AI receptionist for small business in 2026: off-the-shelf vs custom (vs keeping your human)

What's actually at stake

Before any of this matters, the missed-call problem has to be real. For most small businesses, it is. 62.2% of small business calls go unanswered during business hours, and the average small business loses about $126,000 per year to those missed calls (Aira's 2026 missed-call study). For contractors and home-service businesses, individual missed calls are worth $200 to $2,000 each. After a single unanswered call, 78% of callers will move to the next business in the search results, and 93% never call back after reaching voicemail (SchedulingKit, 2026). The economics aren't subtle.

What's also real: the SERP for "AI receptionist for small business" is now uniformly SaaS-vendor content marketing and aggregator listicles. Every piece you read is either selling one specific tool or ranking ten of them as if every small business has the same needs. None of them publish the decision framework that tells you, as a non-technical owner, which of three paths fits — buy off-the-shelf, keep your human and automate the back office around them, or build something custom. The reason that framework is worth reading is that all three paths usually want help with implementation. Most small and mid-sized businesses are intimidated by setting these things up. Even when "buy" is the right answer, you still have to pick the right vendor, integrate it with your CRM and scheduling tools, tune the call flow, run real test calls, and own it once it's live. Even when "keep the human" is the right answer, you still have to wire up the structured-note prompts, build the CRM-to-calendar automation, configure missed-call recovery. Even when "custom" is the right answer, you have to scope it correctly and stop the project from creeping. We do all three kinds of work. Helping a five-person firm integrate Goodcall properly costs a fraction of a custom build. Helping an estate planning firm keep their human receptionist while fixing the back-office data layer is a different engagement again. Building a custom voice agent for a healthcare practice that needs HIPAA-grade handling is the third. The piece you're reading is meant to tell you which one your situation actually calls for — so whichever you land on, you're spending money on the right thing. The implementer-services niche is honest build-vs-buy advisory — the same frame we apply to AI for accounting firms, our own technical stack, and the five-layer framework.

The decision matrix

Here's the question tree. Run through it once with honest answers and you'll know which of the three paths fits your situation. The first question is a gate — if you answer yes, the AI questions below it stop mattering.

Question	If yes →
Gate question: Is your client base meaningfully older (60+) and/or is your brand explicitly built on a personal-touch human voice (estate planning, traditional wealth management, boutique probate or family law, concierge medicine, high-end hospitality)?	Keep the human. Automate the data layer instead. See "The third path" section below.
Do calls need to trigger actions in 3+ existing systems (CRM + scheduling + billing + accounting)?	Lean custom
Do you handle PHI, attorney-client privileged information, financial data under FINRA, or PCI cardholder data over the phone?	Lean custom (or HIPAA-specialty vendor)
Do you need bilingual or multilingual support beyond English-Spanish?	Lean custom
Does the AI need to know the caller's history (past visits, open matters, account standing) to handle the call well?	Lean custom (CRM-integrated)
Is your brand voice — the actual phrasing, humor, register — itself a competitive differentiator clients comment on?	Lean custom
Is your call volume above 2,000/month or growing past it within 12 months?	Lean custom (unit economics flip)
None of the gating + custom-justifying questions are yes?	Buy off-the-shelf.
Exactly one custom-justifying question is yes?	Buy off-the-shelf with a vendor that handles that one (e.g. healthcare-specific Smith.ai for HIPAA, hybrid AI + human Smith.ai for legal intake)
Two or more custom-justifying questions are yes?	Custom is justified.

The reason this matrix works is that the modern off-the-shelf tier has gotten genuinely good. Across 347,609 real business calls analyzed in the most recent industry benchmarking, top-tier AI receptionists resolve 90-95% of calls without human escalation, answer in under 5 seconds, and maintain 99% positive caller sentiment (NextPhone 2026 benchmarking). For routine appointment-booking, intake, and overflow work, that's enough. The cases where it's not enough are the cases the matrix above identifies — and they're the cases where building is genuinely justified, not the cases where some agency just wants to sell you a custom build.

