The full analysis

Where AI really does the work #

The jobs where products really use AI are answering and placing phone calls, replying to customers and reviews, working the inbox, reading documents such as invoices and forms, bookkeeping, and recruiting. In other jobs, like sending appointment reminders or asking customers for reviews, most products sold as “AI” turned out to be simple automation that sends a message when something happens.

Every product we examined was put to two questions: does it do the job with nobody pressing send, and is an AI doing it rather than a fixed rule? On the first question alone, most jobs look crowded. On the second, 96 of the 499 products in our first round dropped out because a fixed rule, not an AI, does the work, 58 of them in the five jobs below. Five more were held because their pages did not settle the question.

The last column includes what the later search of the whole market added, including products sold only to larger organizations. All five jobs share the same shape: the work starts with a trigger, such as a job marked complete, an invoice at thirty days, a slot opening or an appointment tomorrow, and then a message goes out. In the fixed-rule tools, nothing in that step reads, decides or writes anything, and most of it has existed for years. That is automation, and it should be priced as automation.

The rest that dropped out came from other jobs, among them bookkeeping, bill pay, dispatch, shift scheduling, inventory reordering, documents, the browser, and ad and social management. Several of those do read or write something before they act, so not every product that dropped out was a simple messaging tool.

Where AI does the work, the products cluster. Calls are the busiest part of the map: 34 receptionist products, 19 outbound callers, 17 voice platforms and 15 voice ordering products. Replying is next: 26 support resolution agents, 19 review reply products and 17 chat agents. Then recruiting at 25, staff knowledge at 21, data analysis at 16, documents at 14 and the inbox at 15. Across the map, 27 of the 35 jobs have 8 or more AI agents, 7 have fewer, and one, reordering stock, has none.

Doing versus suggesting #

A product that writes you a draft leaves the job with you. You still have to read it, correct it and send it.

A lot of the best-known tools in this market only draft, and you still press send. Advertising shows it well: most of the independent ad tools an owner is likely to be shown only recommend changes (see AI Ad Manager).

The same pattern shows up across the map, and the vendors often say so on their own pages:

• Email and text marketing. Nine of 23 vendors we opened say on their own pages that a person sends (see Campaign Agent).
• Review replies. Draft-and-approve is the default in most products and hands-free sits above the entry price. One vendor is free forever for drafts and charges $9.99 a month for hands-free. Another caps its free plan at 25 AI responses a month and puts automatic reply mode on the paid tier.
• Recruiting. Almost nobody will let the software reject a candidate. Two vendors advertise that a person always makes that call: the product works on its own up to the shortlist and never past it, and nothing is turned down automatically.
• Social media. One well-known vendor states on its own page that nothing publishes without explicit approval.
• Money. Finance vendors say it plainly. One spend platform says its agent recommends approvals and flags what needs review. Another says in plain words that its agents draft and a person approves before anything posts. A third says your team sets the rules and the agents carry them out, and a European one says you write the policy and approve when needed.
• Big marketing platforms. We checked 23 of the best-known marketing products a second time, and 20 of them keep a person approving before anything goes out. One search platform (not on this map) whose agent really does push changes into a live website states that a person confirms every change before it goes live.
• Shift scheduling and inventory. Most products build a draft schedule or a draft purchase order for a person to approve, and for reordering, none passed both tests. See Shift Scheduling Agent and AI Inventory Reordering.

Who it’s sold to #

Of the 459 listings, 317 you can sign up for yourself and 142 are sold only through a sales call or contract. The split depends heavily on the job.

Every chat agent, meeting agent and SEO agent on the map can be bought with a sign-up form. At the other end, some jobs are sold almost entirely through a sales call: all 4 waitlist fill agents, 11 of the 13 customer win-back agents and 11 of the 15 voice ordering products, with most quote follow-up agents (7 of 9) and support resolution agents (16 of 26) sold the same way. For those jobs the AI exists, but you will usually have to talk to a salesperson to buy it, and some vendors set it up for you as a service.

Most vendors do not say who their product is for. On their own pages, 369 of the 459 listings name no business size, 53 say they are built for small businesses and 37 say they are built for larger organizations. A page that names no size can still suit a small business, and a page aimed at bigger companies does not always stop you from signing up.

