Before you sign anything, ask an AI receptionist vendor: does it actually go live, is a real person reviewing what it says, is it a genuine agent or a rebranded chatbot, what happens when it doesn't know an answer, how fast does it really respond, what does it cost including overages, how long until go-live in writing, what happens to the work if you cancel, and are the vendor's own reviews recent and real. Every one of those questions has a documented failure mode behind it, from MIT's and Gartner's own 2025-2026 research on AI project failure, not sales-page skepticism.
Below: the data behind each question, a side-by-side table of what a good answer sounds like versus a red flag, and real published pricing so you know what "reasonable" looks like before you get a quote. Start here: talk to a real person about what a managed build looks like.
Why does an AI receptionist purchase need a checklist at all?
Because the category is having exactly the failure rate the broader AI-agent market is having, and most of it is invisible from a sales page. MIT's Project NANDA published its State of AI in Business 2025 report after studying 300 publicly disclosed AI pilots, 150 leadership interviews, and 350 employee surveys, and found that 95% of enterprise generative AI pilots never reach production or produce a measurable financial return. Gartner's own June 2025 research goes further: it predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls as the three named reasons.
None of that means AI receptionists do not work. It means the gap between a working one and a stalled one is almost entirely in how the vendor builds, supervises, and prices it, which is exactly what a demo does not show you. The 9 questions below are built to expose that gap before you sign, not after.
How did we choose these nine questions?
We started from the two most-cited, most-methodologically-transparent 2025 studies on why AI agent projects fail, MIT NANDA's report (300 disclosed pilots) and Gartner's agentic AI research (a January 2025 poll of 3,412 webinar attendees plus analyst estimates on vendor authenticity), and worked backward to the specific question a buyer can ask that would have surfaced each failure mode in advance. We added two questions grounded in real, currently-published vendor pricing (Smith.ai, Ruby, Abby Connect, PATLive) rather than industry averages, since "what does this actually cost" is where most of the vague-answer problem hides. We dropped several commonly-repeated "questions to ask an AI vendor" listicle items, like "does it have a nice-sounding voice," because they test the demo, not the vendor.
That left nine: three about whether the project reaches production at all, one about supervision, one about escalation behavior, one about real-world speed, two about honest pricing and timeline, and one about what happens if you leave.
9 questions to ask before you buy an AI receptionist, in order
Ask these in this order. The first three disqualify a vendor before you ever discuss price.
"Does it actually go live, or does it die in the pilot phase?"
Why it's first: MIT's Project NANDA studied 300 publicly disclosed AI pilots and 150 leadership interviews for its State of AI in Business 2025 report and found that 95% of enterprise generative AI pilots never reach production or deliver a measurable P&L impact. Only 5% of custom-built pilots ever reach production at all.
What a good answer sounds like: a specific go-live date, named against a written scope, with a described process for what happens if that scope changes. A vague "we'll get you set up soon" is the same sentence every one of the 95% heard.
"Is a real person reviewing what it says, or is it running unsupervised?"
Why it matters: Gartner's June 2025 research names "inadequate risk controls" as one of three specific reasons it expects over 40% of agentic AI projects to be canceled by the end of 2027, alongside escalating costs and unclear business value.
What a good answer sounds like: a described review process, who reviews transcripts, how often, and what triggers a correction, not just the word "managed" repeated back to you. A senior engineer reviewing every line of a new deployment is a specific, checkable claim; "our AI is very advanced" is not.
"Is this actually an AI agent, or a rebranded chatbot wearing an AI label?"
Why it matters: Gartner's analysts warn about this directly, using the term "agent washing," and estimate only around 130 of the thousands of vendors marketing themselves as agentic-AI companies offer genuinely agentic capability, autonomous decision-making, escalation logic, memory across calls, rather than a re-skinned chatbot, IVR menu, or basic automation tool.
