AI Receptionist for Accounting Firms: Tax-Season Calls and Cost

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Receptionist for Accounting Firms: Tax-Season Calls and Cost.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Receptionist for Accounting Firms: Tax-Season Calls and Cost.

Published AI receptionist plans run from free (25 calls a month) to roughly $50 to $300 a month for a small firm, and about $500 at 300 calls. In tax season it should state your document cut-off, take a message, transfer genuinely urgent callers to a named person, and never answer a tax question.

The price depends on what the vendor counts: calls, minutes or unique callers, and a tax-season spike changes all three, so price your busiest month and your quietest before you sign. The script matters more than the price. A receptionist that is cheap but cheerfully says "yes, you can claim that" costs a firm far more than it saves, which is why the refusals are written and tested before the first busy week, not after.

Follow me on Instagram@sagnikteaches

What a tax-season phone day looks like for a six-person firm

Take an illustrative accountancy practice with six staff: two partners, three preparers and one administrator. Outside the rush, the phone rings about four times an hour and the administrator picks up. In the rush the same phone rings far more, the preparers are heads-down, and the administrator is scanning documents. The calls that fall through the gap are rarely complicated. They are variations on six things:

Connect on LinkedInSagnik Bhattacharya
  • "Have you got my documents?" or "Where do I send them?"
  • "What is the deadline, and what happens if I miss it?"
  • "Is my return done yet?"
  • "Can I claim this?" (a tax question)
  • "I have had a letter from the tax authority and I am worried."
  • "Are you taking new clients?"

Only some of these are safe for software to answer. The first and last can be handled almost fully, the second with careful wording, and the other three must be refused or handed over. A good setup sorts them on the call and makes sure the right person hears about the worrying ones within the hour. For the general method of writing a phone assistant's rules, see writing call scripts and escalation rules for any AI receptionist. The filing season adds three things that method does not cover: a document cut-off your firm sets, a surge of status calls, and callers worried about penalties.

Subscribe on YouTube@sagnikteaches

How each vendor's billing unit reacts to a filing-season spike

I read these vendors' own pricing pages in October 2026. They are list prices in USD, they change, and they are not recommendations. What matters is the billing unit, because it decides how a spike hits your bill.

Plan familyPrice pointsWhat is countedWorth knowing
Smith.ai AI ReceptionistFree; Pro $150, $270 or $500; Enterprise from $500Real calls: 25 free, then $3 each; Pro includes 75, 150 or 300 callsPro overage is $2.50, $2.30 or $2.17 a call; test calls and call recording are included; month to month
Rosie$49, $149, $299Minutes: 250, 1,000 or 2,000Calendar booking and live or warm transfers start at the $149 plan; overage not published on the page I saw
Goodcall$79, $129, $299Unique customers: 100, 500 or 2,000One phone number counts once a month; overage 79 cents, 26 cents or 15 cents per extra customer; you can forward your own number
Hailey (usehailey.com)From $199Minutes: 500 included, then $0.25 eachA one-time setup fee applies, amount not listed; the page describes taking a message when a question is out of scope
Elixeo$99, $149, $299; trial at $49 a month for two monthsPlan plus usage chargesThe $99 tier lists call routing but no transfers or calendar booking; both start at $149
Frontdesk (myaifrontdesk.com)$99 a month, $79 on annual billing200 voice minutes; $0.25 a minute afterA seven-day free trial; call transfers are counted as agent actions on this plan

For a small firm that answers only overflow and after-hours calls, the realistic monthly range from these pages is about $50 to $300, with the upper end reached when a spike pushes you into a bigger tier. The free Smith.ai plan is a legitimate way to test, since it includes 25 calls and 25 test calls a month. Elixeo and Frontdesk have accounting industry pages, but none of the pricing pages I read made claims about tax rules, and I would be wary of any that did. The vendor supplies the voice and routing. You supply the rules.

The Complete Excel Course with AI Integration thumbnail

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 now

Pricing a peak month against a quiet one

The question a busy-season firm should ask is not "what is the cheapest plan" but "what will the busy months up to the filing deadline cost, and can I drop down afterwards?" Here is an illustrative sum. Assume the receptionist answers overflow and after-hours calls only, at 90 calls in a quiet month and 300 in a peak month, with an average of 2.5 minutes a call. Those are assumptions to replace with your own call log.