The off-the-shelf shortlist (honest takes)

If the matrix above sends you to the buy side, here's where we'd actually look. We have no affiliate relationships with any of these vendors — this list reflects the tools we recommend (and integrate) for clients whose situations don't justify a custom build. Most of our receptionist work is on the buy side: pick the right vendor for the use case, configure it properly, integrate it into the CRM and scheduling stack the client already runs on, train the call flow against real call distribution, and stay on as the operator. The piece below is calibrated to help you do that yourself if you want to — and to help you know what good integration looks like if you'd rather have us do it.

Smith.ai — hybrid AI + human, best for legal and professional services

Worth understanding what Smith.ai actually is, because the marketing can blur it: this is a hybrid service, not a pure-AI tool. The AI handles routine call work — appointment booking, hours/location queries, simple intake — and complex or ambiguous calls escalate to a U.S.-based live human receptionist trained on legal and professional-services scripts. For firms whose brand is partly built on caller-facing care (law firms, accounting practices, concierge services) the hybrid handoff is the entire point: the AI handles the volume and the human handles the moments that matter. Pricing starts at $292/month for 30 calls and scales up from there (Smith.ai pricing). Where it falls short: the per-call pricing model gets expensive past about 200 calls/month; deep CRM integration (Clio, Karbon, Salesforce) is workable but limited to the integrations Smith.ai has shipped, not arbitrary endpoints. Best-fit scenario: a 2-15 person law firm or accounting practice doing under 200 calls/month, where some human-touch on the phone is part of the brand promise and where the cost of a botched intake is high.

Goodcall — cheapest entry tier, AI-only, simple use cases

Goodcall is the price-leader entry option, with plans at $59/month for 100 unique monthly callers, $79/month Starter, and $99/month Pro. Direct vendor pricing positions Goodcall as the AI-only alternative — no human escalation tier, just AI, with a goal of replacing humans entirely. This works when your call mix is genuinely simple (appointment booking, hours/location info, callback request) and you're comfortable with a 10-15% escalation-fail rate. It does not work when you need orchestration into multiple systems, compliance frameworks, or any kind of judgment-heavy intake. Best-fit scenario: a single-location dental practice, hair salon, restaurant, or service business with under 200 calls/month and a simple intake workflow.

Ruby Receptionists — premium live-human service, the deliberately "no AI" choice

Ruby charges $235-$1,640/month and runs entirely on U.S.-based live human receptionists, not AI (Smith.ai's own Ruby comparison). This is the right choice when your brand depends on every caller getting a live human voice and you have the budget — particularly for firms with older clientele who actively dislike bots, or for practice areas (estate planning, traditional wealth management, family-owned services) where the warmth of a human voice is part of what clients pay for. Worth being explicit: that's a real preference, not a quaint one. Boomer and Silent-Generation callers measurably abandon calls more often when they hit AI prompts, and for a 70-year-old considering an estate plan, the bot answering "Smith Law Firm, how can I help you?" is a brand signal that says they don't think I'm worth a person. Where Ruby gets expensive: the premium tier is roughly equivalent to a fully loaded part-time receptionist ($3,750-$4,000/month for the median U.S. receptionist position fully burdened, per NextPhone's pricing guide). Best-fit scenario: a high-touch professional services firm where the cost of one badly-handled call exceeds 6 months of Ruby fees, and where AI-answered calls would be brand-incongruent. Worth noting: Ruby and your existing in-house receptionist solve the same problem at different price points. The next section is about when keeping your own person — and adding workflow automation around them — is the better answer than swapping to either Ruby or AI.

Synthflow / Trillet — no-code voice AI builders, mid-tier flexibility

Synthflow and Trillet sit in the middle: more configurable than Smith.ai or Goodcall, much cheaper than custom development, with no-code interfaces that let a non-developer build basic call flows, integrate Calendly or HubSpot, and ship in a few days. Pricing varies but typically lands in the $99-$300/month range for SMB volumes. The trap: the no-code interface gives you enough rope to build a call flow that looks fine in testing and falls apart on edge cases your testing didn't catch (regional accents, callers who interrupt, callers asking off-script questions). If you have an in-house technical resource willing to maintain the flow as edge cases surface, these are excellent. If not, you'll be back to a managed vendor or a custom build within 6 months. Best-fit scenario: a 5-25 person agency, brokerage, or services firm with one technical-enough person willing to own the system.

