Modernizing a DTMF IVR to conversational AI

See also

Sources

1. 1
   Amazon Connect publishes bot metrics, including bot conversations completed, percent bot conversations outcome, average bot conversation turns and average bot conversation time, on the Flows and conversational bot performance dashboard and through the GetMetricDataV2 API, with data available from 2 December 2024.

   Connect Customer bot metrics and analytics · Page introduction; Notes under each metric · Checked 2026-09-30
2. 2
   In Amazon Connect bot metrics a single turn is one request from the client application and one response from the bot.

   Connect Customer bot metrics and analytics · Average bot conversation turns · Checked 2026-09-30
3. 3
   When Amazon Connect's Get customer input block runs in Amazon Lex mode, the touchtone buffer is cleared before the bot interaction, buffered digits are not used, and the buffer contents are not passed to the Lex bot.

   Flow block in Connect Customer: Set Touchtone Buffer Behavior · How the buffer interacts with other blocks > Get customer input · Checked 2026-09-30
4. 4
   Amazon Connect's touchtone buffer is cleared automatically when the contact is transferred to an agent or queue, when a Stop and Clear action runs, and when the contact ends.

   Flow block in Connect Customer: Set Touchtone Buffer Behavior · Automatic clearing · Checked 2026-09-30
5. 5
   AWS documents that when a caller enters DTMF quickly across a Transfer to flow boundary between a Get customer input block and a Store customer input block in the next flow, some digits may be dropped because of setup time between flows.

   Flow block in Connect Customer: Get customer input · Error scenarios · Checked 2026-09-30
6. 6
   DTMF input given to a Lex bot through Amazon Connect is exposed as the Lex request attribute x-amz-lex:dtmf-transcript, with a value of at most 1024 characters.

   Flow block in Connect Customer: Get customer input · How to configure this block > Additional configuration tips (last bullet) · Checked 2026-09-30
7. 7
   Amazon Connect's Get customer input block exposes Timeout, Default and Error branches: Timeout when no input arrives in time, Default when input matches no DTMF condition or configured Lex intent, and Error when the block errors or a Lex intent is not fulfilled.

   Flow block in Connect Customer: Get customer input · How to configure this block > Flow block branches · Checked 2026-09-30
8. 8
   In Amazon Connect's Get customer input block the DTMF timeout can be set from 1 to 180 seconds, must be defined statically, and for voice is the time until the first DTMF digit is entered.

   Flow block in Connect Customer: Get customer input · How to configure this block > Configure for DTMF input > Set timeout · Checked 2026-09-30
9. 9
   The Amazon Connect Get customer input block in DTMF mode accepts only a single digit (0-9) or the characters # and *; multi-digit entry is not supported and the documentation directs multi-digit collection to the Store customer input block.

   Flow block in Connect Customer: Get customer input · Description · Checked 2026-09-30
10. 10
    In Amazon Connect, barge-in is enabled globally by default for Amazon Lex (V2) bots and disabled by default for Amazon Lex (Classic) bots, and the Classic barge-in session attribute controls only Lex barge-in, not DTMF barge-in.

    Flow block in Connect Customer: Get customer input · Configure for Amazon Lex input > Barge-in configuration and usage for Amazon Lex · Checked 2026-09-30
11. 11
    When a Lex bot collects keypad input in Amazon Connect and callers are not prompted to finish with #, they wait five seconds before Lex stops waiting for further key presses.

    Flow block in Connect Customer: Get customer input · How to configure this block > Additional configuration tips · Checked 2026-09-30
12. 12
    AWS documents that Amazon Lex bots support both spoken utterances and keypad input when used in an Amazon Connect flow.