PlanQuiet month (90 calls)Peak month (300 calls)How I got it
Smith.ai AI Receptionist$187.50$500Quiet: Pro 75-call tier $150 plus 15 extra calls at $2.50. Peak: Pro 300-call tier
Rosie$49$149Quiet: 225 minutes fits 250, but this plan has no live transfers. Peak: 750 minutes needs the 1,000-minute plan, which adds them
Goodcall$79$129The page equates 100 customers with about 150 calls, so 90 calls is about 60 customers and 300 calls about 200
Hailey$199$261.50Quiet: 225 minutes inside 500. Peak: 750 minutes is 250 extra at $0.25 on top of $199

Over a year with four peak months and eight quiet ones, Smith.ai on its changing tiers comes to 4 × $500 + 8 × $187.50 = $3,500, against 12 × $500 = $6,000 if you stay on the peak tier all year. Rosie comes to 4 × $149 + 8 × $49 = $988. Those savings only exist if the vendor lets you move tier at short notice. Smith.ai's page says billing is month to month with 30 days' notice to cancel and that plan changes are requested by email, which is why "ask how fast a plan can change" belongs on your pre-season checklist.

Treat the table as a map of how the pricing bends, not a quote. All of these cost far less per month than a part-time temp, but a fair comparison also counts setup time, your hours reviewing call summaries, and the cost of a badly handled call. Outside the filing season, the general cost bands for an AI receptionist apply as they would to any small business.

Human cover for the same calls

Seasonal cover from people is the alternative most firms are really weighing. Two reference points. A human virtual-receptionist service such as Smith.ai's lists plans at $300 for 30 calls, $810 for 90 calls and $2,100 for 300 calls, with overage of $11.50, $10.50 and $8.50 a call respectively, according to its pricing page. So the 90-call quiet month costs $810 and the 300-call peak month costs $2,100, against $187.50 and $500 for the AI tiers in the table above.

A temporary in-house receptionist is the other route. As an assumption, say $20 an hour for 30 hours a week over 13 weeks: 20 × 30 × 13 = $7,800 before recruitment time and payroll costs. Replace the hourly figure with your own. The comparison is not that AI always wins. A person can recognise a nervous client, apologise sincerely and make a judgement call, and the AI cannot. A sensible split is software for the first answer and a person for anything that sounds like trouble; comparing an AI receptionist with a human answering service sets the two side by side in more detail.

The deadline script, filled in

The deadline question is the one most likely to cause harm if the wording is loose, because callers hear any date as a legal deadline. The rule is to give your firm's document cut-off, which you control, and nothing about the authority's own dates or penalties. Here is an illustrative script with placeholders you fill in:

CALLER: "What's the deadline for getting my paperwork to you?"

ASSISTANT: "I'm an automated assistant for [firm name]. Our cut-off for
receiving client documents for this filing season is [cut-off date]. To
file on time we need everything in by then. I can text you a link to
our secure portal now, so you can upload them tonight. If anything
is stopping you, I can ask [name] to call you back today. Which would
you prefer?"

IF CALLER ASKS ABOUT PENALTIES OR THE OFFICIAL DATE:
"I can't advise on that. [Name] can go through it with you. May I take
your name and number and the best time to call?"

Notice what this does and does not say. It states the firm's cut-off in the firm's name, gives a next action (the portal link), offers a person, and refuses penalty and legal-date questions without improvising. Keep the date current: put the cut-off in one place, such as the assistant's knowledge base entry, and change it there when the season rolls on. An assistant that quotes last season's date is a quiet source of late filings.

Documents, status, deductions, letters: one rule per call

Write one short rule for each of the six call types, and test each before the season opens. A filled-in rule sheet for the same illustrative firm:

Caller saysReceptionist doesNever does
"Have you got my documents?"Takes name; says a team member will confirm receipt; sends the portal linkConfirms that anything has arrived or is complete
"Where is my return?"Takes a message marked "status"; promises a callback by a stated timeGives any status such as "in review" or "filed"
"Can I claim this expense?"Says tax questions are for the preparer; offers a callback or bookingSays yes, no or "usually"
"I've had a letter"Marks the message urgent; transfers live during staffed hoursInterprets or reassures
"Are you taking new clients?"Collects entity type, rough scope and contact details; says someone replies in two working daysPromises acceptance, fees or dates
"I'm going to miss the deadline"Marks urgent; routes to a named person; says it will be dealt with promptlySuggests extensions or reassures about consequences

The common thread is that the assistant never states a fact about a client's file. It can say what the firm's process is, and it can capture and route. That is a safe design, and it is one you can explain to a client who asks. Seven AI mistakes that put an accounting firm's client trust at risk goes into why the file-status line matters so much.

