Start with four jobs where a wrong answer is cheap to catch: new-client intake checks, document extraction, document chasing and exception flags. Give each a named reviewer, a pass rule and a stop condition. Hold back anything that sends numbers to a client or touches a return until the first four have run clean for a month.
An agent differs from a chatbot in the way that matters here: it takes steps in your systems. It reads the inbox, writes to the tracker, drafts and sometimes sends. So the question is never only whether the answer is right; it is what the agent is allowed to do, and who checks before that becomes final. The answer differs by job, so each one below gets a gate card: starting permission level, reviewer, pass rule, stop condition and log.
What counts as an agent, and the four permission levels
A chatbot answers a question you type. An agent is given a goal, a set of tools and some access, and works through steps without waiting for each instruction. Tools sold for this include the agent modes in general assistants, such as ChatGPT Work, which OpenAI offers on the web for Plus, Pro, Business and Enterprise plans for longer multi-step tasks, and automation platforms with an AI step. The label matters less than the permissions, which come in four levels.
| Level | What the agent may do | Example in a CPA firm |
|---|---|---|
| L0 Read only | Read files or the inbox and answer a staff question | "Which clients still lack a signed engagement letter?" |
| L1 Draft only | Write drafts or notes into a holding area; nothing leaves the firm | A chase email waiting in a review tab |
| L2 Act with approval | Send or post after a named person approves each batch | Digest emails sent once the manager ticks them |
| L3 Act alone | Perform a routine step inside hard limits, without approval | Filing an incoming statement into the right client folder |
Two rules follow. Every job starts at L1, however good the demo looked. And nothing that carries figures to a client goes above L2, because a wrong number in a client's inbox cannot be recalled by pressing undo.
Choosing the first jobs with a four-factor score
Score each candidate job from 1 to 5 on four factors, where 5 always means "safer to automate": cost of a miss (5 is trivial), volume (5 is many a week), ease of checking (5 is a clear pass or fail) and data sensitivity (5 is low). Add the four. Here is an illustrative scoring for a six-person firm:
| Candidate job | Cost of a miss | Volume | Ease of checking | Sensitivity | Total |
|---|---|---|---|---|---|
| Chasing missing documents | 4 | 5 | 5 | 4 | 18 |
| Intake checklist checks | 4 | 3 | 5 | 3 | 15 |
| Extracting figures from statements | 3 | 5 | 4 | 3 | 15 |
| Bookkeeping exception flags | 4 | 4 | 4 | 3 | 15 |
| Meeting notes into tasks | 4 | 3 | 3 | 3 | 13 |
| Workpaper tick-and-tie | 3 | 3 | 3 | 3 | 12 |
| Replying to client emails | 2 | 4 | 2 | 3 | 11 |
| Sending client financial summaries | 1 | 3 | 2 | 2 | 8 |
| Drafting a tax position memo | 1 | 2 | 1 | 2 | 6 |
Take everything scoring 15 or more, which gives the four jobs below. The scores will differ at your firm; the exercise is worth doing because it makes partners argue about the factors, not the technology. A low total is not a permanent ban. The memo job scores 6 because a miss is costly and nobody can check it at a glance, and no amount of prompt polish changes either fact. The wider ranking in ten AI use cases ranked by set-up effort pairs well with this one, since effort and risk are separate questions.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowJob 1: Chasing missing documents
This is the main worked example, because it scores highest and the arithmetic is easy to follow. Take a six-person firm with 140 recurring clients, each needing about three chasers a season. By hand, a chaser takes six minutes: find what is missing, write, send, log. That is 140 × 3 × 6 = 2,520 minutes, or 42 hours. With an agent drafting one digest per client and a manager reviewing each draft for 1.5 minutes, the review time is 630 minutes, 10.5 hours. Add six hours of set-up and the first season saves about 25.5 hours; later seasons save 31.5. These are illustrative figures to replace with your own timings.
| Gate card | Entry |
|---|---|
| Starting level | L1 for two weeks, then L2 with batch approval |
| Agent does | Compares each client's request list with the stored documents, drafts one digest per client, logs the date |
| Reviewer | The manager on that client, in one sitting |
| Pass rule | Every listed item is truly missing, checked against the portal and the folder; the digest states no amounts owed |
| Stop condition | An email to a wrong client or recipient, or any mention of tax owed, refunds or deadlines the firm has not set |
| Log | Tracker row: drafted on, approved by, sent on |
A realistic miss shows why the pass rule says "checked against the portal". An agent that matches by file name chases a client for a bank statement uploaded two days earlier under a different name. The manager catches it, because the manager opens the portal before approving; the fix is to match uploads by date and content before drafting, not by file name. If you want chasing wording and cadence, how to chase missing client records with AI has it.