Voksha — entry-tier $49/mo

The cheapest paid tier in the legitimate vendor set, Voksha starts at $49/month with 24/7 answering, security protections, and lead capture included. This works for a 1-2 person business that needs something better than voicemail, has under 50 calls/month, and is okay with a generic-feeling experience. We'd usually recommend the same person upgrade to Goodcall's $79/mo tier within a year as call volume justifies the better integrations.

The vendors we'd skip for SMB use

RingCentral, Dialpad, and the larger UCaaS platforms have AI receptionist features bolted onto enterprise phone systems. The pricing isn't bad, but the AI is typically the weakest part of the platform — it's a feature checkmark, not a focus. If you're not already a RingCentral or Dialpad customer, don't become one for the AI receptionist alone. Developer platforms like Vapi and Retell are excellent — but they're not products for non-technical owners. They're infrastructure that custom builds run on. If you'd be evaluating Vapi or Retell directly, you're already on the build side of the matrix.

The third path: keep your human receptionist, automate the data layer instead

This is the section most AI-receptionist content omits, because most of that content is selling AI. The honest answer for a real subset of firms is: don't replace your receptionist with anything. Keep them. Then fix the part of their job that's genuinely error-prone — the data layer.

The case for keeping the human is real and specific. If your client base skews older, your brand is built on a personal voice, or your practice area is one where callers measurably react against bots — estate planning, traditional wealth management, probate, certain family-law and elder-law practices, concierge medicine, family-owned services with multi-generational client relationships — the brand-drift cost of a bot answering the phone is much higher than the labor savings. Boomer and Silent-Generation callers abandon AI-answered calls at materially higher rates than younger cohorts, and for a 75-year-old considering whom to trust with their estate, the receptionist's voice is part of the trust signal. We've watched firms in these categories swap to AI to cut costs, watch their inbound conversion drop 20-30%, and quietly swap back six months later. That's not a rounding error — that's the entire AI ROI thesis collapsing because the brand fit was wrong.

But "keep the human" doesn't mean "leave them alone." A human receptionist is genuinely error-prone in ways that hurt the business — and the parts of their job that they're worst at are typically not the call itself, but the data hygiene around the call. The classic failure pattern: a caller leaves a 3-minute message about an estate review request. The receptionist takes notes on a sticky pad. Two days later the partner asks "did anyone call about that estate review?" and the sticky has been thrown away or the name is misspelled or the case type was misclassified or the callback never made it onto anyone's calendar. The call landed perfectly; the data downstream of the call dissolved.

This is the actual problem to solve, and it's solvable without removing the human voice. Human-centered automation — also called augmented receptionist workflows — keeps the human on the phone but automates everything around them. Some of the moves we deploy on these engagements:

• Real-time call transcription with structured-note prompts. The receptionist sees a structured form on screen as they take the call: caller name, callback number, matter type, urgency, summary. Voice-to-text fills it as they talk; they correct what's wrong. The note exists in the CRM the moment they hang up — no sticky pads, no after-the-fact data entry.
• Automatic CRM-to-calendar integration. If the call results in "schedule a callback Tuesday at 2pm with Maria," the system books it on Maria's calendar, sends Maria a Slack ping with the call notes, and sends the caller a confirmation SMS. The receptionist's job becomes "have the conversation" rather than "have the conversation and remember to do five things after."
• Missed-call SMS auto-recovery. Even with a human receptionist, calls go to voicemail when the receptionist is on another line, on lunch, or off for the day. An automated SMS goes out within 30 seconds: "Hi, this is the front desk at Smith Law — sorry we missed you. Were you calling about an existing matter or a new question? Reply here and someone will get back to you within 2 hours." That recovers about 40-50% of the calls that would otherwise have been lost.
• Post-call quality logs. A weekly summary of every call: caller name, matter type, what was promised, whether the promise was kept (callback completed, appointment confirmed, document sent). This is the metric layer the receptionist never had time to build, and it's how you actually find out whether your front desk is working.
• Bilingual call routing without changing voice. If the caller speaks Spanish, the system routes the call to a Spanish-speaking team member or a translated voicemail flow without requiring the AI-to-human handoff Smith.ai-style services use. The first voice is still human.

