AI employee for a small business: 5 jobs (Nov 2026)

An AI employee is a sales name for an agent given one recurring job. Under the name sits a model, a harness that loops until a stop rule, tools the model can call, memory of the business, and a person who approves anything that leaves the company. The software is not on payroll, and a wrong send is still yours.

That is the whole object. The pages that currently answer “what is an AI employee” talk about ownership and a manager. They do not print a November 2026 price, a credit band, or an MCP gate. This page does. The label, with Claude Cowork and ChatGPT Work, is already What is an AI employee?. The small-business question left over is which of five jobs you can hand off, and where the handoff breaks.

On November 12, 2026, 12 of 15 pricing pages for these jobs print a dollar. The median of the seven fixed subscription cards is $33 a month. Three pages, Artisan, Synthflow, and Truewind, print no dollar at all. The row-by-row census is 12 of 15 AI employee products print a price. A shared ChatGPT or Claude seat is a different purchase, on Best AI agent for small business.

The label is the pitch. The harness is what you are buying.

A model chooses the next step. It only sees what the harness puts in the context window: the thread, the retrieved help article, the tool result. A help center you did not connect is not knowledge. The model cannot use a fact it was not given.

The harness is the program around that model. It offers tools, keeps the memory, enforces the permissions, and decides when to stop. Tool use is the call itself: send the email, read the ledger, book the slot. MCP is one way to attach those tools. Lindy’s pricing page lists it. Kick’s pricing page lists it and then requires a confirmation before the write. Goodcall counts an MCP server as one external tool against the plan limit. The protocol is not the employee. It is a socket.

Memory is the part vendors name and rarely price. Sintra calls it Brain AI, shared across the helpers. Lindy’s page says the memory is yours to edit. Memory is why Tuesday’s agent can sound like it “knows” the shop. It is also why a stale note keeps getting reused. Edit it, or the next draft inherits the error.

The stop is a human in the loop, usually an approval gate. Lindy’s pricing page states it cleanly: anything with outside impact waits for approval, including a send, a ticket update, a post, or a published doc. Read-only lookups do not. That sentence is the difference between a draft and a message your customer received. Most of the other pricing pages do not say it. You still need it.

An eval is how you find out the gate is doing work. Take twenty of your own tickets, write the pass rule, and run them before the agent is allowed to send. The method is How do you evaluate an AI agent harness?. A vendor demo is not that test.

Five jobs. One of them is not even sold on its own.

Scheduling is the job the ads imply and the price pages skip. Lindy lists inbox and calendar together. Goodcall books during the call. Marblism’s receptionist is described as booking. None of the fifteen sells the calendar as its own meter. A double-booked slot is a customer you have to call back. Treat the calendar as a tool the phone or inbox agent calls, and keep a person on the exceptions. The separate shelf is Best AI agent for scheduling.

Inbox: the draft is cheap until the credit band shows up

Sintra sells twelve named helpers on one card. The 12-month card is $15.60 a month, beside a struck $52. The 1-month card is $48.50. The FAQ on the same page says pricing starts from $48.50 a month. Every card includes 250 credits, and the page does not say how many credits an email costs. The helpers also print $39 a month and are labeled included with Sintra X. Do not add $39 times twelve unless a checkout charges it. The fetch on November 12 does not contain an approval rule. Put one in front of send anyway.

Marblism is $24 a month on the yearly card, or $44 month to month, for 50 metered hours and seven roles, including the receptionist. The pricing page says hours reset and that you can buy top-ups. It does not print minutes per email. The November 11 guide did that arithmetic from a help article. This page does not repeat it. If the only number you have is “50 hours,” you do not yet know whether your inbox fits.

Lindy is the inbox product that tells you both the meter and the stop. $29.99 per active user buys 3,000 credits. Everyday asks, and the page includes “drafting a reply,” cost 2 to 250 credits. At 2 credits the seat holds 1,500 drafts. At 250 credits it holds 12. Same $29.99. Top-ups are $10 per 1,000 credits, so a credit is a cent once you are paying extra, and a 250-credit draft is $2.50 on that rate. Monthly credits do not roll over. The pool pauses when it is empty. And the send waits. That last sentence is why Lindy is the one to copy even if you buy something else.

Support: read whether the handoff is the bill

A 6-minute ticket at the BLS May 2025 median for customer service representatives, $21.53 an hour, is $2.153 of wages. Fin is $0.99 an outcome. Zendesk is $1.50 if you commit, or $2.00 pay-as-you-go. HubSpot’s customer-agent page, read the same morning and not in the fifteen, is $0.50 a resolution, and only if you already pay for Professional or Enterprise.