A digest the agent wrote, and what the manager changed
This is the instruction the chasing agent runs on, followed by an illustrative draft.
You draft one document-chase email per client. You do not send it.
Input: client first name, the request list with item numbers, the stored document list, and the
date each item was requested. Output: one email listing only items missing from the stored list.
Rules: no amounts owed, no refunds, no deadlines unless supplied in the input, no advice, no
attachments. If the stored list shows an item as received, never list it. End with a one-line
note for the reviewer naming any item you could not match.Subject: Documents still needed for your year-end
Hello [first name],
Three items on our list have not reached us yet:
3. Loan statement for the van, showing the balance at the year-end date
5. Insurance renewal schedule
7. Payroll summary for the final quarter
Could you upload them to the portal by [agreed date]? If any is hard to find, reply and tell
me and we will help.
Kind regards,
[your name]
Reviewer note: item 5 matched no stored file by name; it may have been uploaded under another name.The manager reads the reviewer note first. Item 5 had been stored under the insurer's name, so it came off the list and the email went out with two items. The edit took about 40 seconds, close to the review time assumed in the sum above, and it shows why the instruction makes the agent name what it could not match instead of guessing.
Job 2: Intake and engagement checks
Onboarding a client involves a checklist that people skip when busy: signed engagement letter, identity checks done under the firm's procedure, authority forms, prior-year returns, a tax organiser sent. An agent can compare the new client's folder with the checklist and list what is missing. It reads and compares, which is why it works at the lower levels.
| Checklist line | Evidence found | Status |
|---|---|---|
| Signed engagement letter | PDF dated, signed by the owner | Complete |
| Identity checks recorded | Form present, second director's section blank | Incomplete |
| Prior-year return | Not in folder; client emailed "will send" | Missing |
| Authority to act with the tax authority | Nothing found | Missing |
| Organiser issued | Sent, no reply logged | Waiting on client |
The illustrative table shows the habit to enforce. The prior-year return stays "Missing" although the client emailed that it would follow, because a line is complete only when it points to a stored file.
| Gate card | Entry |
|---|---|
| Starting level | L1; L2 for the internal checklist note only |
| Agent does | Compares the folder with the intake checklist and writes a missing-items list |
| Reviewer | Admin clears the exceptions list; a manager samples one in five completed files in month one |
| Pass rule | Each complete line points to a stored file, never to an email alone |
| Stop condition | A client marked complete while a signed engagement letter is missing |
| Log | Checklist version, date, who cleared it |
Job 3: Extracting figures from client documents
Statements, invoices and payroll summaries arrive as PDFs and photos. Extraction turns them into rows. The tools range from capture products, such as Dext (its pricing for businesses starts at $302.50 a year, billed annually, for 250 documents a month and five users) and Xero's built-in Smart Document Capture, which is free on all plans, to a general assistant reading a file. Whatever reads the document, the gate is the same: the rows must add up to the document's own total.
An illustrative failure shows the value of that rule. A statement listed a deposit of 8,450.00 and the agent read 6,450.00. Nothing in the text looked wrong, but the extracted lines summed to 2,000 less than the movement between opening and closing balances, so the tie-out flagged it before anything was posted. A second illustrative failure is quieter: two accounts printed on one statement are merged into one table, and the tie-out failed because two closing balances cannot match one total.
| Gate card | Entry |
|---|---|
| Starting level | L1 until 200 documents in a row tie out; then L2 with sampling |
| Agent does | Reads each document into a table with the printed control total beside the sum of its own rows |
| Reviewer | Bookkeeper reviews everything at first, then one in ten |
| Pass rule | Rows sum to the control total; dates and period are right; nothing posted without a pass |
| Stop condition | Any posted figure that fails the tie-out, which returns the job to full review |
| Log | Document ID, control total, sum of rows, pass or fail, reviewer |
The comparison of capture tools sits in what automation in accounting includes, which also explains where plain rules do the job without AI.