What a good answer sounds like: a plain description of what the system decides on its own versus what it always escalates. If the answer is "it follows a script," that is an IVR with better marketing, not an agent.
"What actually happens when it doesn't know the answer?"
Why it matters: a confidently wrong AI answer near a patient, client, or customer is the single fastest way this category damages a brand, and it is exactly the "unclear business value, inadequate risk controls" combination Gartner names as a cancellation cause. Neuron's own approach, disclosed here as our own method rather than an industry claim: it escalates to a person instead of guessing, and a correction becomes a reviewed update to what it's allowed to say next time.
What a good answer sounds like: a described escalation path with a real human at the end of it. "It will figure it out" is not a policy, it's a hope.
"How fast does it actually answer, day and night?"
Why it matters: Invoca's platform data, cited in our own analysis of missed-call statistics, shows 62% of calls to a small local business go unanswered by a live person, and fewer than 3% of callers routed to voicemail leave a message. "24/7 coverage" on a features page and a system that actually answers ring one, every time, are different claims. Ask for a live test call at 11pm, not a case study.
What a good answer sounds like: an offer to place a test call with you right now, on the spot, not a promise you'll "see it in the demo."
"What does it actually cost, including overages, and what happens if it doesn't work?"
Real 2026 published pricing for comparison (live-staffed and AI answering services, not custom builds): PATLive starts at $75/month pay-as-you-go with no required minutes, Abby Connect's AI Receptionist plan starts at $99/month, Ruby's starter tier is $250/month, and Smith.ai's Virtual Receptionist plans run $300 to $2,100/month depending on call volume, with per-call overage fees of $8.50 to $11.50 past your plan. A custom-built managed AI agent is typically quoted as a setup fee plus a monthly fee instead of a plan tier.
What a good answer sounds like: the overage rate stated before you ask for it, and a plain answer to "what happens if this doesn't recover more than it costs." A vendor that can't answer the second half of that sentence is pricing a hope, not a product.
"How long does it actually take to go live, and who's building it?"
Why it matters: MIT's 2025 NANDA report found that purchasing from a specialized AI vendor and building an integration partnership succeeds about 67% of the time, roughly three times more often than attempting the equivalent build with an internal team, which the same report found succeeds only about one-third as often. A vendor with a repeatable build process should be able to give you a real date.
What a good answer sounds like: a number of weeks, in writing, tied to a specific scope you both signed off on. "It depends" is true of everything and tells you nothing.
"What happens to the work if I ever cancel?"
Why it matters: this is a contract-terms question, not a marketing one, and it deserves a specific written answer rather than a verbal reassurance in a sales call. Ask what data, scripts, and configuration you keep access to, what the vendor deletes, and how much notice either side owes the other. Get it in the agreement, not the pitch.
What a good answer sounds like: a section of the contract they can point to, not a shrug and "we've never had anyone want to leave."
"Are the vendor's own reviews recent and real?"
Why it matters: BrightLocal's 2026 Local Consumer Review Survey found that 74% of consumers specifically look for reviews written within the last three months, treating an older review pile as a sign the business (or in this case, the vendor) has gone quiet. Before a first call, check the vendor's Trustpilot, G2, or Capterra profile for review dates, not just star count. A vendor with 200 reviews from 2023 and none since is telling you something a sales call won't.
All 9 questions: good answer vs. red flag
Use this table on the actual sales call. Write down the vendor's answer next to each row before you decide anything.