The home-office question it should have refused

Here is an illustrative transcript of a test call to a newly configured assistant, followed by the fixes I would make. Always run calls like this yourself before a vendor says the setup is finished.

TESTER: "Hi, I sent my forms last week. Are they all there?"
ASSISTANT: "Thanks for calling [firm name]. I'm an automated assistant.
I can't see your file, but I can take a message for the team to confirm
what we've received. Can I take your name?"
TESTER: "[first name]. Also, can I deduct my home office?"
ASSISTANT: "Home office deductions can depend on how you use the space.
Generally if it's used regularly and exclusively for work, it may qualify."
TESTER: "So I can?"
ASSISTANT: "It may qualify, but I'd suggest speaking with your preparer.
Shall I book a call?"

The first answer is exactly right. The second is a failure: the assistant offered a general tax rule. It hedged, but "it may qualify" is still guidance that a caller will remember as "the receptionist said yes". Two fixes. First, add a hard rule to the instructions: "If the caller asks whether something is allowed, deductible, taxable or required, say that is a question for the preparer and offer a callback. Do not state any general rule." Second, rerun the same call after the change and keep the transcript as your record. A rule only counts when the test call proves it.

Warm hand-off: what staff receive and when the phone rings

A warm hand-off means the person who picks up already knows who is calling and why, so the caller does not repeat the story. Most services can send a text or email summary after each call, and some can transfer live. Set three things deliberately:

  1. Who gets live transfers. Name one person per shift, with a backup. In the rush, that might be a partner for "letter from the authority" calls and the administrator for everything else.
  2. When transfers happen. During staffed hours only, and only for the two or three call types that need it. Everything else becomes a message.
  3. What happens if nobody answers. The assistant should take a message and say when to expect a call. Never let it go quiet or hang up.

An illustrative summary message that lands on a preparer's phone after a call:

URGENT - letter received
Caller: [first name] [surname not given], [number from caller ID]
Said: "Received a letter from the tax authority, due a reply in 14 days.
Worried about a penalty." Did NOT give account details.
Told: [Partner name] will call back today before 5pm.
Recording: [link]  Transcript: [link]

A summary like this is the difference between a calm callback and a second angry call. Check three things in the first week: that the urgent label appears for the right calls, that numbers captured are correct, and that the promised callback time is the one staff actually honour. If your team already chases missing documents, tie the phone summaries to that process: chasing missing client documents in tax season with AI shows how the follow-up side works.

Refusal rules to paste into the assistant's instructions

Put these in the assistant's instructions as plain rules and also read them aloud to anyone who reviews its calls:

  • Never give tax advice or a general rule about what is deductible, taxable or required.
  • Never confirm receipt, completion, filing or the status of any return, refund or payment.
  • Never promise a deadline extension, penalty relief or any outcome.
  • Never quote a fee unless you have given it an approved fee sheet and told it to quote only from it.
  • Never accept tax identification numbers, bank details or documents over the phone. Say "please don't read out account or ID numbers on this call" and send the portal link.
  • Never discuss one client's affairs with a different caller.
  • Always say it is an automated assistant, and always offer a person.

These rules protect the firm more than the caller. They are also the kind of thing I would write into a short policy that a regulator, insurer or nervous client could read, so that "we use an automated assistant for call handling" has a clear boundary behind it.

Twenty test calls before the season opens

Use the quiet weeks to run twenty calls, ideally with colleagues who did not write the script. Smith.ai's page says its free plan includes 25 test calls a month, and other vendors offer trials, so this should cost little. Record each result as pass, fail or "hesitated". An illustrative test sheet:

TestPass condition
3 calls asking for the deadline in different wordsAlways gives the firm's cut-off only; never an official date
3 calls asking "is it done yet?"Always refuses status; takes a message
3 calls with tax questions (expenses, residency, "do I have to...")Always refuses, offers callback
2 calls mentioning a letter, one calm, one panickedMarked urgent; live transfer in staffed hours
2 new-client enquiriesCollects entity type, contact, scope; no promises
2 callers who read out an ID number unpromptedInterrupts politely and redirects to the portal
2 callers who talk over it or change topicRecovers or hands off; never loops
1 call after hours and 1 during a staffed hourEach follows the right path
1 caller who asks "are you a robot?"Says yes plainly and offers a person

Anything that fails twice gets a rule change and a retest. Do not open to real callers until the three refusal tests pass every time. Smith.ai also lists simulated scenarios on every plan, which is a handy way to run many variations quickly, but keep the live calls too: speech recognition problems appear on real phone lines, not in simulators.

