Why discovery, not a demo
Most AI conversations at accounting firms start backwards: a tool demo looking for a problem. A discovery inverts that. It starts from the firm’s actual workflows, quantifies where the hours go, and only then asks what AI or automation could take off the table. Node8 ran exactly this with a mid-sized CPA firm — the findings are summarized in the engagement overview — and the process below is the repeatable part.
The premise: a CPA firm’s costs are mostly people-hours, and a surprising share of those hours are mechanical — retyping, copying, mapping, filing. You cannot prioritize what you have not measured, and you cannot measure what you have not watched someone actually do.
Step 1: Workflow inventory on real screens
The core of discovery is a long working session — in this case around two hours — where the people who do the work demonstrate it live. Not a slide about the tax process; the actual tax software, the actual spreadsheet, the actual document management system, shared on screen.
The walkthrough covers, in a document-heavy firm:
- Intake: how client documents arrive (portal, email, paper), who touches them first, and how often they arrive wrong.
- Preparation: data entry from source documents (K-1s, W-2s, 1099s) into tax software, trial balance handling, financial statement spreadsheets, and the audit workflow in tools like CCH Engagement.
- Output and filing: how finalized returns get saved, delivered to clients, and archived.
- Communication: how routine client questions, document requests, and compliance queries get answered.
Two rules make the session productive. First, insist on volumes and minutes for everything: “how many K-1s a year, how long does each take, how many people do this?” At this firm, that question turned “data entry is painful” into “about 5,000 K-1s a year, hours per complex return, fifteen people, 15,000+ minutes annually” — a number you can rank. Second, let people show their workarounds. The Excel template someone built, the manual remapping of client account names, the copy-paste from QuickBooks exports — workarounds are a map of what the official systems fail to do.
Step 2: Sort by where AI actually helps
Every candidate workflow gets sorted into one of three honest categories.
Automate: high-volume, rule-describable work. Document extraction and mapping into tax software, roll-forward of trial balances, saving finalized returns into document management when they’re marked final, generating standard document request lists. These often need conventional integration (import formats, APIs, file automation) as much as they need AI — discovery should say which.
Assist: work where AI drafts and a human decides. Responses to routine compliance and client questions, first-pass account mapping suggestions, flagging a misfiled client upload. The person stays in the loop by design, not as a disclaimer.
Leave alone: judgment and sign-off. Return review, audit opinions, tax positions, anything a CPA attests to. Also anything that fails the client-trust test — at this firm, recording client calls for AI transcription was raised and correctly set aside over privacy concerns. Writing these exclusions down matters as much as the opportunity list; it is what makes the rest credible to skeptical partners.
Step 3: Confidentiality and engagement-letter considerations
Before any pilot, three governance questions get answered explicitly:
- What is the firm’s AI data policy? Discovery almost always finds informal individual use of ChatGPT or Claude already happening — it did here. The fix is a firm-managed subscription with reviewed data-handling terms, plus clear internal rules on what client data may and may not enter a chat interface.
- What do engagement letters and professional obligations allow? Client confidentiality obligations and the firm’s own engagement terms shape whether client documents can flow through third-party AI services at all, and under what agreements. This gets reviewed before architecture, not after.
- Where does data physically live? Automation designs favor architectures where documents stay in systems the firm already controls — the document management system, the tax software — with AI applied inside that boundary rather than exporting client files to new places.
None of this is exotic, but skipping it is how firms end up with shadow AI usage and no defensible answer when a client asks.
Step 4: Pick the pilot
Discovery ends with a ranked list and one recommendation. The ranking math is deliberately simple: annual volume × minutes per occurrence × people affected, weighed against implementation difficulty and data sensitivity. The pilot should be:
- Big enough to matter — thousands of minutes a year, not a novelty.
- Measurable — time-per-document before and after, so the result is a number, not a feeling.
- Contained — one workflow, one team, weeks not quarters.
At this firm, K-1/1099 extraction and trial balance roll-forward topped the ranking, with document-filing automation as a fast follow. The discovery closed with a defined research period — Node8 validating the specific tooling against the firm’s actual software stack — and a follow-up committed for roughly a week later. That cadence is part of the playbook: discovery that doesn’t end in a dated next step decays into a nice conversation.
What the firm gets
A prioritized opportunity map with the math shown, explicit not-doing-this exclusions, a confidentiality baseline, a recommended firm-wide tooling approach, and one scoped pilot proposal. Enough to make a small, evidence-based first bet — which is how AI adoption at professional firms actually starts. More on Node8’s accounting work at /acct.
Work with Node8
If you want this run at your firm — a real workflow inventory, honest boundaries, and a pilot chosen on numbers — get in touch. Discovery is designed to be a small commitment that de-risks every decision after it.