The definitions are the product.

Fin charges once per conversation. The page counts an outcome when the customer confirms, when the customer goes quiet, or when Fin completes a Procedure, including a handoff. Asking for a person can be the thing you pay for. The Essential seat beside it prints $29 and $19, with “New customers save 35%.” Use the number the checkout shows.

Zendesk charges the automated resolution only when no human takes the ticket. The comparison prints 5, 5, and 10 included resolutions per agent per month under Support Team ($19 a year, paid yearly), Suite Team ($55), and Suite Professional ($115). A handoff avoids the resolution charge and lands on the human seat. Eight hundred 6-minute tickets are 80 hours, $1,722.40 of wages. Eight hundred pay-as-you-go resolutions are $1,600, which is 93% of that wage, before the seat. At a full part-time queue the “cheap resolution” is almost the person.

Tidio’s Lyro starts at $32.50 a month from 50 AI conversations. A Lyro conversation is one with at least one AI reply. Human handoff is a feature, and a human reply is a different meter. “Starts at” is not a per-ticket rate. If you use all 50, the floor is $0.65 a conversation. A quieter month costs more per conversation, because the $32.50 does not shrink.

The failure mode is a confident answer that is wrong, followed by a bill for the apology. Ground the agent in the help center you actually maintain. Pull a human in on refunds, anger, and anything you would not let a new hire say on day one. Then check, on that vendor’s page, whether the pull is free.

Outbound: $1.50 a person, or $36,000 before the first one

AiSDR prints $1.50 per person prospected. Research, lists, and writing are inside that rate. There is no monthly fee. Pay-as-you-go bills daily. One FAQ says campaigns start at $7.50, which is five people at the rate. A later sentence says campaigns start from $3. Budget $1.50 and treat $7.50 as the floor the FAQ will repeat. Do not budget $3 until a checkout shows it.

11x Growth starts at $36,000 a year, billed annually, for up to five end users and 2,000 new prospects a month. The page charges per lead, not per send, so three touches and thirty cost the same. Divide the year by the full cap and you also get $1.50 a prospect. Divide it by 100 people and you get $30 a person, because the floor does not shrink. One hundred people on AiSDR is $150. The year is twenty times the meter.

On November 11 the same 11x URL showed a $3,750 monthly card beside an FAQ of $36,000. The November 12 HTML prints $36,000 and does not print $3,750. The older figure is a record of that day, in the cost guide. It is not the current card.

Artisan describes the guardrail you want, “approvals, escalation rules, and when a human steps in,” and prints no dollar. A control you cannot cost is not a plan. AiSDR’s approval sentence is not on the $1.50 line either. It is on a $149 managed-campaign add-on. The default list says replies go out in 5–10 minutes or in co-pilot mode. Pick co-pilot until you have read a week of the mail the agent wanted to send. A wrong cold email is the company’s email. The model will not be the one who answers the angry reply.

11x’s page says a fully loaded SDR costs $8,000 to $12,000 a month. That is 11x’s sentence, not a survey. The BLS figure in the customer-service handbook’s related-occupations table is $76,460 a year for wholesale and manufacturing sales representatives. That is a wage for a different job title. It is not a reason to sign a $36,000 year you will not fill.

The phone: the minute is not where the savings are

Goodcall Starter is $79 a month for 100 unique customers, then 79 cents each. A unique customer is one phone number, once a month. Five calls from the same person count as one. The FAQ says the plan is month to month, there is no annual plan, and you should set a human escalation contact. You choose who receives the transfer. The page does not meter minutes. Its own “about 13 cents a minute” uses “about 150 calls” and a four-minute assumption. The invoice unit is $0.79 per customer.

Eighty hours of a receptionist at the May 2025 median, $18.27 an hour, is $1,461.60 in wages before benefits. Goodcall’s $79 is 5.4% of that line. The comparison holds when the alternative is a person waiting for the phone to ring. It does not hold when the agent books the wrong job and someone spends the afternoon undoing it.

Phonely Starter is $50 a month, or $33 billed annually. The card says 200 minutes. A table on the same page says 250 minutes included. Overage is $0.25 a minute. Divide $50 by the card’s 200 minutes and the included minute is also $0.25. One minute of the receptionist wage is $0.3045. The AI minute is 82% of the wage minute. You are not buying a cheaper minute. You are buying the hours the chair is empty, and the concurrency the card marks unlimited.