Job 4: Exception flags on the books
Most exception checks are rules and belong in formulas or your accounting software: duplicate amounts on the same date, round-sum journals, postings on weekends, a new payee above a set amount, uncategorised items older than 30 days. An AI step earns its place on the language side, reading descriptions and memos that rules cannot, such as a payee coded to office supplies whose memo says scaffold hire. Either way the agent writes to a review sheet and changes nothing in the ledger.
| Flag | Why raised | Reviewer's decision |
|---|---|---|
| Two payments of 1,840.00 to one supplier on one day | Possible duplicate | Real error: one reversed |
| Journal of 25,000.00 posted on a Sunday | Round sum, weekend | Fine: owner's loan repayment, documented |
| Hardware supplier coded to office supplies; memo "scaffold hire" | Description and category disagree | Real error: recode |
| New payee, 9,300.00, no invoice attached | Over the new-payee limit | Ask client for the invoice |
Track two numbers. If fewer than 10% of flags turn out to be real across four weeks, the rules are too loose and are wasting review time. If the reviewer finds errors the agent never flagged, the rules are missing a pattern. Both are fixed by changing rules, not by asking the agent to "be more careful". The tutorial on spotting errors in client books with AI has a prompt for the language-based flags.
| Gate card | Entry |
|---|---|
| Starting level | L1 permanently; the agent never edits a posted entry |
| Agent does | Runs rule checks plus a language check on descriptions; writes flagged items to the review sheet |
| Reviewer | Senior bookkeeper clears the sheet weekly |
| Pass rule | Every flag has a decision recorded: real error, fine, or ask the client |
| Stop condition | The agent changes a ledger entry, or real errors fall below 10% of flags for four weeks |
| Log | Flag, rule, decision, decided by, date |
Jobs to hold back, and what would change the answer
| Job | Why it waits | What would make a firm revisit it |
|---|---|---|
| Monthly packs and forecasts sent to clients | A wrong figure arrives in the client's inbox, with your name on it | Three months of clean exception data, then drafts at L1 with a manager reading every page |
| Tax positions and advice letters | Judgement, and a professional's responsibility | Not as an agent job; an assistant may help with research under review |
| Preparing or filing returns | The preparer signs a declaration | Drafting inside tax software with the preparer reviewing every figure; the agent never submits |
| Answering client emails unattended | An answer to "should I buy the van this year?" is advice | Logistics replies only, such as meeting times, at L2 |
| Anything touching bank logins or payments | Harm moves faster than review | No revisit; use exports and uploads |
Where the first four sit on the ladder
| Job | Month 1 | Month 3 | Approver | Sampling |
|---|---|---|---|---|
| Chasing | L1 | L2 | Manager on the client | All digests at first, then batch approval |
| Intake | L1 | L2 for checklist notes | Admin, then manager | One in five files |
| Extraction | L1 | L2 after 200 clean documents | Bookkeeper | One in ten after the threshold |
| Exception flags | L1 | L1 | Senior bookkeeper | Every flag decided |
Review time is a budget, so add it up before starting. At month 3 a six-person firm might see 40 chase digests a week at 1.5 minutes each (60 minutes), 6 intake files at 5 minutes (30), 300 extracted documents of which one in ten is sampled at 2 minutes each (60), and 25 exception flags at 2 minutes (50). That is 200 minutes, about 3.3 hours a week. If nobody has those hours, do not start the job: an unreviewed agent is only a faster way to make errors.
The pattern is that autonomy rises where checks are mechanical (a total ties, a file exists) and stays low where checks need judgement. For a sharper view of one case, what a bookkeeping agent does alone draws the same line inside the ledger.
A 90-day start, week by week
- Weeks 1 and 2: shadow mode. Pick the job, write its gate card, and run the agent beside the manual process. Nothing it produces is used; you compare.
- Weeks 3 to 6: live at L1. Staff use the drafts and review all of them. Record edits per draft and failures against the pass rule.
- Weeks 7 to 10: L2 where earned. Move chasing and intake to batch approval if edits are near zero and nothing has hit a stop condition. Keep extraction at full review until the threshold.
- Weeks 11 and 12: decide. Partners read the logs. Keep, tighten or retire each job, and choose the next candidate from the score table.
Add one job at a time. A firm running four new agents at once cannot tell which one caused a problem, and the reviewer drowns.
Three misconceptions that lead firms to skip the gate
"An agent is a cleverer prompt." A prompt produces text. An agent acts, so its mistakes reach clients and ledgers, which is why permissions matter more than prompt wording. "A review gate means we do not trust it." Gates are how trust is earned with evidence; the 200-document threshold converts "it seems fine" into a number. "If it is right 95% of the time, we can skip review." The question is what the other 5% cost. Five wrong figures in a hundred client emails is unacceptable; five wrong file names in a hundred is a nuisance. The cost of a miss, not the accuracy rate, sets the gate.