    Flow block in Connect Customer: Get customer input · How to configure this block > Additional configuration tips · Checked 2026-09-30
13. 13
    The Amazon Connect admin guide states that the Lex Max Speech Duration (x-amz-lex:audio:max-length-ms) defaults to 12000 milliseconds, with a maximum of 55000 milliseconds, and that a higher value routes the contact down the Error branch.disputed

    Flow block in Connect Customer: Get customer input · Configure for Amazon Lex input > Configurable time-outs for voice input > Amazon Lex tab > Max Speech Duration · Checked 2026-09-30
14. 14
    Amazon Connect's Get customer input block can branch on a Lex sentiment score before the Lex intent; the score reflects only the customer's last utterance, and when both negative and positive thresholds are set the negative one is evaluated first.

    Flow block in Connect Customer: Get customer input · Configure for Amazon Lex input > Intents > Use sentiment override · Checked 2026-09-30
15. 15
    For Lex voice and DTMF session attributes in Amazon Connect, precedence is: attributes supplied by the customer's Lambda function, then attributes defined in the Get customer input block, then service defaults.

    Flow block in Connect Customer: Get customer input · How to configure this block > Additional configuration tips · Checked 2026-09-30
16. 16
    The Amazon Connect admin guide states that the Lex Start Silence Threshold (x-amz-lex:audio:start-timeout-ms) defaults to 3000 milliseconds.disputed

    Flow block in Connect Customer: Get customer input · Configure for Amazon Lex input > Configurable time-outs for voice input > Amazon Lex tab > Start Silence Threshold · Checked 2026-09-30
17. 17
    Amazon Connect's Set Touchtone Buffer Behavior block, when enabled, collects keypad input (0-9, #, *) into a buffer of up to 30 characters as the caller presses keys, including while prompts play and between flow block transitions.

    Flow block in Connect Customer: Set Touchtone Buffer Behavior · Description · Checked 2026-09-30
18. 18
    In Copilot Studio basic voice agents, barge-in supports both DTMF-based and voice-based interruption, and Microsoft advises turning it off for a recently updated message, a compliance message that should not be interrupted, and the first agent message.

    Configure basic voice agents · Turn on barge-in > When to turn off barge-in; Specifications · Checked 2026-09-30
19. 19
    For Copilot Studio basic voice agents, Microsoft states that timers such as the DTMF interdigit timeout and silence detection timeout are not supported when DTMF only for voice is enabled, and that the only customer engagement hub supporting voice agents is Dynamics 365 Customer Service.

    Basic voice agents overview · Tips and limitations > 'DTMF only' and 'Customer engagement hub' · Checked 2026-09-30
20. 20
    Microsoft's real-time agent design guidance frames two first decisions, how the conversation is controlled and how speech is handled end to end, and says these shape cost, latency, flexibility, compliance and operational complexity.

    Overview of real-time agent design optimization · Select the orchestration · Checked 2026-09-30
21. 21
    Microsoft warns that turning off Enable voice on a Copilot Studio agent whose Telephony channel is enabled can break the agent because all DTMF triggers are automatically disabled, and that the change takes effect only when the agent is published.

    Configure basic voice agents · Turn on basic voice > Turn off optimization for voice (Important note and last bullet) · Checked 2026-09-30
22. 22
    Microsoft recommends per-question DTMF mapping over a global DTMF-only setting (System.Conversation.OnlyAllowDTMF) in Copilot Studio, because informational messages sent while DTMF-only mode is on can produce an empty DTMF grammar and telephony errors.

    Configure basic voice agents · Use voice variables > Note following the variables table · Checked 2026-09-30
23. 23
    Copilot Studio voice agents expose System.Activity.UserInputType, which records whether the caller used DTMF or speech, and System.Activity.SpeechRecognition.Confidence, a value between 0 and 1 from the last speech recognition event.

    Configure basic voice agents · Use voice variables > variables table · Checked 2026-09-30
24. 24
    In Copilot Studio basic voice agents the agent-level silence detection timeout defaults to No silence timeout, meaning the agent waits indefinitely for input, and the default fallback action after silence reprompts on a Question node is Hang up, with Escalate available as an alternative.