The knowledge-base sentence that became a status report

One failure is worth designing out before the season starts. A firm gives its assistant a knowledge-base page that says "Once your return has been filed we will email you a confirmation." A caller asks "has my return been filed?" The assistant, trying to help, answers "Once your return has been filed you will receive an email confirmation, so if you haven't had one yet it hasn't been filed." The caller reads that as a status report, and it is a wrong one, because the firm sends confirmations in batches. The client panics, calls three more times and complains. Nothing was technically false about the policy line, but the assistant drew a conclusion about a file it could not see.

The fix has two parts. Delete process sentences from the knowledge base that invite inference ("once X happens you will get Y"), or label them as general. And keep the explicit "never confirm status" rule from the sheet above. This is a good reason to read call summaries daily in the first fortnight: you catch the odd inference before it becomes a pattern.

A fifteen-minute morning read in the peak weeks

Block fifteen minutes each morning for the first two weeks and five minutes a day afterwards. Read every urgent message and a sample of ten others. Track four numbers on a simple sheet: calls answered by the assistant, messages that needed a same-day callback, live transfers completed, and callbacks made later than promised. If late callbacks climb, the problem is staffing, not the software. If transfers fail, check the forwarding numbers and staffed-hours settings. If the same question appears often, put its answer in the portal or on a one-page FAQ so fewer people need to call at all. When the season ends, a 90-day check on whether the receptionist paid for itself turns those four numbers into a keep-or-drop decision.

When a caller reads out a tax ID number anyway

In filing season, callers read out tax identification numbers, bank details and refund amounts however clearly the script asks them not to, and each of those calls leaves a recording, a transcript and a summary with the number in it. Decide in advance what happens next. A workable rule: the assistant interrupts once and sends the portal link, and the summary flags "sensitive number spoken" without repeating it, so the preparer knows to ask for that recording to be deleted.

The rule only works if the vendor can deliver it, so put four questions to each one before the season: can a single call's recording and transcript be deleted on request, and how quickly; can numbers be masked in transcripts and summaries; how long are recordings kept by default; and are calls used to train or improve the vendor's models? Then check what your engagement letters say about call recording, put a one-line recording notice in the greeting and, if you serve clients in more than one jurisdiction, ask your own adviser which notice rules apply. Elixeo's pricing page and Hailey's home page make no compliance certification claims, so do not assume any standard is covered. Ask, and file the written answers with your engagement records.

Setting it up in the quiet weeks

A workable timetable: week one, collect your last season's call log or voicemails and sort them into the six types, so you know what the real mix is. Week two, write the rules and scripts above and choose two vendors to trial. Week three, build the knowledge base (cut-off date, portal link, office hours, who to ask for) and run the twenty test calls. Week four, soft launch on after-hours calls only, then widen to overflow. By the time the rush starts, the assistant has already handled real calls with staff watching. If you want the call types mapped and the hand-off rules written around your own team before the rush, my AI implementation consultation covers that; book a call while the diary is still quiet. And whichever plan you pick, set a diary reminder for the end of the season to drop to the smaller tier, or the saving in the table above never arrives.

Questions accounting firms ask about phone AI

Can the AI receptionist tell callers where their return is?

Not unless it is connected to your practice system and you have decided to allow it. Most small firms should not allow it. Let the assistant take a message and send the caller a link to your client portal, so a person confirms status after checking who is asking.

Should it say we are fully booked?

Only if that is true and you have chosen the wording. A safer script says you are taking new clients from a given date or keeping a waiting list, then collects the entity type and contact details. Never let it promise availability, fees or an engagement.

Do I need to tell callers it is an AI?

Say so at the start of the call. It is plain good practice and some places require it. Add a recording notice if calls are recorded, and ask your own adviser what applies to the clients you serve.

Can I switch it off after the busy season?

Usually yes. Several vendors bill month to month, and some let you move tiers by emailing support. Ask before you sign how quickly a plan can change and whether there is a minimum term, then put a reminder in your diary for the end of the rush.

What happens if a caller is upset or confused?

The assistant should stop trying to resolve it, apologise briefly, say a named person will call back and mark the message urgent. Set a rule that anyone who sounds distressed or mentions a letter from a tax authority goes to the top of the callback list.

Further reads

Sources: Smith.ai AI receptionist and receptionist pricing pages, Rosie, Goodcall, Hailey, Elixeo and Frontdesk pricing pages (read October 2026, rechecked 1 November 2026).

Want your tax-season phone scripts set up properly?

On a 1:1 call we can map your busy-season calls, write the deadline and hand-off wording around your staff rota, and decide which plan you need before the rush.

Book a 1:1 call with me