Can an AI Sales Agent Close a Deal on the Phone?

A few months ago I was choosing a vendor to handle my company's compliance filings. I collected quotes from five firms. The prices were close enough that price was not going to decide it. One salesperson did.

He called me three times in one day. Every time I raised an objection, he had an answer ready, and when he did not, he said he would check with his manager and called back with one. He offered me what he called a "special friends and family discount" and priority handling. Between calls he was clearly going back and forth with his seniors. At the end of the third call he issued an invoice while I was still on the phone, and I paid it.

I was testing him the whole time. Not his pitch, his integrity: whether the company behind him was credible, whether he would still pick up after I paid. I closed with him because I felt I could trust that person.

Every AI sales agent vendor now promises some version of this. So I want to answer the question properly: could an AI have closed me? Not "can AI make sales calls", because it obviously can. Could it have done what he did, end to end?

Here's what matters most:

• That sale was nine separate jobs, not one conversation. AI is already good at some of them and nowhere near others.
• The gap is not mostly intelligence. Handling an objection is the part AI does well. The gaps are capability (doing things in your systems), reliability (doing them right every time) and trust.
• An AI must never invent a discount. Offers need hard limits, set by a person, with anything beyond them going to a manager.
• Persistence and paperwork are where an AI beats most salespeople: calling back on time, remembering every detail, issuing the invoice, sending the payment link.
• The best design today is an AI that does the chasing and preparation, with a named person who closes.

The sale, broken into the jobs he actually did

When people ask whether an AI sales agent can sell, they picture one long conversation. What I experienced was a process. Here are the jobs inside it:

1. Reaching me. He got me on the phone, and kept getting me on the phone.
2. Remembering me. On the third call he still knew my objections from the first.
3. Handling objections live. Whatever I threw at him, answered in the moment.
4. Asking upstairs. When he did not have authority, he went to his manager and came back.
5. Making an offer. The discount and the priority handling.
6. Following up on his schedule and mine. Three calls, at times that worked.
7. Proving credibility. Answering in a way that made the company seem real and accountable.
8. Doing the paperwork. Issuing a correct invoice while I was still on the call.
9. Taking payment. Getting me to pay before the call ended.

Laid out like that, "can an AI do sales" stops being a yes-or-no question. Each job needs something different from the machine, and each fails differently.

Intelligence, capability, reliability, trust: four different gaps

I find it useful to sort the jobs by what they actually demand.

Intelligence: working out what to say. This is the part people worry about most and the part AI handles best. In a preregistered study published in Nature Human Behaviour in 2025, GPT-4 debated people online. When it had a few basic facts about its opponent, it was more persuasive than the human in 64.4 percent of the pairs where one side was more persuasive. Without those facts, the difference was not significant. That was text debate, not phone sales, so do not stretch it. But it tells you that answering "why should I pay more than the other quote" is not where an AI sales agent breaks.

Capability: doing things. Remembering yesterday's call, messaging a manager and waiting for the reply, scheduling a callback that actually happens, creating an invoice. None of this is about how clever the model is. It is about whether someone has connected the AI to the systems where those things live. This is engineering work, and it is solvable.

Reliability: doing it right every time. This is the real gap today, and it is the one vendors talk about least. Salesforce's own research team built a test of business tasks inside a customer records system, called CRMArena-Pro. Leading AI agents succeeded about 58 percent of the time on single-step tasks, and about 35 percent when they had to ask follow-up questions to get the information they needed. In Sierra's τ-Voice benchmark, voice agents doing customer-service tasks did far worse than text agents doing the same tasks, especially with background noise and accents. A salesperson who forgets your objection on one call in three would not last a month.

Trust: whether you will accept it. This is the hardest to engineer. In a field experiment published in Marketing Science in 2019, a financial services firm made outbound sales calls to over 6,000 customers using either people or a chatbot. When customers did not know, the bot sold about as well as experienced staff and four times better than new ones. When customers were told at the start that it was a bot, purchases fell by more than 79 percent. People were curt and assumed it knew less. That is one study, one firm, older voice technology, and I would expect better voices to narrow the penalty. It is still the only hard number I know of, and it lines up with my own experience. I was not buying a filing service. I was buying a person's promise that the filing service would be there.

Map the nine jobs onto those four gaps and you get a clear picture. Objection handling is mostly an intelligence job, and AI is good at it. Memory, asking upstairs, the invoice and payment are capability jobs, and they need building. Getting all of them right on every call is a reliability problem. Proving credibility is a trust problem, and today a person still wins it.

The discount problem is about authority, not persuasion

The moment in my story I would least trust an AI with is the discount.

A salesperson's discount is not improvised. Someone gave him authority up to a limit, and above that limit he had to ask. When he went "back and forth with his seniors", that was an approval process running in real time. An AI sales agent needs the same thing, made explicit: a list of offers it is allowed to make, the conditions for each, and a hard stop that sends anything else to a named person.

The cost of getting this wrong is not theoretical. In 2024 a Canadian tribunal held Air Canada liable after its website chatbot gave a customer wrong information about a bereavement fare. The airline argued the chatbot was responsible for its own words. The tribunal disagreed. In late 2023 a California car dealer's chatbot was talked into agreeing to sell a $76,000 SUV for $1. That one was a prank and no sale happened, but it shows what an AI with no hard limits will agree to when someone pushes.

