The Brief The AI Layer series · piece 6 · Last updated July 2026
Conversational AI taxonomy for the enterprise: a buyer's disaggregation for 2026.
IVA, voicebot, chatbot, copilot, agent. Vendors deliberately blur the terms to defend adjacent bookings. RFPs get written around whichever term the loudest analyst used last quarter. In 2026 the market has consolidated enough to make disaggregated procurement possible — orchestration, domain agent, LLM inference, voice, evaluation, knowledge as separate line items, with a named vendor at each layer.
8 min · The AI Layer · Series piece 6
Questions this article answers
- What is the working definition of IVA, voicebot, chatbot, copilot, and agent?
- Who did Gartner name a Leader in the 2026 MQ for Conversational AI Platforms?
- Which platforms are horizontal versus hyperscaler-native, and when to pick each?
- How did the SoundHound-LivePerson deal and Ultimate.ai absorption change the vendor map?
- What is the practical difference between copilot and agent in production?
- What belongs on separate line items in the RFP?
The conversational-AI category has become a naming problem, not a technology problem. Vendors deliberately blur the terms because that is how adjacent bookings are defended. In 2026 the market has consolidated enough to make disaggregated procurement possible — and to make the RFP a shorter document.
Working definitions first. An IVA — intelligent virtual agent — is the umbrella term, usually voice-first, that predates the LLM era and still lives in RFPs from analysts and BPO buyers. A chatbot is text, usually digital-first, sometimes rules-based, increasingly LLM-grounded. A voicebot is the voice sibling, dependent on STT and TTS around an intent model. A copilot assists a human who executes; an agent executes itself. Agentic AI is the assembly — an LLM plus tools plus a runtime plus a defined authority boundary — that lets multiple agents coordinate to complete a task. Cognigy's own explainer frames the copilot-versus-agent line as assist versus execute; that is the framing the RFP should adopt because it is the line at which authority, licensing, and audit obligations change.
The 2026 vendor map after consolidation
Per the CX Today rundown of the July 2026 Gartner Magic Quadrant for Conversational AI Platforms, Google Cloud, Salesforce (debut), SoundHound (moved from Visionary), and Kore.ai land in the Leaders quadrant. Cognigy — now NiCE Cognigy after the $955M September 2025 close — dropped from Leader to Visionary. LivePerson was removed from the report entirely following the April 21, 2026 SoundHound acquisition announcement. Boost.ai dropped from Leader to Challenger. Treat the CX Today and vendor recaps as secondary summaries; the full quadrant map sits behind paid Gartner access.
Horizontal platforms versus hyperscaler-native builders
Horizontal platforms sit above the cloud choice. NiCE Cognigy, Kore.ai (XO / Artemis), Sinch (Chatlayer + Voice Relay), Sprinklr Service, Yellow.ai, and Ada orchestrate across messaging, voice, and email. Hyperscaler-native builders — Amazon Lex, Google Dialogflow CX, Microsoft Copilot Studio — are tighter and cheaper inside their own service mesh but strand a buyer whose estate spans two clouds. Customer-facing multi-channel work usually lands on a horizontal orchestrator. Internal, employee-facing work in a Microsoft estate usually lands on Copilot Studio, whose Computer-Using Agents hit GA May 13, 2026 and whose A2A protocol is generally available. Salesforce Agentforce is native to the Salesforce data cloud and best-fit when the record system is already there.
What the acquisitions actually changed
Ultimate.ai no longer ships as a standalone product. Per the Zendesk marketplace listing, it is Zendesk AI Agents in Essential (included in Suite) and Advanced (per automated resolution) tiers. Buyers holding a legacy Ultimate MSA should renegotiate at Zendesk's per-resolution pricing surface and stress-test the volume assumptions — well-tuned bot volume can invert the economics against a flat-rate horizontal platform. On the LivePerson side, close is expected in the second half of 2026; treat the post-close roadmap as the SoundHound Amelia 7 and OASYS surface, and treat LivePerson's own Syntrix evaluation platform (March 2026) as an evaluation reference rather than a long-term standalone bet. Cognigy's Leader-to-Visionary drop reflects distraction risk from integration into NICE CXone; ask NICE Cognigy which roadmap items shipped in the twelve months since close before committing to a multi-year renewal.
Copilot versus agent, in production
The line is authority. A copilot suggests, drafts, retrieves. A human presses the button. An agent presses the button itself, within a defined authority boundary. In production, most mid-market deployments in 2026 keep the copilot on drafting and knowledge retrieval while giving the agent narrow autonomy on ticket routing, retrieval, and closed-loop workflows with an audit log. Copilot Studio's A2A GA makes multi-agent workflows contractable, but it also triggers Agent 365 governance licensing ($15/user/mo, May 1, 2026 GA) that many buyers do not scope. Voice adds real-time constraints: a chat-first orchestrator bolted onto a telephony bridge routinely exceeds the sub-one-second loop threshold that determines caller abandonment. Contract for latency at the 95th percentile, not just intent accuracy.
The disaggregation grid
Conversational AI procurement disaggregation grid
- Orchestration / flow (the "brain"). Cognigy, Kore.ai, Sinch, Sprinklr, Copilot Studio, Dialogflow CX. Contract for author-time governance, evaluation harness, and observability — not runtime alone.
