The direct answer: for most small businesses this is not a choice between a person and software. It is a question of which steps of follow-up need a person's judgement, and what the rest is costing you in someone's time today.
Follow-up is not one task. It is a sequence: an inquiry arrives, someone owns it, a next action is promised, that action actually happens on time, the reply is handled, exceptions are routed to a person, and the outcome is recorded. Some of those steps are genuinely mechanical. Some are not. Deciding where the line falls for your business is the whole decision, and this article gives you a workflow and a worksheet to make it rather than a rule of thumb.
This article gives you the real cost of each side, states plainly what automation does badly, and ends with a worksheet that can legitimately tell you to stay in-house. We sell automation. The worksheet still has an in-house answer in it, because sometimes that is the right answer.
What a person actually costs
Start with published figures rather than estimates, because this is the number people guess at most often.
The US Bureau of Labor Statistics reports the median pay for receptionists at $38,010 per year, or $18.27 per hour (May 2025), across about 947,500 jobs. For customer service representatives the median is $44,770 per year, or $21.53 per hour (2025), across about 2,666,000 jobs.
Those are wages. They are not what the role costs you.
BLS also publishes what employers actually spend per hour worked. For private industry workers in June 2026, total compensation was $46.89 per hour worked, of which wages and salaries were $32.82 — 70.0 percent of employer costs — and benefits were $14.07, the remaining 30.0 percent.
That ratio is the part owners forget. Wages are roughly 70 percent of what an employee costs.
Hypothetical calculation — clearly labelled, not an observed cost. Assumptions: the BLS median receptionist wage of $38,010/year; the BLS private-industry ratio of wages at 70.0 percent of total employer compensation cost; no overtime; no recruiting or training cost included; one full-time person. Estimated total employer cost: $38,010 ÷ 0.70 ≈ $54,300 per year, or roughly $4,500 per month. Limitations, stated: the 70/30 ratio is an all-private-industry average, not a receptionist-specific figure, and benefit structures vary enormously between small employers. Your actual number could be materially lower if you offer few benefits, or higher with health coverage. Use your own payroll figures if you have them; this is a starting frame, not a quote.
Two things that number does not include, and which are real:
- Coverage gaps. One person covers roughly 40 hours. A business taking calls evenings and weekends is not covered by one hire, and a second hire is not half the cost of the first.
- Recruiting and training. Every departure restarts it. I am deliberately not attaching a figure to this, because I could not find a primary source for small-business turnover cost that I would be willing to publish. It is real; it is unquantified here on purpose.
One more piece of context from the same source: BLS projects receptionist employment to decline 2 percent between 2025 and 2035 (about −16,000 jobs), and customer service representative employment to decline 5 percent (about −141,800). That is a projection about an occupation, not a prediction about your business, and it is not an argument that you should automate. It is context for the direction of the market you are hiring into.
What automation actually costs
Rather than repeat it here, the full cost structure — the four billing models, what setup fees cover, and the per-call add-ons that do not appear on pricing pages — is in our companion guide on what an AI receptionist actually costs.
What matters more for this comparison is scope, and here our own pricing page is the clearest example of why a single monthly number is not comparable to a salary.
- AI Receptionist — $200/month, no setup fee. Scope, in the page's own words: "Call answering, appointments, and text confirmations." That is inbound coverage. It is not a follow-up programme.
- AI Automation — $999 setup plus $799/month. Scope: "Receptionist plus agreed tasks such as follow-up, admin, and reporting." Follow-up sits here, and only for the tasks actually agreed.
- Both include a scheduled monthly review — the page is explicit that this is "a scheduled review, not continuous human monitoring."
So the honest comparison is not "$200 versus $4,500." A subscription covers agreed tasks, not every duty of an employee. Whether it saves you anything depends on three things a pricing page cannot tell you:
- What scope you actually need. If you need follow-up, the receptionist tier is not the comparison — the automation tier is.
- Whose marginal time it replaces. If follow-up is currently squeezed into an existing person's day, you are not removing a salary. You are freeing hours, and the value is whatever those hours are worth on better work.
- Your own implementation and oversight time. Configuration, deciding the escalation rules, and reviewing what happened are real hours, mostly yours, and they do not disappear after setup.
Work out your own number against the scope you actually need. Anyone presenting the gap between a subscription and a salary as your saving is comparing two different things.
What automation does badly — stated plainly
If you only read one section, read this one. A vendor who will not tell you this is not being straight with you.
The genuinely unusual call. A customer who is upset, confused, or describing a situation that does not fit any of your normal categories. Automation handles the expected well and the unexpected poorly. That is not a temporary limitation to be patched next quarter; it is the shape of the tool.
Judgement calls with money or reputation attached. Whether to waive a fee, whether to squeeze someone in, whether this complaint is about to become a public one. These should reach a person, and your configuration should make sure they do.