The cost: typically $200-$600/month in tooling (CRM, transcription service, SMS gateway, integration platform) plus a one-time integration build of $3,000-$8,000 to wire it all together. That sits on top of your existing receptionist payroll. For a firm currently paying a receptionist $3,750/month and losing 25% of inbound to data-hygiene problems, the math is: ~$4,200/month total cost, recovering an estimated 50-70% of those previously-lost data instances. The brand voice stays human. The error rate drops. Nobody on the client side knows you changed anything except that more callbacks happen on time.

This is also the path where an implementation partner earns its fee differently than on a custom-AI build. Less voice-AI engineering, more workflow architecture, CRM integration, and process design. We've shipped this pattern alongside the non-profit payroll tool for organizations whose people-first culture made AI-fronted intake a non-starter. The principle is the same across both engagements: automate the toil, leave the human contact intact. If the gating question at the top of the decision matrix returned yes for you — keep the human, fix the data layer.

The six cases that justify a custom build

Now the other side. These are the conditions we've actually seen — across the law firm intake we rebuilt in three weeks, the non-profit payroll tool, and the consultations we've passed back to off-the-shelf vendors when they didn't fit — that genuinely justify the custom-build investment.

1. Multi-system orchestration (3+ integrations)

If a successful call requires the AI to: (a) lookup caller in CRM, (b) check calendar availability, (c) book the appointment, (d) post to billing, (e) trigger a confirmation SMS, and (f) update the matter management system — that's six systems. No off-the-shelf tool reliably orchestrates six systems. Two systems? Sure. Three? Maybe. Six? Custom.

2. Industry compliance with BAA-required handling

HIPAA, FINRA, PCI-DSS, attorney-client privilege. Each carries specific technical and contractual requirements. HIPAA in particular requires a Business Associate Agreement with every vendor that touches PHI, and most general-purpose AI receptionists cannot sign BAAs because their underlying voice/LLM/storage stack isn't BAA-covered (see the PMC analysis of HIPAA gaps in AI vendor scope). The technical safeguards are well-defined: end-to-end encryption in transit and at rest, access controls, retention policies, audit logs. The contractual safeguards are the harder part. Healthcare-specialty vendors exist (MyAIFrontDesk healthcare tier, dentalaiassist, savvyagents) and can sign BAAs. Outside healthcare, the BAA-equivalent for legal or financial services is typically harder to source and the path of least resistance is custom on a BAA-covered cloud (AWS HIPAA-eligible services or Azure HealthCare).

3. Bilingual or multilingual beyond English-Spanish

Most off-the-shelf tools handle English well, Spanish acceptably, and everything else as a marketing claim. If your customer base genuinely calls in Mandarin, Vietnamese, Haitian Creole, Tagalog, or Portuguese — or if you need to handle code-switching mid-call (callers who switch between English and Spanish in the same sentence) — the off-the-shelf experience falls apart. Custom builds on Vapi or Retell with language-specific voice models (ElevenLabs multilingual, Cartesia's regional models, the latest 11Labs flash models) can hold up. Off-the-shelf cannot, and pretending otherwise costs you the calls you most needed to win.

4. Caller-history-aware personalization

"Hi, Mrs. Chen — I see you're calling about the matter we filed Tuesday. Do you want to talk to Maria, or shall I take a message?" That sentence requires the AI to: identify the caller from the inbound number, lookup their open matters in the practice management system, retrieve the staff assignment, and route accordingly. That's a CRM-integrated call flow with real-time data lookup mid-conversation. Off-the-shelf tools can do basic caller-ID-keyed routing (which menu the caller hears). They cannot reliably retrieve open-matter context and adapt the conversation around it.

5. Brand voice as a competitive moat

This one is rarer than agencies pretend. Most businesses' "brand voice" is generic enough that a polished off-the-shelf voice handles it fine. But occasionally — luxury concierge services, specialty professional firms, brands where the phone manner is itself part of what clients pay for — the voice matters enough that a custom-cloned voice (via ElevenLabs voice cloning or similar), specific phrasing patterns, and trained-in humor or warmth are worth the build investment. The test: would a current client recognize within 10 seconds whether they got the AI or a human? If the answer is "yes, and they'd care," custom is justified.