    Configure basic voice agents · Agent-level settings table, row 'Silence detection timeout'; Configure silence detection and timeouts, step 6 'Fallback action' · Checked 2026-09-30
25. 25
    Microsoft documents that Copilot Studio voice agents support speech and DTMF input, context variables, call transfer, and speech and DTMF customization.

    Use interactive voice response in your agents · Opening paragraphs (fourth paragraph) · Checked 2026-09-30
26. 26
    Copilot Studio basic voice agent defaults are: DTMF interdigit timeout 3,000 ms, DTMF termination timeout 2,000 ms, utterance end timeout 1,500 ms (values above 3,000 ms are reduced to 3,000 ms), and speech recognition timeout 12,000 ms.

    Configure basic voice agents · Agent-level voice options reference > Agent-level settings table · Checked 2026-09-30
27. 27
    Copilot Studio basic voice agents support blind transfer to a PSTN phone number and to a direct routing number; an optional SIP UUI header sends only its first 256 characters, and transfers carrying SIP UUI must use direct routing because PSTN numbers do not support SIP UUI header transfers.

    Configure basic voice agents · Transfer a call to a representative or external phone number · Checked 2026-09-30
28. 28
    Copilot Studio offers two kinds of voice-enabled IVR agent: basic voice agents, which use classic orchestration with natural language understanding models, and real-time voice agents, which are used with generative orchestration and a speech-to-speech real-time model.

    Use interactive voice response in your agents · Opening paragraphs · Checked 2026-09-30
29. 29
    Dialogflow CX experiments support automatic rollout that raises a variant's traffic in defined steps under rollout rules and halts on configured stop conditions, and automatic rollout supports only one variant flow.

    Experiments · Auto rollout and steps section · Checked 2026-09-30
30. 30
    In Dialogflow CX experiments the Contained metric is the count of sessions that reached END_SESSION without triggering live agent handoff, callback or abandonment, and it is available only to agents using a telephony integration.

    Experiments · Metrics table, row 'Contained' · Checked 2026-09-30
31. 31
    Dialogflow CX DTMF settings include a finish digit that ends input immediately and is excluded from the value the agent receives, an interdigit timeout, an endpointing timeout, and a Disable speech detection option that makes the agent ignore speech and use DTMF-only input.

    DTMF for telephony integrations · DTMF settings: Finish digit; Disable speech detection; Interdigit timeout duration; Endpointing timeout duration · Checked 2026-09-30
32. 32
    Google's voice agent design guidance recommends a combined no-match and no-input maximum of 3 on every page.

    Voice agent design best practices (Dialogflow CX) · No-match error handling · Checked 2026-09-30
33. 33
    Dialogflow CX experiment metrics also include live agent handoff rate, callback rate, abandoned rate, session end rate, total no-match count, total turn count and average turn count, with callback and abandoned rates limited to telephony integrations.

    Experiments · Metrics table · Checked 2026-09-30
34. 34
    Running a Dialogflow CX experiment requires the agent's Enable conversation history setting to be on, and a flow can have only one experiment running at a time.

    Experiments · Prerequisites / preparation section; limitations · Checked 2026-09-30
35. 35
    Dialogflow CX experiments compare one or more flow versions (variants) against a control version while serving live traffic, with a chosen share of traffic allocated to each version.

    Experiments · Page introduction · Checked 2026-09-30
36. 36
    In Dialogflow CX, when the pause between DTMF digits exceeds the interdigit timeout, the digits captured so far are sent to the agent, which the documentation says is likely to trigger a no-match event.

    DTMF for telephony integrations · Interdigit timeout duration · Checked 2026-09-30
37. 37
    Google's voice agent design guidance lists misroute rate, first call resolution, average handling time, customer satisfaction surveys, number of turns per task, and user churn or disengagement as measures of agent quality.

    Voice agent design best practices (Dialogflow CX) · Measure agent quality · Checked 2026-09-30
38. 38
    Google's voice agent design guidance recommends repeating the original question unchanged on a first no-input, and rephrasing the question concisely on a first no-match, with more effort shown on the second occurrence.

