The State of Conversational AI, Q3 2026

Architecture, adoption, and the open-source shift

Published 28 August 2026 · Ethora Research · Watch the 90-second summary

1. Executive summary

Conversational AI in 2026 is no longer a question of "will teams build with language models" but "which layer of the stack do they buy, and which do they own". Four things stand out in this year's data:

1. Agents became the story. US search for "ai agents" grew from about 60 searches a month in early 2023 to roughly 50,000 a month in 2026 - about a 700-fold rise - and now leads the category's vocabulary.
2. A new layer appeared from nothing. "mcp server" (the Model Context Protocol) had effectively zero search volume until late 2024 and reached tens of thousands of monthly searches within about eighteen months, with its developer libraries already among the most-downloaded packages in the field. A tools-and-protocol layer standardised in a single year.
3. Self-hosting is gaining ground, at least among developers. Some of the most-starred open-source projects in the space are the ones that let teams run models and chat on their own infrastructure (Ollama, llama.cpp, Open WebUI, vLLM), and they are among the largest developer communities we measured. This is developer and early-adopter momentum: much of it is experimentation and tech-savvy startups, and most enterprises are still figuring out cloud-hosted AI, let alone running production loads on infrastructure they own. How far real commercial self-hosting follows is a trend worth tracking in future editions.
4. The category is now answered by AI first. Of the 28 buying-intent search terms we tracked, 24 trigger a Google AI Overview, and the assistants draw their recommendations largely from third-party "best of" lists rather than from vendors directly. How a product gets discovered is changing underneath the market.
5. The labour market has a clear shape. The US shows on the order of 300,000 open AI/LLM roles ("AI agent" already a top-three category) but only ~112,000 core AI/ML engineers - a talent importer - while India, with ~95,000 practitioners and far fewer local openings, is the world's exporter. By concentration, tiny Singapore leads both the demand and the supply per-capita tables.

The through-line: the raw model is commoditising, and durable differentiation is moving up (to memory, agents, and proactivity) and down (to who controls the infrastructure and the data).

2. How to read this report

We describe a conversational AI system as five layers (Section 3) and use that structure throughout. Wherever we cite adoption, we mean a specific, reproducible measurement - GitHub stars, Hugging Face downloads, package-manager installs, search demand, or sampled AI-assistant answers - and we state the limits of each. These are proxies for interest and usage, not revenue or market share. Numbers are point-in-time snapshots (August 2026). Full method: Section 10 and the source log.

Who this is for. Builders and buyers deciding how to add chat or AI agents to a product, and anyone tracking where the conversational-AI field is heading.

What this report is, and is not. This is a free, vendor-neutral map of the whole conversational-AI stack, built from primary data rather than analyst scoring, and it treats open-source and self-hosted tools as first-class rather than a footnote. It is deliberately not a vendor ranking or a buyer's shortlist for one use case. If you are selecting an enterprise platform for a specific application such as customer service, the analyst evaluations are built for exactly that and score named vendors on offering and strategy: see the Gartner Magic Quadrant for Conversational AI Platforms and the Forrester Wave for Conversational AI Platforms for Customer Service. This report instead sits alongside the free, primary-data landscape reports - among them the State of AI Report, the Stack Overflow Developer Survey, GitHub Octoverse, the Stanford AI Index, and Hugging Face's State of Open Source - and differs from those by focusing specifically on conversational AI, framing the field by architecture layer, and measuring how the AI assistants themselves now answer the category.

How to cite. "The State of Conversational AI, Q3 2026" with a link to the source page. The charts and tables are free to reuse with attribution, and every figure links to its full underlying data.

3. Anatomy of a conversational AI system

Every conversational AI product, from a weekend chatbot to a regulated enterprise assistant, is assembled from the same five layers. Naming them makes the rest of the report legible.

1. Interface layer - where a person meets the system. Vendor apps (ChatGPT, Claude, Gemini), the terminal, voice, messaging channels (WhatsApp, Slack, Teams), and chat embedded directly inside another product via SDKs and widgets.
2. Orchestration / harness layer - the application logic wrapped around the model: routing, prompting, agent orchestration, tool use, guardrails, evaluation. This is where most "AI app" frameworks live.
3. Model layer - the language model itself, reached either through a cloud provider's API or run from an open-weight model on infrastructure you control.
4. Data and memory layer - retrieval (RAG), vector stores, short- and long-term memory, and per-agent persistent identity. What the model knows beyond its training and its current context window.
5. Infrastructure layer - where it all runs: a provider's managed cloud, your own cloud, on-premises, or the edge, plus the serving stack that hosts open models.