6. Call volumes above 2,000/month

Off-the-shelf pricing models are typically per-call or per-minute. At low volumes that's cheap. At 2,000+ calls/month, the math flips. Smith.ai at 2,000 calls runs into low-five-figures monthly. Goodcall would max its plan tiers. Custom builds on Vapi or Retell at $0.07-$0.20/minute all-in (per Retell's 2026 pricing breakdown) settle into the $1,500-$3,000/month range plus the amortized build cost. Above 2,000 calls/month, owning the stack saves money. Below that, you're paying more for the privilege of owning it.

Real cost comparison

Here's what the three options actually cost across realistic SMB call volumes. Everything is fully loaded — including the costs vendors don't put on the pricing page.

Option	100 calls/mo	500 calls/mo	2,000 calls/mo
Human receptionist alone (median, fully loaded)	$3,750-$4,000/mo	$3,750-$4,000/mo	$3,750-$4,000/mo + overflow
Human + workflow augmentation (year 1, amortized)	$4,150-$4,750/mo	$4,200-$4,800/mo	$4,300-$5,000/mo
Human + workflow augmentation (year 2+, ongoing only)	$3,950-$4,500/mo	$4,000-$4,550/mo	$4,100-$4,700/mo
Off-the-shelf AI — Goodcall / Synthflow	$59-$99/mo	$199-$299/mo	$500-$800/mo (often plan-capped)
Off-the-shelf hybrid — Smith.ai (AI + human)	$292/mo	$650-$900/mo	$2,500-$4,500/mo
Off-the-shelf live-human — Ruby	$235-$400/mo	$700-$1,200/mo	$1,640+/mo (plan-capped)
Custom AI build (year 1, amortized)	~$900-$1,200/mo	~$1,200-$1,600/mo	~$1,500-$2,500/mo
Custom AI build (year 2+, ongoing only)	~$400-$700/mo	~$600-$900/mo	~$900-$1,500/mo

Worth noting: the human-plus-augmentation row looks expensive next to pure-AI options because it includes the human's salary, which is the entire point. The comparison that matters for a firm choosing this path is "human alone" vs "human + augmentation" — and there, the math is $400-$1,000/month additional in year one to recover an estimated 50-70% of previously-lost-to-data-hygiene leads. For a firm where each lost lead is worth $2,000-$10,000 in lifetime client value, the augmentation pays for itself off a single recovered case per quarter.

Custom build math, shown explicitly: $5,000 setup amortized over 12 months ($417/mo) + $400-$700/mo platform/operations + variable per-minute. At 500 calls × 4 minutes average × $0.18/minute all-in = $360/mo variable. Year-1 total: roughly $1,177-$1,477/mo. Year-2+: roughly $760-$1,060/mo. The crossover with Smith.ai sits around 1,000-1,200 calls/month. The crossover with Ruby sits around 800 calls/month.

Worth noting: industry data on AI ROI runs about $3.50 per dollar invested, with AI cutting missed calls by 75% in deployments where the call flow is properly tuned (NextPhone's 2026 customer-service stats roundup). Both off-the-shelf and custom can hit those numbers if implemented well. Both can also fail to hit them if implemented poorly. The build-vs-buy decision is upstream of implementation quality, not a substitute for it. We've written more about realistic ROI math in our guide to AI automation ROI in 2026.

Implementation timeline expectations

How long it actually takes — not what the vendor's marketing says.

Off-the-shelf, basic config: 2-5 days from sign-up to live. The vendor's onboarding flow handles a generic call script; you tweak it for hours/business name; you point your phone forwarding at their number. Calls work. The first 50 calls reveal the script's gaps; you iterate.

Off-the-shelf with proper CRM integration: 2-3 weeks. You're configuring the integration (Calendly, HubSpot, Karbon, Clio), tuning escalation rules, training the call flow on your specific intake questions, recording any custom voice prompts, and running test calls. This is the deployment shape we'd recommend for any business serious about getting it right.

