Key takeaways
- An assistant answers and hands the work to a person; an agent finishes the task during the call, inside limits you set.
- On a booking call the difference is a scheduled job versus a callback, and callbacks leak because 80% of US adults say they generally do not answer unknown numbers[2].
- Agents need more setup and fail differently: a bad rule means a wrong booking or transfer, not just a slow reply.
- Choose per call type, not per product: routine bookings and urgent transfers suit an agent; quotes, complaints and patient-specific questions suit assistant behavior.
- An SI Employee works as an agent where you have connected and tested the calendar and transfer path, and falls back to taking a request everywhere else.
Same caller, two endings
Put the same caller through both designs and the difference shows within a minute. Example: a homeowner rings a five-truck plumbing company at 7:15 pm because the kitchen sink has backed up for the second time this week. She wants someone out tomorrow, and she is calling the first company on the list that answers.
An assistant-style phone line answers politely, confirms the company covers her ZIP code, takes her name, address and number, and tells her the office will call in the morning to arrange a time. An agent-style line does all of that, then checks the schedule the owner connected, offers two windows for a drain cleaning, books the one she picks and reads it back.
Example endings. Assistant: "I've passed your details to the office, and someone will call you tomorrow morning to set a time." Agent: "You're booked for a drain cleaning tomorrow between 8 and 10 am at 22 Birch Lane. Is this number the best one to reach you on the day?"
Both calls were answered, and both callers were treated well. Only one ended with a job on the schedule. The other ended with a promise to call back, and callbacks are where bookings leak: in Pew's survey, 80% of US adults said they generally do not answer cellphone calls from unknown numbers[2]. By the time the office rings at 8:30, she may have booked the company that answered second.
What separates an agent from an assistant
The useful definition is practical. An assistant responds: it answers questions and records what the caller wants. An agent acts: it completes a step in another system, such as a booking or a transfer, while the caller is still on the line. Everything else follows from that one difference, including how much setup each needs and how each goes wrong.
| AI assistant | AI agent | |
|---|---|---|
| Caller wants to book | Takes the request for staff | Books into a connected calendar and reads it back |
| It does not know the answer | Says so and takes a message | Says so and takes a message |
| Urgent caller | Takes details, promises a callback | Transfers to the approved on-call number |
| After the call | A message someone must act on | A finished task plus a record of it |
| Setup needed | Information and a greeting | Information, rules, calendar and a transfer path |
| If set up badly | A slow callback | A wrong booking or a bad transfer |
The second row is the same on purpose: neither design should guess.
The last row is the one vendors skip. An agent's mistakes land in your schedule. A booking in the wrong window, a job outside your service area or a transfer to a number nobody answers on Sundays all cost more than a slow callback. That is why an agent needs tighter rules, a read-back on every booking and someone checking the day's bookings each morning.
Callers' own attitudes point the same way. In Pew's 2025 survey, 73% of US adults said they would let AI assist them at least a little with day-to-day activities, but only 13% would let it assist a lot, and 27% would not let it assist at all[1]. People accept help with a task; fewer want to hand over the whole thing. On the phone, that argues for an agent that confirms out loud and offers a person, not one that rushes.
Pick the design per call type
The mistake is to choose one design for the whole phone line. Most businesses want agent behavior on some calls and assistant behavior on others, and a good setup lets you decide call type by call type.
| Call type | Design | Why |
|---|---|---|
| Routine booking with clear rules | Agent | Saves a callback the caller may never answer |
| Urgent problem after hours | Agent: transfer | Needs a person now, not a message for the morning |
| Quote for complex work | Assistant | Needs a site visit or an estimator's judgment |
| Patient-specific medical question | Assistant, routed to staff | Health details stay with the practice |
| Table for a large party | Assistant | The host confirms; nobody promises a table |
| Complaint or billing dispute | Assistant, then transfer | A manager owns the outcome |
Assistant behavior is the right rule more often than vendors admit. Restaurants still take a large share of reservations by phone: in one Cornell study of leading independent restaurants, 49% of reservations came by telephone, and diners who had not booked online most often said they preferred to talk to someone[3]. Yet a phone line that confirms tables without checking the book creates double bookings, so a request the host confirms is the safer design; the restaurant guide to what to expect from AI reservation assistants covers this. Medical offices are similar for a different reason. Before patient information goes to an outside vendor, HIPAA requires a written business associate agreement[4], so patient-specific calls go to staff.
Where an agent earns its keep
Agents pay off on calls with clear rules and a cost to waiting: routine visits, consultations, estimate appointments, and urgent calls that need a person tonight. Trades are the obvious case. BLS notes that plumbers often work nights and weekends and are frequently on call for emergencies[5], which is exactly when nobody is free to return a message.
- Caller
- SI Employee
- System
- Your team
01Caller
Caller asks for a visit
Sink backed up again; wants someone tomorrow.
02SI Employee
Checks your rules
Service area by ZIP code, job type, and urgency against your urgent list.
03SI Employee
Offers open windows
Two windows from the schedule you connected and tested.
04SI Employee
Books and reads back
Date, window, address and callback number confirmed aloud.
05System
Summary to the office
Email with the recording, transcript and booking details.
06Your team
Dispatcher checks the day
Reviews new bookings each morning and moves any that need it.
Example flow for a plumbing company; an urgent problem would follow the on-call transfer path instead.
Notice the last step. An agent does not remove the dispatcher; it moves the dispatcher's work from returning calls to checking bookings that already exist. In a five-truck shop that is a ten-minute morning job instead of an hour of phone tag.