    Voice agent design best practices (Dialogflow CX) · No-match error handling; No-input error handling · Checked 2026-09-30
39. 39
    Google warns that changing the version in a Dialogflow CX production environment while sessions are active may cause agent errors, and recommends making version changes during off-peak hours.

    Versions and environments · Production traffic guidance (note on changing versions with active sessions) · Checked 2026-09-30
40. 40
    A Dialogflow CX version is an immutable snapshot of a flow and its associated agent data, and Google advises serving production traffic from versions rather than the draft, which changes with every edit and can be broken unintentionally.

    Versions and environments · Version definition; production traffic guidance · Checked 2026-09-30
41. 41
    The vendor success measures read for this packet are not interchangeable: Dialogflow CX 'Contained' counts sessions ending without handoff, callback or abandonment; Lex V2 conversation 'Success' depends on the final intent being fulfilled; and the Webex Virtual Agent V2 'Handled' outcome fires when virtual agent execution completes; so a containment figure cannot be compared across platforms, or with a legacy DTMF IVR, without restating the definition.inferred

    Experiments · Metrics table, row 'Contained' · Checked 2026-09-30
42. 42
    Across the Google, AWS, Microsoft and Cisco documentation read for this packet, speech and DTMF are treated as coexisting input modalities on the same conversational agent, with speech-off or DTMF-only as a per-step option, which suggests vendors expect a modernized IVR to remain hybrid rather than speech-only.inferred

    Configuring timeouts for capturing user input with a Lex V2 bot · Allow audio input; Allow DTMF input (both default True) · Checked 2026-09-30
43. 43
    The vendor tooling read for this packet supports a staged cutover pattern of baseline test set, immutable version, partial live-traffic experiment against a control, then stepped rollout with stop conditions, but none of these sources prescribes a sequence for migrating a DTMF IVR to conversational AI.inferred

    Experiments · Page introduction; auto rollout and steps section · Checked 2026-09-30
44. 44
    Because Amazon Connect clears the touchtone buffer before a Lex interaction and does not pass buffered digits to the bot, callers who habitually key ahead through a legacy menu will probably lose those key presses at the point where a DTMF menu step is replaced by a Lex bot.inferred

    Flow block in Connect Customer: Set Touchtone Buffer Behavior · How the buffer interacts with other blocks > Get customer input; Automatic clearing · Checked 2026-09-30
45. 45
    Lex V2's x-amz-lex:allow-interrupt attribute defaults to True and can be turned off selectively per intent and slot to stop callers interrupting a prompt.

    Configuring timeouts for capturing user input with a Lex V2 bot · How interrupt behavior works in a Lex V2 bot · Checked 2026-09-30
46. 46
    Lex V2 analytics classify a conversation as Success when its final intent is a success, Failed when the final intent failed or Lex defaulted to AMAZON.FallbackIntent, and Dropped when the customer stops responding before either outcome.

    Key definitions · Conversations · Checked 2026-09-30
47. 47
    Lex V2 DTMF defaults are: end character #, deletion character *, end timeout 5000 milliseconds after the last digit, and a maximum of 1024 DTMF digits per utterance that cannot be increased.

    Configuring timeouts for capturing user input with a Lex V2 bot · Set configuration for DTMF input · Checked 2026-09-30
48. 48
    The Lex V2 silence timeout x-amz-lex:audio:end-timeout-ms, the wait after the caller stops speaking before the utterance is treated as finished, defaults to 600 milliseconds.

    Configuring timeouts for capturing user input with a Lex V2 bot · Set the timeouts for voice input > Silence timeout · Checked 2026-09-30
49. 49
    The Amazon Lex V2 developer guide states that x-amz-lex:audio:max-length-ms defaults to 55,000 milliseconds, which is also the maximum, and that a higher setting falls back to 55,000 milliseconds.disputed

    Configuring timeouts for capturing user input with a Lex V2 bot · Set the timeouts for voice input > Maximum utterance length · Checked 2026-09-30
50. 50
    Lex V2 has per-intent and per-slot session attributes allow-dtmf-input and allow-audio-input, both defaulting to True, so either modality can be switched off for a specific intent or slot.