Figure 3.1. The five-layer anatomy of a conversational AI system, from the interface a person meets to the infrastructure it runs on. View larger, with a text description and a PNG download.

A useful way to read the rest of the report: each subsequent section examines one layer's adoption and the choices a buyer faces there.

4. Adoption and landscape

We tracked a fixed set of open-source projects across the five layers and captured GitHub stars, forks, and activity, alongside Hugging Face model downloads and npm/PyPI installs. Stars measure developer mindshare rather than production usage or revenue, and closed commercial products have no repository to count - so we read this section as "where open developer attention concentrates", and cover the commercial names through search demand and assistant answers instead.

Table 4.1 - Most-starred projects in the tracked set (top 15 by conversational-AI-relevant stars, Aug 2026). Full data: all 34 tracked projects, by layer.

On n8n. The single most-starred project we track is n8n (200,617 stars), but it is a general workflow-automation platform, not a conversational-AI tool, so counting its stars at face value would overstate its place here. To keep the comparison fair, we normalise it: of n8n's 11,772 community workflow templates, about 17% are conversational AI (1,293 chatbot plus 758 retrieval/RAG, per n8n.io in August 2026). Scaling its stars by that share puts its conversational-AI-relevant mindshare near 35,000. Its raw figure is in the full data.

Figure 4.1. Open developer attention by GitHub stars, top tracked conversational-AI projects (n8n shown at its conversational-AI-adjusted figure).

Reading: open developer energy concentrates at the two ends of the stack - running models (serving/self-hosting: Ollama, llama.cpp, vLLM) and wiring them together (orchestration and the new protocol layer) - more than on any single interface. The Model Context Protocol servers repository reached about 90k stars less than a year after its late-2024 creation - the fastest riser we tracked.

Install counts corroborate and sharpen the picture, because they reflect usage rather than mindshare.

Table 4.2 - SDK / library downloads, last month (Aug 2026). Full data: all tracked SDKs & libraries.

Two things stand out. LiteLLM - a library that routes calls across many model providers - being pulled hundreds of millions of times a month says multi-provider abstraction is now default plumbing; few teams commit to a single model vendor. And MCP appearing near the top of both registries, barely a year after launch, confirms the protocol layer is real adoption, not just conversation.

Search demand confirms the same shifts from the buyer side. US search for "ai agents" rose roughly 700-fold since early 2023, "mcp server" from effectively zero before late 2024, and "self hosted llm" on a steady climb (by "self-hosted LLM" we mean running an open-weight model on your own infrastructure - on-premises, private cloud, or local), while "rag" doubled and then plateaued into assumed infrastructure. Notably, 24 of the 28 category terms we tracked now trigger a Google AI Overview. Full data: search demand and 3-year trends for 28 terms.

Figure 4.2. Three-year US search-volume trend for headline category terms (ai agents, mcp server, self-hosted LLM, rag). Volumes are third-party estimates.

5. The model layer: cloud versus self-hosted

The sharpest structural question of 2026 is whether to call a model over an API or run an open-weight model yourself. The open-weight side is now a serious field of its own.

Table 5.1 - Open-weight model families by Hugging Face downloads (top-8 models per publisher, 30-day, Aug 2026). Click a column header to sort by downloads or likes. Full data: families + top models.

Downloads and affection diverge, and the gap is itself a finding. Qwen and Gemma lead on downloads (what people run in volume), but on likes (a mindshare signal) the order flips - DeepSeek, Llama, Kimi, and Mistral lead while Qwen sits far lower. Heavy production pull toward Qwen and Gemma; enthusiasm concentrated on DeepSeek, Llama, and Kimi. There are two different "leading" open models depending on whether you measure what people run or what people love.

Figure 5.1. Open-weight models: 30-day Hugging Face downloads (what people run) versus likes (a mindshare signal). The two orderings diverge.

Licence terms vary widely, and can constrain commercial use. Most families are commercially self-hostable, which is what makes self-hosting a realistic option rather than research-only, but it is important to understand the exceptions and caveats.