Firm policy points: data, clients and accountability
Write one page covering four things. Which data each agent may see, and where the documents live, ideally in the firm's own storage. Which plan the tools run on, because business plans of the major assistants are set not to train on business content by default, which is worth confirming in your workspace before client files pass through. Who owns each job and signs its monthly log. And what clients are told, in the engagement letter, using wording like that in the engagement letter tutorial. Keep a log from the first day, with one line per output batch and one per stop condition triggered. An illustrative week shows how little paperwork the gate needs:
| Day | Job | Output | Reviewer | Result |
|---|---|---|---|---|
| Monday | Chasing | 38 digests drafted | Manager A | 36 approved, 2 edited (item already uploaded) |
| Tuesday | Extraction | 61 documents | Bookkeeper | 59 tie out, 2 fail (two accounts merged on one statement) |
| Wednesday | Intake | 4 client files checked | Admin | 3 complete, 1 missing the signed letter, client chased |
| Thursday | Exception flags | 22 flags | Senior bookkeeper | 5 real errors (23%), 14 fine, 3 queries to clients |
| Friday | Log review | Week's entries | Partner | No stop condition triggered |
If a stop condition does trigger, add a line saying what caused it and what changed before the job restarts.
Questions CPA firms ask before giving an agent a job
Do we need a dedicated agent product, or will our practice software do?
Start with the rule-based features in your practice software. Karbon lists Automatic Client Reminders on its Business plan, and TaxDome's pricing page lists document request lists and automated reminders. Those are predictable and cheap. Add an AI step only where reading or wording is the bottleneck, such as extracting figures or drafting a digest.
How many documents should pass before we sample instead of checking all?
That is your policy to set. One workable rule: review every document until 200 in a row tie out to their own control totals with no failures, then check one in ten. Return to full review the moment one fails, and keep a running count of failures so partners can see the evidence behind the change.
Can an agent sign into a client's bank or portal for us?
Keep credentials away from agents. Ask clients to upload documents or send read-only exports, and let the agent work from the files. A shared login makes it impossible to show who did what, and an agent with live access to a bank or portal can do harm faster than anyone can review.
What should an engagement letter say about agents?
Say in plain words that the firm uses AI tools to read documents, draft routine messages and flag exceptions; that a named person reviews the output; where documents are stored; and that client data is not used to train models under the plan you use. Ask your own adviser to confirm the final wording.
Further reads
- AI in Accounting Examples: Six Worked Jobs for a Small Practice — Six more concrete jobs, each walked through end to end.
- Accountants AI Risk Checklist: Error Points, Each With a Named Check — A named check for every error point in the workflow.
- Accounting Firm Standard Operating Procedures With AI: Record to Publish — Writing each agent job into a procedure the team follows.
- TaxDome vs Karbon for a Small Firm: AI, Cost and Fit — Which practice platform already covers requests and reminders.
- AI for Accounting Teams in a Small Firm: Roles, Prompts and Sign-Off — Who writes prompts, who reviews, and how sign-off is shared.
- Can AI Help Accountants? What It Handles and What You Review — What AI handles reliably and what still needs an accountant.
- AI Agents for Auditors: Chasing Evidence Requests From Clients — The same chasing job under audit-style sign-off.
- Accounting Software for CPA Firms: Automation Features Compared — TaxDome, Karbon and Canopy compared on real automation, with list prices, plan gating and what is still beta, plus a five-seat cost sum.
- AI Chatbot for Accountants: Website FAQs Without Giving Tax Advice — Scope a website chatbot to fees, documents and dates, with refusal wording, a hand-off flow and a weekly log review for a small accountancy practice.
- AI Chatbot for Finance Firms: Client Questions Without Advice Risk — What a finance firm's website chatbot may answer, what must go to an adviser, and the disclosure, logging and sign-off a compliance lead will expect.
- AI for Real Estate Attorneys: Closing Checklists and Title Review — Build a checklist per transaction type, summarise title exceptions with page cites, and keep a deadline calendar, with a verification step for every exception.
- AI for Real Estate Brokers: Compliance File Review Step by Step — A checklist-driven way to review transaction files with AI: exception report, human sign-off and the data-handling rules that keep agent files safe.
- AI for Small CPA Firms: Chasing Missing Client Documents in Tax Season — Turn each client's organiser into a missing-documents list and a reminder sequence with AI, using placeholders so return data never enters the tool.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Karbon and TaxDome pricing pages; Dext pricing; Xero Smart Document Capture announcement; OpenAI and Anthropic pricing and business privacy pages. All read in November 2026.