The model that works looks more like Walmart's. Walmart used Pactum, an automated negotiation system, to agree terms with smaller suppliers. It negotiated inside limits Walmart set in advance. According to figures reported by PYMNTS and the company, it closed deals with 64 percent of suppliers in its first pilot of 89 suppliers, and later with 68 percent at an average saving of 3 percent. It ran by chat, not voice, on small contracts. The point is not the numbers. The point is the design: the machine negotiates freely inside a box a person drew.

One more thing about the "friends and family discount". An AI can say that line. If the discount genuinely exists for a reason, fine. If it is an invented scarcity trick, then it is a lie, and when an AI tells it, the lie is your company's on the record, in a transcript, at scale. I would not let an AI use pressure tactics a regulator could read back to you.

Persistence and paperwork are where AI beats people

Now the encouraging part. The jobs my salesperson did best, and that most salespeople do worst, are the ones an AI does well when it is built properly.

Calling back on time, every time. Most follow-up dies because people get busy. An AI does not. The engineering requirement is that the callback survives anything: a restart, a crash, a weekend. Quo, a business phone company, describes building its voice agent so that every call is a long-running job that can pick up exactly where it left off. That is the difference between "I'll call you back in ten minutes" as a promise and as a guarantee.

Remembering everything. Vendors such as Retell now keep a short running brief per contact and give it to the agent at the start of the next call. Done well, the third call starts where the second ended, with the objections already on the table.

Issuing the invoice. Creating an invoice in your billing system during the call is a straightforward connection. The care needed is on the details: company names and email addresses are exactly what voice agents mishear. In the τ-Voice research, agents often failed to capture names and emails even when callers spelled them out. The safe design reads the details back, or texts the customer a short form to confirm them.

Taking payment. Do not let the AI hear a card number. Payment specialists such as Paytia describe intercepting the keypad tones so that neither the agent nor the recording ever receives the digits, which keeps the call out of the scope of card security rules. The simpler option for most businesses is to text a payment link while the customer is still on the line.

Checking its own claims. In τ-Voice, one agent told a caller it had updated their address without ever making the change. In a sale, "your invoice is on its way" when nothing was sent is fatal. Any system I build lets the AI claim only what its own action record confirms.

How I would design an AI sales agent today

If I rebuilt my compliance vendor's sales process with what exists today, it would look like this.

The AI answers every enquiry within a minute, day or night, qualifies it, and books a call with a named person. It handles the first round of common objections from approved answers. It follows up at the times the prospect chose, never forgets a detail, and logs everything. When the prospect asks for a better price, it offers what it is allowed to, and anything beyond that becomes a request to a manager, who approves or declines in a message, and the AI calls back with the answer. When the prospect is ready, it issues the invoice and texts the payment link. The named person takes the moments where trust is being tested, which in my case was the whole relationship.

That design keeps the AI alongside the sales team rather than in place of it. The salesperson stops spending their day chasing and typing, and spends it on the conversations that decide the deal.

Two boundaries matter. First, this works for people who asked to hear from you. AI cold calling to strangers has the weakest return and the biggest legal exposure; in the US, the FCC treats an AI voice as an artificial voice under robocall rules, which I cover in AI outbound calling: is it legal and what should you automate. Second, the gap closes fastest where price and speed decide the sale, and slowest where, as in my case, five quotes are similar and trust decides.

You can hear the conversation side for yourself below. This demo agent qualifies a heating and cooling enquiry and agrees next steps. Push it on price.

If you are deciding what an AI should and should not take off your sales team, that is the kind of system we design when we build AI voice agents. For where a person should step in during an AI call, read human in the loop AI on the phone.

FAQ

Can AI agents do sales?

They can do much of the work: answering enquiries instantly, qualifying, handling common objections, following up on time, issuing invoices and sending payment links. The weaker areas are reliability on multi-step tasks and trust in considered purchases. Research such as Salesforce's CRMArena-Pro found agents' success rates fall sharply when tasks need follow-up questions, so a person should still own the decisive moments.

Can an AI sales agent offer discounts?

It can, but only from a list of approved offers with clear conditions and a hard limit. Anything beyond that should go to a named manager for approval. A Canadian tribunal held Air Canada liable for what its chatbot told a customer, so treat everything your AI says as a company commitment.

How much does an AI sales agent cost?

Pricing changes quickly and depends on what the agent must do. Voice platforms usually charge per minute, and a headline platform rate often excludes the speech, model and phone charges added on top. The bigger cost is connecting it to your records, calendar and billing so it can act, and that work varies by business.

Will customers buy from an AI?

Some will, especially for quick, low-risk purchases. In a 2019 field experiment on outbound sales calls, purchases fell by more than 79 percent when customers were told at the start they were speaking to a bot. Better voices may reduce that penalty, but for trust-heavy sales a named person should handle the close.

Is AI cold calling legal?

In the US, the FCC ruled in February 2024 that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act, so robocall consent rules apply. Calls to people who never agreed to be contacted that way carry real legal risk. Use AI calling for people who asked to hear from you, and check state rules with a telecom attorney.

Read next: AI Sales & Outreach Engine