Human + workflow augmentation: 2-4 weeks. The receptionist stays exactly as they are — no retraining, no behavior change for the caller. The work is around them: provisioning the structured-note interface, wiring CRM-to-calendar integration, configuring missed-call SMS auto-recovery, building the post-call quality log. Faster than a custom AI build because there's no voice-AI engineering or call-flow tuning; slower than a basic off-the-shelf deploy because every integration touches your existing systems and has to be configured to your specific data model.

Custom build, well-scoped MVP: 4-8 weeks. Includes provisioning telephony, building the conversational flow, integrating CRM/scheduling/billing endpoints, voice selection and tuning, error handling and human escalation, observability and call-recording, and at least two weeks of test-call iteration before going live. We've shipped this shape of project in three weeks (the law firm intake) when scope was tight and the client moved fast on review cycles. Six weeks is more typical.

Custom build with compliance scope (HIPAA / FINRA / PCI): 12-16 weeks. The compliance work isn't the AI — it's the BAA negotiations, the security review, the audit-log architecture, the encryption-key management, and the compliance documentation. Anyone quoting under 8 weeks for a HIPAA-grade custom AI receptionist is either lying about scope or planning to deliver something you'll regret signing off on.

How each option fails (the practitioner-voice section)

What we've actually seen go wrong, in client work and in our own experiments.

Off-the-shelf failure mode #1: the silent escalation gap. Vendor says "we escalate complex calls to humans." Vendor's escalation queue gets backed up at 11pm Friday. Caller hears five minutes of hold music and hangs up. You don't know it happened until the angry email Monday. Mitigation: insist on real-time escalation SLAs in writing and review them quarterly.

Off-the-shelf failure mode #2: the integration cliff. The vendor "integrates with your CRM" — but the integration only writes new contacts, not new opportunities, and not into the custom field you actually care about. You discover this six weeks in. Mitigation: build your test-call list to include the specific data-flow you need, and run it before signing the annual contract.

Off-the-shelf failure mode #3: the price surprise. The $99/mo plan covers 100 calls. Your tax season hits 280 calls in March. You roll into overage pricing at $4/call, and your bill is $820 in March. Mitigation: pick a plan tier sized for your peak month, not your average month.

Custom-build failure mode #1: the demo-quality build. The agency's demo sounded great. The actual production system handles 60% of calls but routes 40% to "I'm sorry, let me transfer you" — including calls the demo handled. Mitigation: insist that the test-call set used during build mirrors your actual call distribution by intent and difficulty. We've documented this exact pattern in our broader small business automation challenges piece — the gap between demo and production is the single most common failure mode across automation projects.

Custom-build failure mode #2: orphaned ownership. Build is delivered. The agency moves on. A year later, an LLM provider deprecates a model, a TTS vendor changes their API, or a new compliance requirement emerges. Nobody owns it. The system degrades or breaks. Mitigation: contract a clear retainer for ongoing maintenance, or build with the explicit intent to bring it in-house and budget for an internal owner.

Custom-build failure mode #3: scope creep mid-build. Project scoped at 6 weeks. By week 3, you've added "oh, can it also do the Spanish line" and "can it post to QuickBooks too." Project finishes at 14 weeks at 2x budget. Mitigation: lock scope at the start, treat add-ons as a phase 2 conversation.

Human-only failure mode #1: data hygiene erosion. The receptionist takes the call perfectly. The notes never make it into the CRM, or make it in three days late, or make it in with the wrong matter type, or the callback never gets onto the partner's calendar. This is the single biggest hidden cost of a human-only setup, and it's invisible until you actually measure it. Mitigation: workflow augmentation (the third-path section above) — structured-note prompts and automated CRM-to-calendar integration eliminate most of this without changing the caller experience.

Human-only failure mode #2: scaling cliffs. Receptionist handles 80 calls a day fine. On a Monday after a big mailer goes out, calls hit 140. Half go to voicemail. Half of those callers never call back. The volume spike isn't the receptionist's fault; the architecture is the problem. Mitigation: missed-call SMS auto-recovery + an after-hours overflow vendor (off-the-shelf AI is fine here) for the spike days. The human stays on the routine load.

Human-only failure mode #3: PTO and sick-day blackouts. Receptionist is out for a week. Calls go to a voicemail nobody monitors. By the time anyone realizes, ~40 leads have been lost. Mitigation: same as #2 — automated SMS recovery plus a backstop vendor that activates when the human's phone forwarding flips on, configurable in any modern PBX or VoIP platform.