| Question | Good answer sounds like | Red flag sounds like |
|---|---|---|
| 1. Does it go live? | A dated go-live tied to a written scope | "We'll get you set up soon" |
| 2. Is it supervised? | A described human review process | "Our AI is very advanced" |
| 3. Real agent or chatbot? | Names what it decides vs. escalates | "It follows a script" |
| 4. What if it doesn't know? | A named escalation path to a person | "It will figure it out" |
| 5. How fast, really? | Offers a live test call now | "You'll see it in the demo" |
| 6. Real cost + overages? | States the overage rate unprompted | Price only available "on a call" |
| 7. Real timeline? | A number of weeks, in writing | "It depends" |
| 8. What if I cancel? | Points to a contract clause | "We've never had anyone leave" |
| 9. Reviews recent? | Recent, dated, verifiable reviews | Old reviews, none in months |
Key takeaway: every red flag in this table is a vague version of a question that has a specific, checkable answer. A vendor who answers all nine specifically is not automatically the right fit, but a vendor who can't is showing you the same failure mode MIT and Gartner already measured at scale. See what a specific, dated answer to all nine looks like →
You asked all nine. Now what?
Write the nine answers down and read them back without the vendor in the room. The pattern that separates the 5% of pilots MIT found actually reaching production from the 95% that don't is rarely one dramatic red flag, it's usually two or three vague answers stacked together: a soft go-live date, an unnamed review process, and a price that's "on a call." Any one of those alone might be an early-stage vendor being honest about limits. All three together is the profile of a project that stalls.
If the answers were specific and checkable, ask for them in writing before you sign, the same nine questions, as an email you can hold the vendor to later.
We'll answer all nine, in writing, before you sign anything.
Tell us your business and what's currently happening to calls you miss, and we'll walk through a specific go-live date, our escalation process, real pricing, and what happens if you ever want to leave.
See the full AI Agents service on the Neuron SEO page, or start from the Neuron HQ homepage. A real reply from the people who'll build it, usually within one business day.
Frequently asked questions
What percentage of AI agent projects actually fail?
MIT's Project NANDA, in its 2025 State of AI in Business report (300 disclosed pilots, 150 leadership interviews, 350 employee surveys), found that 95% of enterprise generative AI pilots never reach production or deliver a measurable financial return. Separately, Gartner's June 2025 research predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.
How much does an AI receptionist actually cost per month?
Published 2026 pricing for comparable answering services ranges from PATLive's pay-as-you-go floor of $75/month, to Abby Connect's AI Receptionist plan starting at $99/month, to Ruby's $250/month starter tier, to Smith.ai's $300 to $2,100/month range depending on call volume. A managed, custom-built AI agent is typically quoted as a setup fee plus a monthly fee rather than a fixed plan tier, so ask for both numbers and what happens past your included volume.
Is an AI receptionist actually supervised, or does it run unsupervised?
It depends entirely on the vendor, and this is the single most important question to ask directly. A managed, human-in-the-loop setup escalates a question it cannot confidently answer to a real person rather than guessing, and a senior engineer reviews call transcripts to correct and improve it. A fully unsupervised bot has no such review step. Gartner's 2025 research flags inadequate risk controls as one of the three named reasons agentic AI projects get canceled, so ask the vendor to describe their specific escalation and review process, not just confirm the word "managed" appears on their site.
Is a rebranded chatbot the same thing as an AI agent?
No, and Gartner's own analysts warn about this directly. Gartner estimates only around 130 of the thousands of vendors marketing themselves as "agentic AI" companies offer genuinely agentic capability, a practice its analysts call "agent washing": rebranding an existing chatbot, IVR system, or basic automation tool without adding real agentic capability like autonomous decision-making, escalation logic, or memory across calls.
How long does it actually take to launch a custom AI agent?
There is no universal number, but the data points toward buying from a specialized vendor over building it yourself. MIT's 2025 NANDA report found that purchasing from a specialized AI vendor and building an integration partnership succeeds about 67% of the time, while attempting the equivalent build internally succeeds only about one-third as often. Ask any vendor for a specific go-live date in writing, not a range, and ask what happens to your fee if that date slips.
The done-for-you, managed, human-in-the-loop version of what this checklist is testing for.
Cost, how it works, and whether it's right for your business.
Real 2026 pricing and Trustpilot ratings for the vendors named in this post.
A worked example of what a managed, 24/7 build actually looks like.