    Configuring timeouts for capturing user input with a Lex V2 bot · Allow audio input; Allow DTMF input · Checked 2026-09-30
51. 51
    The Amazon Lex V2 developer guide states that the voice timeout x-amz-lex:audio:start-timeout-ms defaults to 4,000 milliseconds.disputed

    Configuring timeouts for capturing user input with a Lex V2 bot · Set the timeouts for voice input > Voice timeout · Checked 2026-09-30
52. 52
    A Lex V2 Test Workbench test set can be created by uploading a CSV file or by generating it from conversation logs, in audio or text modality, with an import limit printed as text files up to 50k and up to 5 hours of audio.

    Generate a test set for Test Workbench · Opening paragraph; 'To create a test set' step 10 note · Checked 2026-09-30
53. 53
    The Lex V2 Test Workbench establishes baseline intent and slot performance for single-input and conversation test sets, and a loaded test set can be run against pre-production or production bots before deployment.

    Evaluating Lex V2 bot performance with the Test Workbench · Opening paragraphs · Checked 2026-09-30
54. 54
    Lex V2 analytics classify an utterance as Detected when Lex recognises it as an attempt to invoke a configured intent and as Missed when Lex does not recognise it.

    Key definitions · Utterances · Checked 2026-09-30
55. 55
    RFC 4733 specifies how DTMF signalling, other tone signals and telephony events are carried in RTP packets, and it obsoletes RFC 2833.

    RTP Payload for DTMF Digits, Telephony Tones, and Telephony Signals · Abstract; header 'Obsoletes: 2833' · Checked 2026-09-30
56. 56
    VoiceXML 2.0, a W3C Recommendation dated 16 March 2004, requires a platform to be able to detect and report character (DTMF) and spoken input simultaneously.

    Voice Extensible Markup Language (VoiceXML) Version 2.0 · Section 1.2.5 (platform requirements) · Checked 2026-09-30
57. 57
    In Webex AI Agent Studio the Agent handover action, which lets the AI agent escalate the conversation to a human agent, is enabled by default and can be turned off with a toggle.

    Webex AI Agent Studio Administration guide · Actions > Agent handover · Checked 2026-09-30
58. 58
    Webex Contact Center Virtual Agent Voice default system settings are: DTMF termination symbol #, DTMF no-input timeout 5 seconds, DTMF inter-digit timeout 3 seconds, and barge-in enabled.

    Configure Virtual Agent-Voice in Webex Contact Center · Default system settings list in the Virtual Agent V2 activity section · Checked 2026-09-30
59. 59
    Webex Contact Center Virtual Agent Voice can be backed by Google Dialogflow CX, Google Dialogflow ES, or a Bring Your Own Virtual Agent (BYOVA) integration.

    Configure Virtual Agent-Voice in Webex Contact Center · Supported integrations (overview) · Checked 2026-09-30
60. 60
    The Webex Contact Center Virtual Agent V2 flow activity has three outcomes: Handled when the virtual agent execution completes, Escalated when the call must go to a human agent, and Errored for any error scenario.

    Configure Virtual Agent-Voice in Webex Contact Center · Virtual Agent V2 activity > Outcomes · Checked 2026-09-30

Documents

Cite this page

WarmTransfer. (2026, September 30). Modernizing a DTMF IVR to conversational AI. WarmTransfer. https://warmtransfer.net/knowledge/ivr-modernization

@misc{warmtransfer-ivr-modernization,
  title  = {Modernizing a DTMF IVR to conversational AI},
  author = {{WarmTransfer}},
  year   = {2026},
  url    = {https://warmtransfer.net/knowledge/ivr-modernization},
  note   = {Verified 2026-09-30}
}