Table 5.2 - Licence and self-hosting footprint (verify per model card before relying on any cell). Full data: licences & footprint notes.

Cohere's Command family is non-commercial; Mistral's newest flagship models are research-only unless separately licensed; Llama and Gemma carry custom terms with real restrictions. A team that picks a model on capability alone can back into a licensing problem. Footprint spans two orders of magnitude - Phi and small Qwen/Gemma variants run on modest hardware, while the large DeepSeek and Kimi mixture-of-experts models are data-centre-class.

(Caveat: download counts skew toward small/quantised variants, and one deployment can trigger many downloads. We report family-level trends, not install counts. Licence terms vary per model and change with releases.)

6. The harness and application layer

Turning a raw model into a product means solving a recurring set of problems: the interface it is delivered through (apps, widgets, messaging channels), integration into existing systems and data, retrieval (RAG), memory (short-term context and long-term persistent identity), multi-user and channels, files, tool use and protocols (function calling, MCP), proactivity (scheduled or triggered actions), identity and access, and deployment control. We verified these capabilities across eighteen notable systems against their current documentation (August 2026).

First, the whole stack at a glance - the notable tools in each layer, open-source ones shaded by adoption:

Figure 6.1. The Periodic Table of Conversational AI: the notable tools by layer, open-source tiles shaded by GitHub-star tier and sorted by adoption within each column. Illustrative selection, not exhaustive.

And the capability detail behind it:

Table 6.1 - The Periodic Table of Conversational AI. Y = native, P = partial (add-on / paid tier / product-not-API), N = not documented. Full matrix + per-cell sources.

Self-hosted assistant UIs

Frameworks / builders

CX / chatbot platforms

Commercial cloud (developer APIs + agent platforms)

(Confidence and columns: each capability cell was verified against the product's current official documentation (August 2026) on a conservative bar - where the docs did not clearly show a capability, we mark it N or P rather than assume it. The same nine columns apply to every sub-table.)

What the verified data shows:

• Tool use and MCP are now table stakes. Nearly every system supports them; the rare exceptions offer native tools but not standard MCP. A capability that did not exist before late 2024 is already near-universal.
• Proactivity is the real differentiator. Acting on a schedule or trigger, rather than only replying, cleanly splits the field. This is where a chatbot becomes an agent.
• Persistent memory is common in name but unfinished in practice - often a recent add-on, a paid tier, or a consumer-product feature not exposed to developers (hence the cluster of P in the commercial-cloud rows).
• Deploy and control is the cleanest fault line - open-source and self-hostable versus cloud-only, with a few hybrids (an on-premises path here, a formerly-self-hostable product that moved to cloud there).

The synthesis: no single system is strong on control, conversation, and agency at once. Self-hosted tools own control; commercial platforms own managed polish and proactivity; memory sits unfinished across both. That unresolved middle is the category's current frontier.

7. How the AI assistants answer the category

Because most category searches now surface an AI answer, we sampled what the assistants actually say. Using a fixed set of buyer questions, we recorded which tools ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, and Copilot name, and which sources they cite.

The assistants hold two separate mental models of the category, keyed to the question's wording.

Table 7.1 - What the assistants commonly recommend, by question type. This reflects what the assistants actually named in our sample, not a market ranking. Full data: assistant share of voice, brands named, and cited sources.

Where the answers come from is worth noting. The assistants build these recommendations disproportionately from a handful of third-party "best of" lists, directories, and aggregators - g2.com, alternativeto.net, toolradar.com, opensourcechoice.com, ossalt.com, aimultiple.com - alongside GitHub and vendor documentation, rather than from vendors' own pages. Because they lean on a small set of easy-pick sources, their recommendations can inherit those sources' biases and omissions: a tool absent from the popular lists can be missing from the answer even where it is a strong fit, and a well-listed tool can appear more dominant than it is. In 2026, presence on the lists the models read is a distinct discipline from ranking in classic search.

Different answer engines source their answers differently, and it changes who they name. Google's AI Mode and AI Overviews are grounded in web search and lean on pages that rank; the standalone assistants (ChatGPT, Perplexity, Gemini, Copilot) lean harder on curated "best of" lists and cite them almost uniformly. A product visible in classic search can appear in the search-grounded answers while being absent from the assistants that work off lists, and vice versa. Where a tool shows up depends on which kind of answer engine you ask - a fragmentation of "discovery" that did not exist when there was one search box.