Anything requiring real context about a specific customer's history. Unless it is connected to a system that holds that history, an assistant starts every conversation new. Sometimes that is fine. For a long-standing customer with an ongoing problem, it is not.
Being wrong confidently. An automated system that misunderstands a question can still produce a fluent, confident answer, because fluency and accuracy are separate things in these systems. Test this before you buy rather than taking anyone's word for it: ask the assistant something your business does not do, and something phrased ambiguously, and see whether it says it does not know or invents an answer. That single test tells you more than any comparison table, and it is why the escalation path below matters.
Silent failure. If nobody checks, an assistant can mishandle calls for weeks without anyone noticing, because there is no frustrated employee to complain. Whatever you choose, someone has to review what actually happened on a regular schedule.
What a person does better
Anything ambiguous. A person hears hesitation, changes approach, and asks a question that was not in any script.
Relationship. For high-value or repeat customers, being known matters and is difficult to fake.
Catching the thing nobody anticipated. New problems arrive before anyone has configured a response to them. A person can notice that and adapt; an automated sequence only does so if someone built that branch into it. Neither is guaranteed — an inattentive person misses it too.
Accountability you can talk to. When something goes wrong, there is someone who can explain what happened and change how it is handled tomorrow.
The follow-up workflow, step by step
This is a proposed, configurable workflow, not a description of any product. Use it to decide which steps you want a person on. Nothing here is automatic in any plan unless you have agreed that specific task with a provider.
1. Capture the inquiry. Every call, form, text and email lands in one place with the source, the time, and what they asked for. If inquiries live in three inboxes and a notepad, nothing downstream works — and this step alone often fixes more than automation does.
2. Assign an owner immediately. One named person, not "the team". Unowned inquiries are the ones that go quiet.
3. Record the promised next action and when. Not "follow up soon" — the specific action and a date. This is the field the whole workflow turns on, and it is usually the one missing.
4. Make the agreed follow-up, through an approved channel. Whatever you promised, on the day you promised, by a channel the customer agreed to. Decide in advance which channels you use and which you do not.
5. Handle the reply — including "stop". A reply routes back to the owner. A request to stop contacting them is honoured immediately and recorded, and no further follow-up goes out. Build the stop path before the sending path; it is easier to get right at the start than to retrofit.
6. Route exceptions to a person. Anything with a complaint, a price exception, an existing job gone wrong, or an explicit request for a human. Write the exception list down; if it lives only in someone's head, it is not a rule.
7. Verify the outcome. Did the promised action actually happen? Did it end in a booking, a no, or silence? Silence is an outcome and should be recorded as one, because a workflow that never records failures cannot be improved.
Which steps are safe to automate
| Step | Automate? | Why |
|---|---|---|
| 1 Capture | Usually yes | Mechanical, and often the first gap worth closing where missed enquiries are the observed problem — verify that against your own numbers before assuming it |
| 2 Assign owner | Usually yes | Rule-based, as long as the rules are written down |
| 3 Record next action | Yes, with a caveat | The system records it; deciding what to promise is judgement |
| 4 Send the follow-up | Depends | Fine for routine confirmations and reminders; riskier where the message needs to reflect a specific conversation |
| 5 Handle replies / stop | Partly | The stop path must be reliable; genuine replies should reach a person quickly |
| 6 Route exceptions | Rule-based routing yes, handling no | The routing is mechanical; the conversation is not |
| 7 Verify outcomes | Yes for recording | Someone still has to read it and act on the pattern |
The follow-up worksheet
Answer these about your own business, in order. They map to the steps above.
- Where do inquiries currently land, and how many separate places is that?
- Who owns an inquiry today? Can you name the person for the last five?
- Do you record a specific promised next action and date, or a general intention?
- Of last week's inquiries, how many got the follow-up that was promised, on time?
- Which channels has the customer actually agreed to?
- What is your stop process today if someone asks not to be contacted?
- Write your exception list now, in one sentence each. How many are there?
- Of your last twenty inquiries, how many outcomes can you state? If you cannot, start here.
If you cannot answer 3, 4 and 8, that is your finding, and it is not an automation problem. A workflow nobody can describe will not be fixed by automating it — you will get the same gaps, faster. Fix the record-keeping first; then decide what to hand over.
The hybrid, and the escalation rule
Most small businesses that get this right do not choose a side. They automate the predictable and route the rest, which means the important decision is not whether to automate but where the line sits.
A workable default:
- Automate: answering at all hours, the repeated factual questions (hours, service area, rough pricing, directions), capturing who called and why, and booking straightforward appointments.
- Route to a person: anything involving a complaint, a price exception, an existing job that has gone wrong, or a caller who asks for a human. That last one is not negotiable — if someone asks for a person, the system should stop trying to help and hand over.
