Key takeaways
- A working business voice agent is nine parts, from call routing to consent notices, and every part needs an owner after launch.
- In everyday conversation, answers to yes-or-no questions come about 208 milliseconds after the question on average[1]; every part of a build adds delay.
- The build bill is mostly hours: software developers earned a median $64.44 an hour in 2025[2], before upkeep and weekend cover.
- Buy for ordinary calls, build only with your own developer and an unusual call flow, skip if you rarely miss calls.
- A middle route exists: buy the voice layer and point your developer at webhooks and your own API.
Three routes, and who each one suits
Small businesses weighing an AI voice agent have three sensible routes. Buy a product that answers calls and configure it to your business. Build your own from parts, with a developer stitching together telephony, speech recognition, a language model and a synthetic voice. Or skip it for now and fix the basics: a better voicemail greeting, a callback routine and staggered lunch breaks. Each route is right for somebody.
Building is more realistic than it was two years ago. The parts are sold by the minute, tutorials are everywhere, and a capable developer can have a demo answering a phone over a weekend. The demo is not the hard part. The hard part is everything after it: the caller who mumbles a street name, the transfer that rings out at 9 pm, the change to your Saturday hours, the recording rules in the state your customer is calling from.
Interest is uneven by size. Census data for late 2025 to May 2026 show 37% of firms with 250 or more employees using AI, against under 20% of firms with four or fewer[3]. Large firms have developers and budgets for experiments. Most small businesses are really choosing between buying and skipping, and should think hard before building.
What a build actually includes
List the parts before anyone writes code. Each one is a decision now and a maintenance job later, and the maintenance is what most build estimates leave out.
| Part | What it does | Upkeep after launch |
|---|---|---|
| Phone number and routing | Receives forwarded calls, transfers them, handles busy and no-answer | Forwarding rules, transfer numbers when staff change, carrier issues |
| Speech to text | Turns the caller's words into text as they speak | Accuracy checks on accents, background noise and numbers |
| Language model and instructions | Decides what to say and which action to take | Instruction changes, retesting after every model update |
| Text to speech | Speaks each reply in a natural voice | Pronunciation of names, streets and your own products |
| Business knowledge | Hours, services, the prices you publish, policies | Every change to hours, staff or services |
| Actions | Booking, transfers, messages, lookups in your systems | Breakage whenever a calendar or API changes |
| Records | Recordings, transcripts, summaries, notifications | Storage, retention and who can see what |
| Monitoring | Alerts when calls fail or quality drops | Someone on call to act on the alerts |
| Consent and disclosure | Greeting notices, recording consent, opt-outs | Rules that differ by state and change over time |
Speed ties the parts together. In recorded everyday conversation across 10 languages, the average gap before an answer to a yes-or-no question was 208 milliseconds, and the most common gap was zero[1]. People notice a pause long before they could name it. A build that runs speech recognition, the model and speech synthesis one after another can easily take two seconds to reply, and callers start talking over it or hang up. Buying does not abolish this problem; it hands it to someone whose whole job is solving it.
Source: Proceedings of the National Academy of Sciences (via PubMed Central), 2009[1]
Show the data
| Item | Value |
|---|---|
| Japanese (fastest) | 7 ms |
| Mean, 10 languages | 208 ms |
| Italian | 310 ms |
| Danish (slowest) | 469 ms |
The build bill is mostly hours
Usage fees for the parts are visible and easy to compare. The hours are not. Software developers, quality assurance analysts and testers earned a median of $64.44 an hour in 2025[2], and that is wages alone, before benefits or an agency's margin.
Example: a developer spends 120 hours taking a voice agent from weekend demo to something you would let near a paying customer: call routing, one booking action, transfers, recordings, a set of 50 test calls and a page to read transcripts. At $64.44 an hour that is about $7,700 in wages. Upkeep then runs at, say, six hours a month for changes, model updates and fixes, about $390 a month. The hours here are assumptions; ask your own developer to estimate each row of the table above, and add a third for the rows nobody thought of.