(Method note: sampled across ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, and Copilot, US, August-2026. Brand matching for vendors whose name is a common English word over-counts and needs distinctive-token matching; those figures are treated as unreliable. One assistant returned no brand data this cycle and is excluded. Share-of-voice counts a brand as present when named anywhere in an answer or its cited sources, so it inflates where answers cite large multi-vendor lists; own-domain citation is the stricter signal.)

8. Talent and geography

Two views of where conversational AI is being built: the demand side (open jobs) and the supply side (where the practitioners are).

Demand: where the AI jobs are. We counted open AI/LLM roles across nine countries. The United States is the unambiguous centre of gravity - roughly 300,000 open roles, and the only market that leads on both raw volume and concentration.

Table 8.1 - Open AI/LLM roles by country (Aug 2026). Keyword counts overlap, so totals are directional demand, not unique postings. Click a column header to sort by total roles or per-capita. Full data: demand matrix by keyword and country.

The per-capita reframe surfaces a different leader. By concentration, Singapore tops the table at about 1,134 open AI roles per million people, ahead of the United States (900), Canada (301), and the United Kingdom (266); India, huge in absolute terms, sits last per capita (25). Raw volume points to the US; concentration points to a set of small, dense markets.

Figure 8.1. Open AI/LLM roles by country in absolute terms: the US dominates the raw-volume picture.

Figure 8.2. The same roles per million people: concentration reorders the ranking, and Singapore leads.

What roles, from the employer side. In the US, the largest categories are generative AI (~81k), AI engineer (~61k), and - notably - AI agent (~61k), followed by LLM (~35k), prompt engineer (~20k), and RAG (~19k). The agent and retrieval themes visible earlier in developer interest and search demand now show up directly in hiring: "AI agent" is already a top-three hiring category, and RAG is a named requirement in tens of thousands of roles.

The labour market is barbelled: serious roles are hired in-house, the low end has commoditised into gig work. Alongside the salaried employment market, we sampled two freelance marketplaces, and they show the two ends of that barbell. On Freelancer.com, AI projects are overwhelmingly small and fixed-price - a median advertised budget around $250 (hourly work around $14/hr), priced mostly in Indian rupees, for chatbot-integration, "connect ChatGPT to X", and automation tasks. As advertised, these under-state real engagement value: fixed-price posts often grow in scope, and experienced freelancers negotiate higher rates up front. On Upwork, the other end shows through: across roughly 100 open AI chatbot and LLM listings (August 2026), the work is almost entirely hourly (over 95% of postings), at a median around $30/hour (middle half $24-$40, up to $60+), and skews Expert and Intermediate, with almost nothing at entry level. The same word "AI" spans a six-figure principal-engineer role, a skilled $30-$60/hr Upwork contract, and a $250 Freelancer.com integration gig; serious conversational-AI capability is hired in-house or on ongoing skilled contracts, while low-end implementation has become offshore piecework. Full data: freelance market.

Supply: where the practitioners are. Counting LinkedIn profiles by job title across the same countries, the United States and India dominate the pool of core AI/ML engineers - roughly 112,000 and 95,000 respectively (on the "AI Engineer" and "Machine Learning Engineer" titles), an order of magnitude ahead of the next tier (United Kingdom ~15,800, Germany ~14,200, Canada ~13,400). A market difference shows in the titles themselves: India carries more "AI Engineer" profiles than the US (59k vs 44k), while the US leads on "Machine Learning Engineer" (68k vs 36k) - naming conventions differ by market.

Table 8.2 - Core AI/ML engineers per million people (top). Full data: talent by title and country.

Figure 8.3. Core AI/ML engineers by country in absolute terms: the US and India dominate the supply pool.

Figure 8.4. The same practitioners per million people: Singapore leads both the supply and the demand concentration tables.

The concentration leader is, again, Singapore - about 876 AI/ML engineers per million people, ahead of the UAE (410), Switzerland (407), Ireland (348), and the Netherlands (341), with the US and Canada at ~334. Strikingly, Singapore tops both the supply and the demand concentration tables - a genuinely dense hub, even though its absolute numbers are small.