- Review weekly at first: read what actually happened, and check it against what you assumed was "routine" when you set the rules. Treat the first few weeks as a test of that assumption rather than as a settled configuration.
Three questions to settle before you sign anything, because they decide whether the hybrid works or quietly fails:
- What happens to a call the assistant cannot resolve? Message, live transfer, immediate text to you, or silent end? Get the specific behaviour, not a reassurance.
- Is escalation billed as an extra event? Ask the vendor directly and get the answer in writing — on per-call or per-minute pricing, how a transfer or handoff is counted varies by vendor and is not something you should assume either way.
- How do you find out when it got something wrong? If the answer is "you'll notice," that is not an answer.
The decision worksheet
Answer honestly. Scoring is at the end, and staying in-house is a real outcome.
Volume and coverage
- How many inbound calls and messages do you get in a normal week? In your busiest week?
- How many arrive outside your working hours?
- What share are the same handful of questions, versus genuinely different conversations?
What is actually being lost
- How many inquiries go unanswered in a normal week? If you do not know, that is your first finding — count for two weeks before deciding anything.
- What is one new customer worth to you, roughly?
- Is the problem missed contacts, slow follow-up, or disorganised follow-up? These have different fixes, and only the first two are really about availability.
The people side
- Is someone doing this today alongside another job? What is it displacing?
- If that person left tomorrow, what happens?
- Could you afford the estimated ~$4,500/month for a dedicated full-time person, at your margin?
Fit
- What share of your calls need judgement rather than information?
- Do your customers expect a relationship, or do they mostly want a fast answer?
- Will someone review what the system actually did each week? Name them.
Scoring
Stay in-house if: your volume is low and inside working hours; most conversations need judgement; you already have someone with genuine capacity; or nobody will review an automated system. That last one is decisive on its own — unmonitored automation is worse than a person who occasionally misses a call, because the failures are invisible.
Automate the front door if: contacts arrive outside your hours; a large share are the same few questions; the person doing it now is being pulled off higher-value work; or you are losing inquiries you can count.
Hybrid — a common arrangement, though I have no data on how often it is chosen — if: volume is real but conversations vary; you can name who reviews it; and you can state the escalation rule in one sentence.
If you are genuinely unsure, the cheapest next step is not buying anything. Count your inquiries for two weeks and note how many arrive after hours. That single number resolves this decision more often than any comparison table, including this one.
What I could not source, and therefore did not claim
You will encounter confident statistics in this category — that responding within five minutes multiplies your odds of qualifying a lead by some large factor, or that a specific percentage of leads are never contacted at all. These numbers circulate widely.
I could not trace them to primary evidence I was willing to publish, so this article does not use them. They may well be directionally true. But an article arguing that you should demand real numbers from vendors cannot itself rest on statistics it has not verified.
What is sourced here: the wage figures, the employer-cost ratio, and the employment projections, all from BLS, all cited below with their reference dates. The cost estimate is explicitly labelled hypothetical with its assumptions and limitations shown. Everything else is reasoning you can check against your own business, which is the only evidence that actually decides this.
Where AlgoEthos fits
We sell the automated side, and it is worth being precise about which part, because the scope distinction in this article applies to us too:
Current plans are on our pricing page.
- AI Receptionist, $200/month no setup fee, covers call answering, appointments and text confirmations. That is inbound coverage, not a follow-up programme — it is not the plan to compare against the workflow above.
- AI Automation, $999 setup plus $799/month, is receptionist plus agreed tasks such as follow-up, admin and reporting. Follow-up work sits here, and only for the tasks actually agreed.
- Both include a scheduled monthly review. That is a scheduled review, not continuous human monitoring — which matters, because step 7 of the workflow still needs someone reading outcomes.
We also told you above, in detail, what our category does badly. If the worksheet points you at "stay in-house," take that answer — you will have spent nothing and learned something true about your own business. And if you cannot answer worksheet questions 3, 4 and 8, fix your record-keeping before buying anything from anyone, including us.
Next step: book a free walkthrough. We will run the worksheet against your real numbers, and if the answer is that you do not need us yet, we will say so.
Related reading on this site
- What an AI Receptionist Actually Costs — the full cost structure
- AI Receptionist — the service
- AI Automation for Small Business — beyond the phone
- AI for Small Business — the overview
- Pricing · Contact
Sources
All checked 12 September 2026.
- BLS Occupational Outlook Handbook — Receptionists — median pay $38,010/year, $18.27/hour (May 2025); employment 947,500; projected −2% 2025–35 (−16,000).
- BLS Occupational Outlook Handbook — Customer Service Representatives — median pay $44,770/year, $21.53/hour (2025); employment 2,666,000; projected −5% 2025–35 (−141,800).
- BLS Employer Costs for Employee Compensation — private industry, June 2026: total compensation $46.89/hour worked; wages and salaries $32.82 (70.0%); benefits $14.07 (30.0%).