Then there is the Saturday night problem. Phone lines fail at inconvenient times, and support work follows them: BLS notes that some computer support specialists work nights or weekends because support may need to be available 24 hours a day[4]. If your developer is a contractor who answers email on weekdays, your phone line has a weekday-only repair service. Decide who is on call before launch, and what callers hear while the agent is down. The sensible fallback is your old voicemail, reached by a forward that kicks in when the agent does not answer.
Before you approve a build, ask the developer three questions in writing. What happens to a call when one of the parts is down? How will you test a change before it reaches real callers? And who else could take over the code if you are unavailable for a month? Vague answers to any of the three are the real cost estimate.
Risks you inherit when you build
A product vendor carries some of the legal and security work for you. When you build, all of it is yours. Four areas come up on almost every business line.
- Recording consent. Federal law allows recording when one party to the call consents[5], but some states require everyone's consent. Washington, for example, treats consent as given when an announcement that the call is being recorded is made and is itself recorded[6]. Your greeting has to work for callers from every state you serve.
- Calling back. Answering calls is the easy case. If your agent places calls, the FCC treats AI-generated voices as artificial under the TCPA, so those calls need the called party's prior express consent unless an exemption applies[7].
- Card numbers. If callers read out card details, PCI rules say the security code cannot be kept after authorization, including in digital audio recordings that can be searched[8]. Most builds should refuse card details on the call altogether.
- Uneven accuracy. Across five commercial speech recognition systems, word error rates averaged 0.35 for Black speakers against 0.19 for white speakers[9]. Test with voices that sound like your customers, not only the developer's.
None of these is a reason to give up on building. They are a reason to write each one into the project plan with a named owner and a test, the same way a decent vendor does. And if you buy instead, ask the vendor how each of the four is handled, then check the answer on a call rather than in the contract.
Buy, build or skip: a decision grid
| Build it yourself | Buy a product | |
|---|---|---|
| First real call | Weeks to months, depending on the developer | Days, once your rules are written and tested |
| Who fixes it at 9 pm | Your developer, if reachable | The vendor's support and monitoring |
| Changing hours or services | An edit by someone technical | An edit by you or your team |
| Call records | Built and stored by you | Recording, transcript and summary included |
| Compliance work | All yours | Shared; check how the vendor handles it |
| Unusual call flows | Anything you can build | Limited to what the product supports |
Building wins on control; buying wins on time and upkeep. Most small businesses sit in the right-hand column.
Build when three things are true: you already employ or retain a developer with time to maintain it, your call flow is unusual enough that products cannot handle it, and you are prepared to own the risks above. A dispatch company with its own job system and a deep integration need might qualify. A dental office with ordinary booking calls almost certainly does not.
Buy when your calls are ordinary for your trade: questions, bookings, job details, urgent transfers. That describes most service businesses. Before you shop, sort your calls with the four-pile method and put these 15 questions to every vendor on your list.
Skip, for now, when you rarely miss calls, when your schedule is already booked weeks out, or when nobody will own the instructions. If budget is the only thing in the way, read what free AI receptionist tools leave out before settling on one.
Skipping well still takes some work: a voicemail greeting that says when you will call back, a rule that every missed call is returned within the hour, and a lunch rota so the desk is never empty at noon. Count missed calls for a month. If the number climbs, come back to this decision with real figures in hand.
Size the phone work before you decide
Whichever route you take, the payoff is phone time your team gets back, plus calls that stop going to voicemail. Count the first before you spend anything. The example below is a fictional seven-person property management office.
| Input | Example |
|---|---|
| Calls per working day | 40 |
| Minutes per call | 3 |
| Share of calls the agent would complete | 50% |
| Staff cost per hour | $36.73 |
| Calls a month | 868 |
| Staff hours on these calls | 43 |
| Hours handed off | 22 |
| Monthly value of that staff time | $797 |
How it is worked out: Calls a day × 21.7 working days × minutes per call ÷ 60 × share handled = hours; hours × hourly cost.