Bridging into the verticals

Every cell of this build-vs-buy matrix has a vertical-specific shape. Quick paragraphs on the four that come up most.

Accounting firms: Tax-season call volume is the load test that breaks most off-the-shelf tools. Most firms want CRM-integrated intake (Karbon or Canopy lookup), but compliance scope is lighter than law firms — usually no BAA-equivalent required. We've covered the broader accounting-firm AI implementation question in the AI for accountants pillar; the receptionist-specific recommendation for a 5-25 person firm is Smith.ai's professional-services tier or a Synthflow build with Karbon integration, with custom only justified above ~150 partners or in firms with embedded wealth management workflows.

Law firms: Attorney-client privilege requires the receptionist to NOT collect details that could become discoverable. Most off-the-shelf tools collect too much. Custom builds (or Smith.ai's law-specific tier) can be configured to ask only for callback information and matter type, deferring substantive intake to the attorney. Our three-week intake rebuild documents the privilege handling in detail.

Real estate brokerages: The use case is agent overflow — when the listed agent is showing a property, the call needs to go somewhere productive instead of voicemail. Off-the-shelf works fine for routing and lead capture; custom is only justified if you're doing specific listing-aware lookups (which property is the caller asking about, what's its current status, who's the showing agent today).

Insurance agencies: Claims intake is privacy-sensitive but typically doesn't trigger HIPAA. Renewal handling is high-volume and routine — perfect off-the-shelf territory. Custom is only justified for agencies handling commercial lines with multi-product orchestration (auto + homeowner + umbrella + business in one call).

Estate planning, traditional wealth management, elder-care practices: This is where the third-path "keep your human, automate the data layer" recommendation almost always wins. The clientele is older by definition. The brand promise is built on personal relationships across decades, often across generations of the same family. Bot-answered phones are an active brand mismatch and we've seen them measurably depress inbound conversion. The high-leverage work in these practices is making the human receptionist's job better — structured intake notes, automatic calendar wiring, missed-call recovery, post-call follow-up tracking. None of that requires touching the caller experience. All of it makes the firm tangibly better at the part of the job clients judge them on (did the callback happen? did the meeting actually get booked? did the document arrive when promised?). For these practices specifically, the question is rarely "which AI" — it's "what's wrong with how my human handles the back end of every call."

FAQ

What's the cheapest AI receptionist for a small business in 2026?

Voksha at $49/month and Goodcall at $59/month sit at the lowest paid tier for legitimate AI receptionist services. Both include 24/7 call answering, basic lead capture, and security protections. For free-tier exploration, most vendors offer 14-day trials. But "cheapest" is the wrong frame — pick by call volume, integrations needed, and compliance scope. A $59/mo tool that doesn't integrate with your CRM costs more than a $200/mo tool that does, in lost rework time and missed-data follow-ups.

Should I buy Smith.ai or build a custom AI receptionist?

Buy Smith.ai if you need legal or professional-services receptionist work with hybrid AI + human escalation, your call volume is under 500/mo, and you don't need deep CRM orchestration. Build custom if you need three or more system integrations, industry-specific compliance (HIPAA, FINRA, SOC 2), bilingual support beyond English-Spanish, or caller-history-aware personalization. Most small businesses land on the buy side. The decision matrix at the top of this article walks through the six conditions in detail.

How much does a custom AI receptionist actually cost in 2026?

$3,000-$7,000 for setup plus $300-$700/month platform plus $0.12-$0.35/minute all-in (covering speech-to-text, LLM inference, text-to-speech, and telephony) per Retell's 2026 pricing breakdown. Pencil it out for 500 calls/month at 4 minutes average and $0.20/minute: $400/mo variable + $500/mo platform + $5,000 setup amortized over 12 months ≈ $1,317/mo true cost in year one. Year-2+ drops to $760-$1,060/mo as the setup cost rolls off. That's roughly equivalent to Smith.ai's mid-tier pricing, with the tradeoff being you own the system and can extend it.

Can an AI receptionist be HIPAA-compliant?