- Domain agent (the "worker"). Ada, Zendesk AI Agents (ex-Ultimate), Salesforce Agentforce for CX, ServiceNow Now Assist for ITSM. Contract per resolution with a human-handoff SLA.
- LLM inference. Bring-your-own (OpenAI, Anthropic, Google, Meta) or vendor-hosted. Require model-swap portability in the contract so a foundation-model regression does not strand the workload.
- Voice (STT/TTS/telephony bridge). Sinch Voice Relay, SoundHound, Google Cloud STT, ElevenLabs, Deepgram. Separate SKU with sub-one-second loop latency SLA as the 2026 baseline.
- Evaluation and observability. LivePerson Syntrix, LangSmith, vendor-native eval. Contract for pre-production simulation, not just post-hoc drift monitoring.
- Knowledge (RAG grounding). Glean, Guru, Notion, SharePoint. Require attribution in responses and a stale-content aging policy.
- Compliance. PII redaction pre-prompt, retention policy, two-party consent handling for any voice workload.
- Naming discipline. Replace "conversational AI platform" in the RFP with the specific layer(s) being bought. Require every vendor to declare which layers they own versus partner-deliver.
Where these deployments break
Failure mode one. Category rebranding hides feature debt. A vendor whose parent has just closed a large acquisition ships a rebranded product page and a thinner backlog. Post-close, ask for the actual roadmap items shipped since close, not the announced pipeline.
Failure mode two. Per-resolution pricing tiers invert on well-tuned bots. Volume spikes on a bot with a high containment rate can push the annual bill above a flat-rate horizontal platform. Model per-resolution economics at three volume scenarios — plan, upside, and 2× peak — before signing.
Failure mode three. Multi-agent workflows trigger governance licensing the buyer did not scope. Copilot Studio A2A is powerful, and Agent 365 is required to govern the agents at scale. Same shape appears in horizontal platforms: agent governance is a separate SKU almost everywhere in 2026.
Three rules for a mid-market operator
Rule one. Buy layers, not platforms. The bundle discount is real; the substitution risk is worse.
Rule two. Contract model-swap portability. A foundation-model regression is a when, not an if. The workload must survive it.
Rule three. Name the line between copilot and agent in the RFP. That line is where authority, licensing, audit, and adoption all change at once.
In short
- Gartner's July 2026 MQ for CAIP names Google Cloud, Salesforce (debut), SoundHound, and Kore.ai as Leaders; Cognigy fell to Visionary; LivePerson removed post-SoundHound announcement.
- Ultimate.ai ships as Zendesk AI Agents in Essential and Advanced tiers; scope per-resolution economics at three volume scenarios before signing.
- Copilot Studio Computer-Using Agents hit GA May 13, 2026; A2A is generally available; Agent 365 governance licensing ($15/user/mo, May 1, 2026 GA) is a separate scope item.
- Sinch Voice Relay (March 2026) sets the reference for the text-LLM-to-live-voice bridge with sub-one-second loop latency as the 2026 baseline.
- Procure orchestration, domain agent, LLM inference, voice, evaluation, knowledge, and compliance as separate line items. Bundle discounts hide substitution risk.
- Copilot assists; agent executes. Name the line between them in the RFP or the licensing will name it for you.
Sources
- CX Today, “Gartner Magic Quadrant for Conversational AI Platforms 2026,” rundown. cxtoday.com
- SoundHound, “Named a Leader in the 2026 Gartner MQ for CAIP,” press release July 13, 2026. soundhound.com
- Salesforce, “Named a Leader in the 2026 Gartner MQ for CAIP.” salesforce.com
- SoundHound, “SoundHound to Acquire LivePerson,” press release April 21, 2026. globenewswire.com
- NiCE Cognigy, platform page. cognigy.com
- Kore.ai, “Launches Artemis, the New Generation of the Kore.ai Agent Platform,” press release May 21, 2026. businesswire.com
- Sinch, “Voice Relay launches to bring AI agents a voice,” press release. sinch.com
- Sprinklr, “Spring '26 (26.4) release,” newsroom April 7, 2026. sprinklr.com
- Zendesk, marketplace listing for Ultimate. zendesk.com
- Microsoft, “Copilot Studio updates: Computer-Using Agents and A2A,” May 2026. microsoft.com
- Cognigy, “AI Copilots vs Agentic AI,” explainer. cognigy.com
All linked sources were live at time of publish (July 2026). Verify before quoting in a procurement document.
Want a calibration on your conversational-AI stack?
Run the Tier 1 benchmark.
Submit the current RFP or the incumbent MSA. We return a benchmark PDF in five business days mapping every line item to the seven-layer grid — orchestration, domain agent, LLM inference, voice, evaluation, knowledge, compliance — and flag which layers are duplicated, missing, or bundled in a way that hides substitution risk. Free. No follow-up sales drip.
Run the Tier 1 benchmark →Series · The AI Layer
Read the anchor, Where AI actually fits in your tech stack. The CCaaS-side sibling, The AI layer on top of CCaaS. The voice-agent decision at Voice AI agents in 2026: where it works. And the Method framing at the Cardinal Method.