Assumptions
- Hourly cost of $36.73: private employers' average total compensation per hour worked for office and administrative support workers in June 2026.[10]
- 40 calls a day at three minutes each, half of them completed without staff, are invented for a seven-person property management office; use a week of your own call log.
Example only: about 22 hours of office time a month in this model. Set it against build hours and upkeep, or against a vendor's quote.
Now set that against the build example. If about 22 hours a month of office time is the prize, upkeep of six developer hours a month at nearly twice the hourly rate eats about half of it before a single usage fee. That arithmetic is why buying wins for most small offices, and why a build only makes sense when the agent must do something no product will.
The middle route: buy the voice, build the connections
Businesses with a developer have a fourth option. Buy the part that is hard to keep good, the live conversation, and point the developer at the connections, which are specific to you anyway. That is how the SI Employee is designed to fit. You forward your number, approve the greeting and rules, and can listen in, whisper guidance or take over a call while it is happening. After each call the summary arrives by email, with the recording and transcript in your account.
For the connections, signed webhooks pass every call to your own systems or through Zapier, Make or n8n, and the SI Employee can run live actions against your own API during a call once they are set up and tested on your account. The real estate team example shows the webhook route into a CRM, and the real estate guide to call summaries and CRM notes covers which fields are worth sending.
For the wider category, from basic AI receptionists to digital staff that take on more of the front desk, start with what an SI Employee is.
Frequently asked questions
How long does it take to build an AI voice agent?
A demo can take a weekend. Something safe for paying customers, with routing, transfers, records, testing and a fallback, takes weeks of a developer's time, and upkeep never stops. Ask for an estimate per part, not one number.
Can I build one without a developer?
No-code tools can produce a working demo, but the risks do not shrink: recording consent, transfers that fail, card details on calls and weekend outages still need an owner. Without someone technical on call, buying is the safer route.
How is an AI voice agent different from an AI receptionist?
The terms overlap. Voice agent usually describes the technology, which can be pointed at any call flow; AI receptionist describes the job of answering a business line. Most small businesses want the job done, whatever it is called.
Can a voice agent make outbound calls for my business?
It can, but the rules are stricter than for answering. Calls using an AI-generated voice need the called party's prior express consent unless an exemption applies, so get legal advice before any outbound campaign.
What should my developer work on if we buy?
The connections: sending call records into your own systems through webhooks, and giving the agent safe read or write access to your own API for things like job status or booking.
References
Proceedings of the National Academy of Sciences (via PubMed Central) (2009). Universals and cultural variation in turn-taking in conversation (PNAS 106(26)) (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2025). Software Developers, Quality Assurance Analysts, and Testers: Occupational Outlook Handbook (opens in a new tab) ↑
U.S. Census Bureau (2026). Large Firms With at Least 20 Employees Biggest AI Users (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2025). Computer Support Specialists: Occupational Outlook Handbook (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (2026). 18 U.S. Code 2511: Interception and disclosure of wire, oral, or electronic communications prohibited (opens in a new tab) ↑
Washington State Legislature (2026). RCW 9.73.030: Intercepting, recording, or divulging private communication; consent required (opens in a new tab) ↑
Federal Communications Commission (2024). Declaratory Ruling FCC 24-17: Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts (opens in a new tab) ↑
PCI Security Standards Council (2011). Information Supplement: Protecting Telephone-based Payment Card Data (opens in a new tab) ↑
Proceedings of the National Academy of Sciences (via PubMed Central) (2020). Racial disparities in automated speech recognition (PNAS 117(14)) (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2026). Employer Costs for Employee Compensation, Table 4: private industry workers by occupational and industry group, June 2026 (opens in a new tab) ↑