Yes, but only if the vendor signs a Business Associate Agreement (BAA) and the underlying voice, LLM, and storage stack is BAA-covered. Most general-purpose off-the-shelf receptionists (Voksha, Trillet, default Synthflow) cannot sign BAAs because their underlying infrastructure isn't BAA-covered. Healthcare-specific vendors (MyAIFrontDesk healthcare tier, dentalaiassist, savvyagents healthcare config, Retell's healthcare deployment) can. Custom builds on AWS HIPAA-eligible services or Azure HealthCare with proper architecture (end-to-end encryption, access controls, retention policies, audit logs) can. Always verify the BAA before any call traffic touches the system.

What percentage of small business calls can an AI receptionist actually handle?

60-80% of routine calls per current industry data, with top-tier deployments hitting 90-95% resolution across recent benchmarking of 347,609 real business calls (NextPhone 2026). The 20-40% that need human escalation are typically: emotional or distressed callers, ambiguous-intent calls, complaints, large transactions over a comfort threshold, and callers with accents or dialects outside the model's training distribution. The right buy decision sets the human-escalation path before you go live, not after.

How long does it take to deploy a custom AI receptionist vs an off-the-shelf one?

Off-the-shelf with basic config: 2-5 days. Off-the-shelf with proper CRM integration: 2-3 weeks. Human-plus-workflow-augmentation deploy: 2-4 weeks. Custom build well-scoped MVP: 4-8 weeks. Custom build with compliance scope (HIPAA, FINRA, PCI): 12-16 weeks. Anyone quoting a custom HIPAA-grade build under 8 weeks is either misrepresenting scope or planning to ship something off-the-shelf with a custom invoice. Anyone quoting a non-compliance custom build under 4 weeks is doing the same.

Should I just keep my human receptionist instead of using AI at all?

For some firms, yes — and it's a defensible answer, not a quaint one. The "keep your human" path is the right call when your client base skews older (60+), your brand is explicitly built on a personal voice, or your practice area is one where callers measurably react against bots: estate planning, traditional wealth management, probate, certain family-law and elder-law practices, concierge medicine, multi-generational family-owned services. For these practices, the brand-drift cost of an AI-answered phone is usually higher than the labor savings, and we've watched firms in these categories swap to AI, see inbound conversion drop 20-30%, and quietly swap back six months later. But "keep the human" doesn't mean "leave them alone." Human receptionists are typically excellent at the call itself and error-prone at the data layer around the call — sticky-pad notes that vanish, callbacks that miss the calendar, matter types that get misclassified. The right move is keep the human voice, automate the workflow around them: structured-note prompts during the call, automatic CRM-to-calendar integration, missed-call SMS auto-recovery. Cost: $400-$1,000/mo additional in year one on top of existing receptionist payroll. The brand voice stays human; the error rate drops. See the third-path section above for details.

The takeaway

There are three honest answers to "should I buy or build an AI receptionist for my small business," and the right one depends on your specific situation. For most SMBs, buy off-the-shelf — the 2026 tier is good enough for the cases most businesses actually have. For firms with older clientele or a personal-voice brand, keep your human and automate the data layer instead — the brand-fit cost of bot-answered calls is real, and the back-office workflow is where the human receptionist genuinely needs help. For situations that hit two or more of the six custom-justifying conditions, custom is the right call — but that's the smallest of the three buckets, not the default. If you're staring at a vendor pitch for "fully custom AI receptionist starting at $25,000," run your situation through the matrix at the top of this piece. If two or more conditions apply, custom is real. If only one applies, find a vendor that handles that one. If the gating brand/clientele question is yes, the answer isn't AI at all — it's keeping your person and fixing the data hygiene around them.

If you'd like a second opinion before committing to any of the three: that's literally what our free build-vs-buy assessment is. Tell us your call volume, your industry, your client demographics, and your existing systems; we'll tell you which path fits and what the implementation looks like. Either way, the engagement on our side is roughly the same shape: we help you pick the right answer, we get it integrated into the systems you already run on, and we stay on long enough to make sure it's actually working in production. Off-the-shelf integration engagements and human-augmentation engagements are smaller and faster than custom builds — and most of the time, one of those two is the right outcome.

How an Automaton engagement works · Talk to us about your specific scenario