connects to practice management, time and billing, and ledger data, runs the same pulls monthly, computes the metrics you defined, and drafts a written per-client review with flagged exceptions. You review and sign off. Before you scope anything, read your platform’s official API documentation and confirm it exposes the fields this needs — read scopes covering time entries, invoices, write-offs, fee schedules, and WIP. Several practice-management APIs expose contacts, jobs, and tasks but not the billing detail that drives realization. If yours doesn’t, you’re on scheduled exports instead, which is workable but changes the build.

Reports and dashboards (buy/configure)
Live quickly. Numbers tie to source and are reproducible. Little or no review gate. But: single-system view, fixed or hand-coded metric definitions, no written analysis, and someone still has to interpret the chart. Best when one platform holds nearly all your time, billing, and fee data.
AI agent over your own data (build)
Joins billing data to GL cost data and to your own allocation rules. Produces narrative, not just numbers. But: real build and ongoing maintenance effort, a required human review step, and it inherits every data quality problem you have. Best when your economics live in three or more systems and a partner actually reads the memo.

Where an agent fits — and where it must stop

What it does well: consistent monthly extraction, cross-system joins, exception detection against thresholds you set, and a first-draft explanation per client that a partner edits in minutes instead of writing from scratch. It is a drafting and triage layer.

Where it breaks: it cannot distinguish “budget overrun” from “the client sent us a shoebox in March” unless someone recorded that somewhere. It will confidently narrate a variance caused by a time-entry coding error. And it has no business taking actions — no writing to invoices, no client emails, no fee schedule changes.

An agent that drafts the profitability memo is useful. An agent that acts on it is a liability — repricing is a relationship decision, not a data decision.

Building the agentic version, step by step

  1. Write the metric definitions first

    Realization, effective rate, WIP aging, write-off %, cost-to-serve. Define each in one sentence with the exact source field. If partners disagree on a definition, the agent will just automate the disagreement.

  2. Expose data read-only via MCP

    MCP (Model Context Protocol) is an open standard for giving an AI assistant governed access to your systems. As of 2026 there are MCP servers for several accounting platforms, and you can build a custom one over your own data — see connecting an AI assistant to QuickBooks or Xero and practice management APIs vs MCP servers. Scope it to read-only, firm-level financial data.

  3. Package the analysis as a skill

    A skill is a reusable instruction set: pull order, metric formulas, exception thresholds, memo format, and the explicit rule that unexplained variances get flagged rather than guessed at. Same output shape every month.

  4. Validate against three closed months with a stated pass threshold

    Rebuild those months by hand and define what counts as passing before you look — for example: per-client realization within 1 percentage point of the hand model, total billed hours matching exactly, and zero unexplained client omissions or additions. Any client the agent silently drops is a fail, not a rounding issue.

  5. Plan for the ways it degrades quietly

    Vendor API updates change schemas and a field disappears; a new engagement type (a fractional CFO retainer, a fixed-fee advisory block) isn’t in the skill and lands in the wrong bucket; a pull returns zero rows and the memo reads beautifully clean. Add a row-count and client-count sanity check that fails loudly, name one person who owns re-validation, and re-run the hand comparison quarterly and after any vendor release note.

  6. Route to a human sign-off gate

    Output lands as a draft with the partner as reviewer. Mandatory edit-and-approve before anyone acts. The AI review gate pattern applies here as much as to client deliverables.

  7. Log what ran and decide what leaves your environment

    Record which queries ran, which records were read, what the agent concluded. The real exposure in this workflow isn’t client ledgers — it’s that client names and fee amounts travel inside prompts, tool calls, and retained logs at a model provider. Decide deliberately whether to send identifiers at all or substitute client codes, and check retention terms. The IRS’s Publication 4557, Safeguarding Taxpayer Data, is the reference point for protecting taxpayer information; confirm your specific configuration with your IT/security advisor.

Modeling the payoff without inventing numbers

Don’t take anyone’s published savings figure, including ours — there isn’t one. Build the case from your own inputs:

(hrs/month × loaded rate) × 12
Analyst time currently spent rebuilding the model
Worked example — use your own figures
# clients repriced × fee delta × 12
Captured revenue from acting on findings sooner
Worked example — use your own figures
WIP days reduced × avg daily WIP
Cash released from faster billing of aged WIP
Worked example — use your own figures

Then subtract honestly: build or configuration time, monthly AI/API costs, re-validation time, and the reviewer’s time per cycle. Recovered analyst hours only count if they’re reallocated to billable or advisory work — if they go to inbox, they’re not savings. If your whole book is 25 clients on one platform, the built-in report probably wins this math and you should stop here.

Will AI replace the people doing this analysis?

No, and the framing is wrong. An agent that drafts a profitability memo doesn’t make the pricing call, hold the client conversation, or carry the licensure. What changes is who does the assembly work — the extract-join-format layer that currently eats a senior’s Friday. The roles that shrink are data-gathering roles; the ones that grow interpret and decide. If you’re mapping team skills, training on AI skills versus traditional automation courses is a more useful lens.

What the biggest firms run, and why it doesn’t settle your decision

The large firms publicly describe proprietary audit and data platforms — Deloitte’s Omnia, EY’s Canvas, KPMG’s Clara, PwC’s Aura among them — sitting on enterprise ERP and data warehouses. That architecture reflects their scale and audit obligations, not yours. A 15-person firm copying it buys cost without benefit; the relevant comparison is what a small firm’s agents can actually do versus Big 4 tooling.

Not sure where to start?

Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.

Get a free automation audit

Related guides