Some products are built for one industry, 75 in all. The most common are restaurants with 19, home services with 19, e-commerce with 12, healthcare with 7 and dental with 5. Most of them are receptionists, voice ordering, quote follow-up and customer win-back, the jobs closest to a phone call and a booking.

Every product on the map can be used by a US business: its website is in English and it can be paid for in US dollars. Most vendor pages (265 of 459) do not say where the product operates. Of those that do, 120 say the US, 44 say they operate across several countries, and 30 name a home market outside the US, led by India with 7 and Canada with 6. Fifteen products whose sites were not in English, or that could not be paid for in dollars, were left off.

What it costs #

All figures are entry prices read from vendor pages in September and October 2026. They change often.

One price pattern holds across the map, and it is about how you buy. If you can sign up yourself, you can usually see the price first. If the product is sold through a sales call, you usually cannot. A few products have a free plan but sell the agent itself through a sales call.

The $39 typical price mixes different bases. Some monthly products charge one flat fee; others charge per user, per location, per seat, per agent or per driver, so a five-person shop multiplies a per-user figure by five. Others charge by use rather than by the month: by the minute, the resolved ticket, the conversation, the message, the order or the task, and one debt collection agent charges a share of what it collects.

Five patterns are worth knowing before any demo.

The rate on the page is often not the whole cost. One voice vendor publishes the arithmetic against its rivals on its own pricing page: a competitor’s advertised rate covers only that company’s own software, and once you add the AI model, the part that turns speech into text, the part that reads the answer back aloud and the phone line itself, a working setup lands at roughly $0.13 to $0.30 a minute. Comparing $0.05 against $0.14 is not comparing like with like.

Ready-to-use products cost more per minute than the raw rate. Two voice platforms charge $0.10 a minute with everything included. One voice platform’s paid tier is $149 a month for 600 minutes, about 25 cents a minute. The difference pays for a product that works when you sign up, rather than one you have to build.

Most plans include a fixed allowance, and you pay extra past it. One receptionist plan at $24.95 includes 30 calls, then $1.50 per extra call. Another at $149 includes 200 minutes, then 50 cents a minute. One review plan at $99 includes 500 requests a month, another at $10 includes 50. A $5 SEO plan is one post a week.

Billing by credits is behind many of the billing complaints. An agent that gets stuck going round in circles keeps spending credits, and you find out afterwards (see AI Agent Builder).

The entry plan is often not the plan you need. One ad product publishes $66 a month billed yearly but recommends its $335 plan, about five times as much. One restaurant platform charges $249 a month plus 5 percent of every online order, so a month with $20,000 of online orders costs $1,249, against $499 on a flat plan with no per-order fee.

Owners raise a simpler objection on the forums: each tool fixes something too small to justify another $20, $30 or $50 a month, and the charges add up. The typical monthly plan here is $39, and more plans start above $50 (92 of 219) than between $20 and $50 (76).

Trust and control #

Owners want to know how much access to give an AI. One asked other owners whether they would let AI into their bank accounts, payroll, invoices and books, and drew the line between letting it read the numbers and letting it change them.

Vendors mostly handle this by keeping a person’s approval on any step that cannot be undone, and they sell that approval as a feature.

• In bill pay, a person approves the payment before the software sends it in 11 of the 12 (see Bill Pay Agent).
• Three voice and support vendors sell protection against the same risk, the agent saying the wrong thing: one leads with a claim that it never makes things up, one guarantees a resolution rate or refunds the month, and one sells a sealed record of what the agent said that cannot be altered afterwards.

Vendors also publish limits on their own products, and those limits are lower than the marketing suggests. In support, three separately published numbers put the share a product really does handle start to finish at roughly two thirds of contacts: one advertises resolving up to 90 percent but guarantees only 65 percent, and a customer example from another vendor shows 64 to 71 percent.

The general agents go further and publish their own failure rates. One warns that its agent can be talked into the wrong thing by instructions hidden inside a web page. Another reports cutting the success rate of those attacks from 23.6 percent to 11.2 percent when the agent is running unwatched.

Browser agent vendors add their own warnings: one page is titled a trial run and recommends avoiding financial, legal and medical sites.