Phonely also says post-transfer minutes can still bill if its connection stays up after the caller reaches a person. The transfer you wanted, the human, can be the meter. SIP REFER is the exception the page names. Ask which transfer you have before you forward the business line.

Synthflow’s pricing page says there is no per-minute rate and no trial. The product is an annual enterprise agreement. Escalation and handoffs are in the scope, without a price. If you need a card you can cancel, that page is the wrong door. Developer voice stacks, where you bring the model and pay the minute yourself, are a different bill, on what a voice AI agent costs per call. HubSpot’s customer-agent page prices voice at $0.50 a minute, above the wage minute, and only on Professional or Enterprise.

The books: categorization is the cheap part

Digits Essentials is $65 a month for 24/7 AI bookkeeping and reconciliation, invoicing, bill pay, and Ask Digits. API and MCP are on every plan. Bill pay is “drag, drop, approve, pay.” The page says verified financials, the version a lender or a tax authority will trust, may still need a professional. That sentence is the product limit. The $65 is the doer. The signature is a person.

Kick Basic prints $40 a month, and $480 billed annually, which is the same rate. The page’s own line is “doer to reviewer.” MCP can read and write the books, and approvals preview the change before it posts. That is the right shape for a ledger: the agent proposes, the human commits. Free is $0 until 250 transactions a year. Do not mistake the free cap for the $40 plan.

Eighty hours of a bookkeeping clerk at $24.36 is $1,948.80. Digits at $65 is 3.3% of that wage. Kick at $40 is 2.1%. The percentage is real and easy to misuse. A clerk also closes the month, answers the owner, and notices that a transfer was a loan. The agent categorizes. Pilot Essentials, read the same morning and outside the fifteen, is $99 a month up to $100,000 of expenses and says the AI categorizes, reconciles, and closes monthly. Its Core price did not render. A $0 on that card is a placeholder, not a discount.

Truewind describes reviewer routing and an audit trail, and it does not print a price. If the vendor will not say the dollar, you do not have a budget. You have a meeting.

What this means for a shop that is about to connect the inbox

Buy one job you can count. Count the emails, the tickets, the prospects, the unique callers, or the transactions. If you cannot count it, you cannot tell a credit band from a bargain.

Use the fixed card when the job is the phone or the books and a person still reviews the result. Goodcall at $79 and Digits at $65 and Kick at $40 are that shape. Use the per-person meter when the job is outbound, and do not sign $36,000 to find out whether you like the drafts. Use a resolution price only after you have read the handoff sentence out loud.

Then set four controls before the first real send:

1. An approval gate on anything that leaves. Lindy already has one. Make the others match it. A draft in a folder is a safe default. A send is a different grant.
2. A tool list you can point at. MCP is a socket, not a personality. Kick confirms the write. Goodcall counts the server against a limit. A tool you did not connect should be a tool the agent does not have.
3. An eval on your own work. Twenty past tickets, a written pass rule, run before go-live. A demo on the vendor’s data will pass.
4. A handoff you have priced. Fin can charge it. Zendesk uses it to avoid the resolution charge. HubSpot’s $0.50 resolution is a conversation that was not handed to a person for 72 hours, or a lead the agent qualified. Those are three different bills.

The AI agent ROI calculator will take your task count, your minutes, and your wage. The AI agent cost calculator prices model tokens, which is a different meter from every card on this page. If the vendor did not print the unit, neither calculator can invent it. Walk away, or make them print it.

Frequently asked questions

What is an AI employee?

An agent with a job title. The working parts are a model, a harness, tools, memory, and an approval gate. Digital worker is the same label. The longer definition, including Cowork and ChatGPT Work, is What is an AI employee?. This page is the small-business job list.

What is AI staff?

Software you assign a role to, sold as if it were a hire. The invoice is still a seat, a credit, an hour, a resolution, a minute, or a person contacted. Twelve of the fifteen pages checked November 12 print that invoice. Three do not.

Can it replace a part-time receptionist or bookkeeper?

It can take the empty hours and the repetitive pass. It does not take the redo, the door, or the signature. Goodcall at $79 is 5.4% of 80 hours of median receptionist wages. A Phonely minute is 82% of a wage minute. Digits at $65 is 3.3% of 80 hours of median clerk wages, and the page still points you at a professional for verified financials.

What should a small business not hand over?

The send, the payment, the ledger write, and the cold email, until a person has approved a sample and the meter for a mistake is known. Scheduling, as its own employee, is not even on the price list. Leave the exceptions with a person.