Before you switch on agent behavior for any call type, write down what it is allowed to do. An agent with vague rules is worse than an assistant. The minimum list:
- The job types it may book, and how long each one takes on the schedule.
- Your service area as a list of ZIP codes, and what to tell callers outside it.
- The urgent list: the problems that skip booking and go straight to a person.
- The on-call order: who gets the transfer first, who is the backup, and what the caller hears if nobody answers.
- What it must never promise: prices you have not approved, arrival times, or work your team does not do.
What the callback gap is worth
The money question is simple: how many bookings are lost between "someone will call you" and the call itself. The calculator uses the missed-calls model with the labels changed to fit, and it is set up for the fictional plumbing company in the example above.
| Input | Example |
|---|---|
| Calls asking for a visit, per week | 40 |
| Share that end as a callback promise | 30% |
| New customers among them | 50% |
| Bookings lost waiting for a callback | 40% |
| Typical job value | $500 |
| Unanswered calls a month | 52 |
| From new customers | 26 |
| Bookings that would have followed | 10 |
| Monthly value at risk | $5,196 |
How it is worked out: Calls a week × 4.33 weeks × unanswered share × new-customer share × booking rate × average value.
Assumptions
- Typical job value of $500: the American Housing Survey's median homeowner spend on plumbing fixture projects, 2021 to 2023.[6]
- Losing 40% of callback bookings is a cautious figure: most adults say they generally do not answer unknown numbers, though many check a voicemail if one is left.[2]
- The 40 calls a week, 30% ending as callback promises and 50% new customers are invented; take them from a week of your own call notes.
Example only: about $5,200 a month of plumbing work in this model depends on whether the booking happens during the call or after it.
In this example, halving the loss rate still leaves about $2,600 a month. Against that, set the extra setup work an agent needs: connecting the schedule, writing the urgent list, testing the transfer path. The five-truck plumbing company example models the after-hours half of this in more detail.
Test which one you are buying
Vendor pages use both words loosely, so test behavior rather than reading labels. Four calls, made after the vendor has loaded your real information, tell you which design you have.
A routine booking
Ask for a standard visit next week. An agent offers real times and reads the booking back; an assistant promises a callback. Then check where the booking landed.
An out-of-area caller
Give an address one town past your boundary. Either design should decline politely and say why, without booking anything.
An urgent call with nobody answering
Describe something on your urgent list at night, with the on-call phone switched off. Listen to what the caller is told and look for the urgent message afterwards.
A question it cannot answer
Ask something that is not in your information. The right answer from both designs is the same: it says so and takes a message.
Score each call on two things: did the caller leave knowing exactly what happens next, and did the office receive everything needed to act without ringing back? An agent that books but leaves the office guessing is not finished, and an assistant that takes a perfect message may be all some call types need.
Where an SI Employee sits
An SI Employee is built as an agent with assistant fallbacks. Where the owner has connected and tested a calendar, it books routine work and reads it back, and an email invite goes out if the caller gives an email address. Where there is no tested connection, it records a booking request with the preferred window for staff to confirm. Urgent calls follow the transfer path you approve; if nobody picks up, the call comes back, the caller is told plainly and an urgent message is saved. If you run your own booking or job system with an API, actions can be set up to run against it during the call, while named connectors to field-service software are still planned. At any point a person can listen in, whisper guidance the caller does not hear, or take over.
To see each of these steps in order on a single call, read inside one call. For what the labels on vendor pages usually mean, see digital employee, AI agent and SI Employee untangled; for the mechanics underneath, how an AI receptionist works; and for the trade-offs of either design, AI receptionist pros and cons.
Frequently asked questions
What is the difference between an AI agent and an AI assistant?
An assistant answers questions and records what the caller wants for a person to act on. An agent completes the step itself during the call, such as booking a visit or transferring an urgent caller, within limits you set.
Is an AI agent riskier on my phone line?
It fails differently. A misconfigured assistant gives you a slow callback; a misconfigured agent can book the wrong window or transfer to a number nobody answers. Tight rules, a read-back on every booking and a morning check of new bookings keep that risk small.
Can one phone system behave as both?
Yes, and it should. Set agent behavior for routine bookings and urgent transfers, and assistant behavior for quotes, complaints and anything that needs a person's judgment.
What happens if a booking is wrong?
Every booking is read back to the caller and lands in the office's inbox with the recording and transcript, so staff can spot and correct it the same day. Review new bookings daily for the first few weeks.
Do callers mind talking to an agent?
Most accept help but want control: in Pew's survey 73% would let AI assist at least a little, while 13% would let it assist a lot[1]. A clear greeting, a read-back and an easy way to reach a person matter more than the label.
References
Pew Research Center (2025). Americans' awareness of AI and views of use in daily life, control over it (opens in a new tab) ↑
Pew Research Center (2020). Most Americans don't answer cellphone calls from unknown numbers (opens in a new tab) ↑
Cornell University Center for Hospitality Research (2012). The Role of Multi-Restaurant Reservation Sites in Restaurant Distribution Management (Cornell Hospitality Report, Vol. 12 No. 3) (opens in a new tab) ↑
Legal Information Institute, Cornell Law School (2026). 45 CFR § 164.502 - Uses and disclosures of protected health information: General rules (opens in a new tab) ↑
U.S. Bureau of Labor Statistics (2026). Plumbers, Pipefitters, and Steamfitters: Occupational Outlook Handbook (opens in a new tab) ↑
U.S. Census Bureau and U.S. Department of Housing and Urban Development (2023). Home Improvements (2023 American Housing Survey infographic) (opens in a new tab) ↑