One costly failure came from an ad setting doing exactly what it was designed to do. One advertiser turned on the setting that lets the ad platform widen the searches an ad can appear on, for one week on one campaign, and changed nothing else. The ads showed on 955 different searches instead of 30, and 72 percent of the spend went on searches that had nothing to do with the product. Separately, an owner says he followed the AI’s advice and rebuilt his campaigns just before the month he had been counting on, and it came in as his worst.

An ad agent can spend on more searches than you can check by hand, and you see where the money went only afterwards.

The people who build these agents say what keeps them under control is plain code, the kind that does the same thing every time, rather than better instructions to the AI. Every rule on one builder’s list of what survived real use works that way. The AI proposes what to do, and the code checks the amount, the recipient, the permissions and the current state before anything that cannot be undone happens. In another builder’s account, an agent that was blocked from deleting data found a way around the block, which is why the checks have to be built carefully.

Following the rules on contacting people is a real cost, and the cheap plans do not explain it. Vendors sell consent handling and quiet hours as features you pay for. One outbound calling vendor puts compliance review on a higher tier than its entry plan. One vendor cites 224 class actions under the US telemarketing law in a single month, September 2025, against 79 a year earlier.

Where products stop short #

Owners on the forums say AI tools often do most of a job and hand back a last piece that takes as long as doing it themselves. That last step is what this research tested: if a product leaves the final step to you, it is writing you a draft, whatever the marketing calls it.

People who build and buy these agents describe the same gap:

• A builder who sets up automations for small companies says a client’s supplier-order agent looks like it runs on its own, but it gets stuck about twice a week and he fixes it before the client notices.
• An owner was pitched a $2,000 a month chatbot. What he needed was 5,000 badly labeled supplier part numbers entered correctly into his inventory system. He estimated that paying someone a few hundred dollars a month to type them in would cost less and produce fewer mistakes.

These are individual accounts, not measurements. Three more places products stop short:

In support, products lose track of the customer. Someone who writes in by chat, then emails, then calls is treated as three different people and has to explain the problem again each time, because the product links the conversation to the chat window rather than to the customer.

Voice agents get numbers wrong, and the advertised accuracy figure hides it. A team running voice agents for real customers reports that an advertised 95 percent accuracy at understanding speech concealed the failure that matters. Roughly half the time somebody spoke a bank code into the phone, at least one character came back wrong. Numbers said out loud were the worst of it, because “double five”, “fifty five” and “five five” were each heard as something different. Taking down a phone number, an address or a card number is most of what a receptionist is for, so getting numbers wrong matters far more than the overall accuracy figure suggests.

In reporting, a simple agent lasted longer than an ambitious one. A builder made a weekly client report agent that gathered the figures, explained them and proposed what to do next. It looked good in a demonstration and did not last in real use. When its reasoning came out wrong, nobody could see at what point it had gone astray, and when it came out right, it still did not get trusted. The version still running puts the figures into the same headings every week and writes one flat sentence under each saying which way the number moved.

How fast it’s changing #

Products changed while the research was running. One inventory reordering product and one agent builder shut down while the research was running, and one more has gone since: ChatGPT agent mode, which OpenAI retired after this research and before this was published. It is on What changed. One voice platform that started out selling to small businesses now publishes contracts from $30,000 a year. One AI calling company from 2024 no longer sells a voice agent. Three products changed the name they trade under.

New companies keep arriving, but few sell to owners directly. We looked at 538 recent companies from the best-known startup accelerator and checked each one’s home page for a price or a sign-up you complete yourself. Seven passed. Most are building parts for other software companies, services where people do the work with AI, or software for a single industry sold through a sales call. Of 107 newly launched companies we read closely, about half sell only through a demo, a large contract or an invitation.

What this means for you #

Where it is ready. Answering the phone and booking the job is the busiest part of the map. Closing a support ticket is real, with a limit the vendors publish themselves at roughly two thirds of contacts. Taking a restaurant order through to the register is real. Placing outbound calls is real. Replying to reviews is real, though posting without your say-so is usually a tier up. Joining the call and writing the notes is real and often free. Reading a document and putting the details into another system is real. Reading your bills and coding them to the right account is real, though paying still waits for your approval, by design. In those jobs the vendors say plainly what the product does on its own, and many can be bought without a sales call.

Where it exists but isn’t sold to you yet. Filling a canceled appointment has 4 AI agents, and all 4 are sold to healthcare and dental organizations through a sales call. Winning back past customers and taking phone orders are mostly sold the same way. If one of these is your job, the question to ask is whether the vendor will sell to a business your size at all.

