If you've ever tried to take a card payment over the phone, you know how painful it is. The customer spells out 16 digits. You ask them to repeat. They get the fourth one wrong. You start over. Three minutes later, the authorization clears — and you've blocked your only phone line the entire time.
That's exactly the kind of moment where the 2024 AI voice agent would say: "Sorry, let me transfer you to a human." Except there's no human available on a Saturday night.
That changed in April 2026. ElevenLabs added DTMF support to its conversational agents API. Microsoft, Famulor, Retell AI and most serious platforms followed within weeks. The bottom line: your voice agent can now read keypad presses from the caller — and even press keys itself when calling another number.
For a Quebec SMB, this isn't a technical footnote. It's an entire category of use cases that just opened up. Here's what changed, and why 2026 is probably the year to stop saying "no, we don't take phone payments."
So what is DTMF, exactly?
DTMF stands for Dual-Tone Multi-Frequency. Every key on your phone — 0 through 9, star, pound — produces a unique combination of two frequencies. When you press a key, those tones travel down the line. It's the technology that's powered "press 1 for billing, 2 for tech support" menus since the 1980s.
Until recently, AI voice agents were deaf to these tones. They listened for speech, full stop. If a caller pressed a key, the agent ignored it. If the agent called another system that asked "press 1 to confirm," it just sat there waiting for a voice answer that never came.
With native DTMF support, the agent does two new things: it receives digits typed by the caller (perfect for PINs, card numbers, confirmation codes), and it sends tones itself when navigating a phone menu. Both directions, in real time, without a human needing to step in.
Why this happened now
Three things converged over a few months.
First, OpenAI shipped gpt-realtime and its mini variant, which dropped the cost of an AI conversation by roughly 20%. That freed up budget at the platform vendors to add features instead of just chasing margins.
Second, 2025 production deployments surfaced the same feedback everywhere: customers refuse to dictate their card number out loud. Reasonable. You're at a restaurant, in an open office, on the bus — you're not going to shout "4532 9182 7644" in front of strangers. One bank published the numbers: its card authorization completion rate went from 52% to 78% the day it added a keypad option alongside voice. That's a massive lift for a feature this mundane.
Third, the voice agent market exploded. 97% of enterprises have adopted some form of voice AI in 2026, and 80% of SMBs plan to integrate it into customer service. When the market grows, expected features grow too. DTMF, long considered a 1980s legacy, became a mandatory checkbox.
What this unlocks concretely for a Quebec SMB
These are use cases that were unsolvable six months ago and work now.
1. Take card payments without a human in the loop
Classic scenario: a mechanic finishes a repair at 6pm. The customer wants to pay so they can pick up their car. Everyone's gone home. The AI voice agent picks up, confirms the amount, asks the customer to type their card number, CVV, expiry date — on the keypad, silently — and pushes everything to a compatible payment terminal. The customer drives off with their car, the mechanic has the cash in the account before opening the next morning.
The detail that matters: the typed digits don't get recorded in the call audio. That's gold for PCI-DSS compliance, because voice recordings containing card numbers are a real grey area. With DTMF, the number never appears in the audio.
2. Frictionless authentication
For insurance brokers, accountants, notary offices — anyone who has to confirm the identity of a caller before sharing confidential info — DTMF solves an old problem. The agent can ask: "Type the last four digits of your policy number followed by your postal code." The system validates, and the conversation continues. No need to ask ten security questions, no need to transfer to a human.
3. Navigating IVRs when the agent calls on your behalf
This is the inverse case, and the one most people underestimate. You ask your AI voice agent to call your telecom provider to cancel a line. The provider has a menu: "Press 1 for billing, 2 for service changes, 3 to cancel." Before 2026, the agent got stuck. Now, it listens, understands, presses the right key, navigates until it reaches a human or completes the transaction. For a firm that spends its day calling vendors, insurers, or banks on behalf of clients, the time savings are massive.
4. Hybrid voice + keypad in the same conversation
The user decides. The agent says: "You can tell me your reference number or type it." The caller picks what's comfortable. Someone with a strong accent or in a noisy environment types. Someone driving hands-free dictates. Same conversation, two input modes. That's the flexibility old button-press IVRs never offered, and pure voice agents didn't offer either, until now.
The Law 25 angle — often missed
Quick reminder for Quebec SMBs: any processing of personal information by a voice agent falls under Law 25. That includes call recordings, transcriptions, and data entered during the conversation.
DTMF helps here, because it separates two streams: voice (which is recorded and transcribed, so subject to retention and minimization rules) and typed digits (which can be processed separately, encrypted in transit, and purged right after use). Practically, when configuring an agent to take a payment, you program it to never log the raw digits — only the transaction hash and its status.
That separation makes compliance easier to demonstrate, because you can tell your DPO: "The card number never entered our systems in cleartext, nor in our recordings." It's a concrete argument, not marketing.
What this doesn't fix
Let's be honest. DTMF doesn't turn your voice agent into a magician.
Some conversations are still better as voice. Asking someone to type their last name on a numeric keypad is cruel. Nokia 3310 multi-tap input is dead, and that's not a nostalgia we want to bring back. Anything non-numeric remains better served by speech recognition.
And DTMF needs phone infrastructure that supports it cleanly. If your SMB uses basic poorly-configured VoIP, tones can get mangled in transit (the classic G.729 codec that "crushes" certain frequencies problem). Good vendors handle this with RFC 2833 or SIP INFO, but it's not guaranteed everywhere. That's typically the kind of detail you fix at setup — for our clients, the TECHMA team handles it, so a customer never has to debug phantom keypad tones themselves.
Platforms that support DTMF in April 2026
The ecosystem moves fast, but here's where things stand as we write:
ElevenLabs (April 2026): native DTMF support in its agents API, bidirectional. Microsoft Copilot Studio: DTMF configuration at the global level and per conversation node. Famulor: new keypad input feature for secure data collection. Retell AI: DTMF available with intelligent fallback between voice and keypad. VAPI and OpenAI Realtime: support via platform tools.
For SMBs, that means whichever platform you pick for your AI voice agent, DTMF is probably already available. The real question is no longer "does the tech exist," but "have we configured our use cases to take advantage of it."
How to know if your SMB should bother
Three simple questions to ask yourself.
How many times a week do you transfer a call to a human just to collect a card number, a PIN, or a reference code? If the answer is "more than five," DTMF solves that.
Are your customers asking to pay by phone and you refuse because it's too hard to secure? If yes, you're leaving money on the table — and frustrating customers who don't want to come into the storefront.
Are you spending time calling vendors or banks where you have to navigate menus before reaching a human? An AI voice agent that can press keys can make those calls for you and hand back the line when a human answers.
If you answered yes to at least one, the ROI is probably already positive. The detailed math, we've covered elsewhere: see the 4-step method to calculate AI voice agent ROI.
What's next
DTMF isn't the last new thing. 2026 voice agents also add multimodal input (images mid-conversation), persistent memory across calls, and real-time emotion detection. We're in a fast evolution wave that's making everything deployed before 2024 obsolete.
For a Quebec SMB, the right move isn't to deploy everything at once. It's to identify the use case that costs you the most in lost time — usually exactly the card payment or authentication scenario — and start there. The rest follows naturally once the agent proves itself on something concrete.
If you want to see what this looks like for your specific situation, we can set it up. DTMF, Law 25 compliance, integration with your existing payment terminal, French-English bilingualism — that's our team handling all of it for the client. Nothing to code, nothing to wire up yourself. We show you the agent in action on your own calls before you sign anything.