Where it is not ready. Building the staff schedule has 4 products against 12,100 searches a month. Placing the purchase order without a click has no product on the map at all: every one we checked builds a draft order for a person to approve or reorders at a level you set, and the newest companies in that space mostly sell through a sales call or to larger companies. An advertising agent can spend across hundreds of different searches in a week, more than you could check by hand. And the browser agent vendors themselves hedge on leaving an agent to work unwatched on accounts that matter (see Browser Agent).

The questions to ask before you buy anything. Every one is answerable on a demo or a pricing page.

1. Does an AI do this, or does a rule? If the answer is that when X happens it sends Y, you are buying automation, and it should not cost AI prices.
2. Does it send, or does it draft? Ask for the sentence on the page. In the inbox, 5 of the 15 products on the map send the reply themselves, and the other 10 sort, label, archive and route without being asked.
3. Does it do it on its own on this plan, or on the one above? In review replies, the setting that lets it post without you is usually a tier up from the default.
4. What does it do with the hard cases? Many products hold the difficult ones for you by default, such as low-star reviews and unresolved chats (see Review Reply Agent).
5. What is the cap, and what happens past it? Work out the bill at the volume you actually do, not at the included allowance.
6. What is the unit? Per month, per user, per location, per minute, per credit or per resolved ticket.
7. Who does the setup, and is it in the price?
8. What happens to the changes it made when I stop paying? Some tools make changes that stay only while you subscribe.
9. Who carries the legal side? On outbound calling and win-back texting, the smallest buyer often gets the calling system and is left to follow the consent rules himself.

One next step #

Start with the job you would most like off your plate, not with the product you saw advertised. Write down the one task costing you the most time or money this month, in one plain sentence: for example, missed calls, estimates that never come back, reviews that go unanswered, or bills keyed in by hand. Find that job in the 35 jobs above, read what to know before you buy, and take the nine questions into the first demo, starting with the first: does it do this itself, or write me a draft? Buy one thing, for one job, and see it work before you buy a second. If the products for your job stop short today, you will know what to look for as new ones arrive.

How we did this #

Who. Published by Iterategy. The research was carried out with AI agents that read each vendor’s own website and applied the same two tests to every product. Iterategy’s CEO reviewed and approved the method, the rules and the results. Iterategy takes no money from vendors: no product paid to be listed, and we earn nothing from the links on these pages.

When. September and October 2026. The research was done in the week of September 21, 2026, and every page was read between September 23 and October 2. Every price and claim is what the vendor published on its own page when we checked it.

What counts. Two things had to be true. The product does the job itself, with nobody pressing send, and an AI does the work rather than a fixed rule. A product that is started by an event, such as a canceled appointment, still counts if an AI does the work after that point. Tools that only suggest, tools that follow fixed rules, and services where people do the work are left out.

What we looked at. Our first round examined 499 products and kept 398. We then searched the whole market again, including products sold only to larger organizations, industry-specific products, products from outside the US and products charged by use, and relabeled every product by how it is sold. A final check then re-read every product on the map against both tests on the vendor’s own page and removed 22 whose pages did not show both, most often because a person approves or sends every action, or because the page says “automatic” without saying an AI does the work. The map holds 459 listings of 419 products from 388 companies; a product that does more than one job is listed once for each. Every label (who it is built for, how you buy it, where it operates, its industry) was read from the vendor’s own page. Where a vendor’s site blocked our research tool, the page was read in a regular web browser.

What we did not do. No product was signed up for, run or tested. Whether a product is any good is not in this research. Prices are entry prices and they move. A per-minute rate and a monthly fee are not comparable. Monthly ranges use plans quoted as a flat monthly price. Where a product appears in more than one job, its labels come from that job’s page and can differ.

Limits. Seventeen products we looked at are not on the map because we could not find a sentence on their own pages saying what their AI does by itself. We left off products a US business cannot use: sites not in English, or no way to pay in US dollars. The search was organized job by job, so a product doing a job outside these 35 jobs would be easy to miss.

Disclosure. Iterategy, who produced this report, sets AI up for businesses.

Every listing, with its labels, price and the vendor’s own sentence, is in the data download.