A caveat on the supply data, and a fuller picture elsewhere. These supply figures come from LinkedIn profile counts, which under-represent markets where LinkedIn is not the dominant professional network - most importantly China, which is largely absent here despite being one of the world's largest sources of AI talent. Read this as a LinkedIn-visible view, not a global census, and expect China to rank far higher on the supply side than it appears here. The Stanford AI Index 2026 is a good complement on a different methodology: it also finds Switzerland and Singapore leading the world in AI researchers and developers per capita - the same small, dense hubs our data surfaces - while covering the large absolute-supply markets, China included, that a LinkedIn-based count cannot see.

Supply meets demand: who imports talent and who exports it. Setting open roles against the local practitioner pool gives a rough market-tightness read (the two are measured differently, so treat it as directional). The United States is markedly demand-heavy - about 2.7 open roles per core practitioner - a talent-hungry importer; Singapore (1.3) and the UK (1.1) lean the same way. India is the opposite at 0.4 - far more practitioners than local openings, a net talent exporter - as are France (0.3) and the Netherlands (0.3). That mirrors the freelance picture: India supplies much of the world's AI implementation capacity.

(Method: employment demand from Adzuna open-job counts by keyword x country; technology-demand read from Remotive, Jobicy, and Hacker News "Who is hiring"; freelance demand from Freelancer.com's active-projects API (skill-filtered to AI) and a manual sample of about 100 Upwork AI chatbot/LLM listings (August 2026); supply from LinkedIn Recruiter profile counts by job title, one country at a time with a constant title filter (the core figure uses the "AI Engineer" and "Machine Learning Engineer" titles - the cleanest as job titles; broader skill terms such as "LLM" or "RAG" are rarely titles and undercount). Coverage varies by market - which likely inflates Singapore's per-capita on both sides, understates India, where much AI work is offshore/contract, and largely omits China, where LinkedIn is not the dominant professional network. The supply-vs-demand ratio compares differently-measured quantities and is directional only. Population figures approximate.)

9. Trends and outlook

Drawing the threads together, the evidence points to a few directions rather than predictions:

• The model is commoditising; the abstraction layer won. Multi-provider routing at massive install scale means few teams marry a single model vendor. Differentiation moves off the model.
• The protocol layer standardised in one year. MCP went from nonexistent to top-of-registry, which suggests tools and interoperability - not raw capability - are the near-term battleground.
• Self-hosting is gaining adoption as a durable stance, driven by cost at scale, data control, and compliance - so far led by developers and technical teams, with broader enterprise adoption still emerging. The open-weight model field is now deep enough to support it.
• Discovery is shifting to AI answers, which are assembled differently by different engines (search-grounded vs curated lists). Vendors that treat "being cited by the assistant" as a discipline will have an advantage the way early SEO adopters did.
• The frontier is memory, proactivity, interface and integration depth, and multi-tenant control - the capabilities where every system is still partial.

10. Methodology and appendix

• Methodology and data availability: how every figure was measured - sources, tools, collection dates, and limitations - is documented on the methodology page. Every figure links to a full-data page, and every number traces to a dated source line.
• Reproducibility and reuse: this report and its data tables are released under CC BY 4.0. The processed datasets and the collection scripts are published alongside the report (see the citation box below), so any figure can be independently reproduced and re-run in future editions.
• Limits: stars and downloads are interest proxies, not usage or revenue; search-volume figures are estimates; assistant sampling is point-in-time and platform-dependent. Each section states its own caveats.

How to cite, license, and data availability

How to cite this report. Ethora Research (2026). The State of Conversational AI, Q3 2026. Dappros Ltd. DOI: 10.5281/zenodo.22131590. https://ethora.com/research/state-of-conversational-ai-2026-q3/

BibTeX:

@techreport{ethora_convai_q3_2026,
  title  = {The State of Conversational AI, Q3 2026},
  author = {{Ethora Research}},
  institution = {Dappros Ltd},
  year   = {2026},
  month  = {aug},
  doi    = {10.5281/zenodo.22131590},
  url    = {https://ethora.com/research/state-of-conversational-ai-2026-q3/}
}

License. This report and its data tables are released under CC BY 4.0: free to reuse, quote, and republish (including the charts and tables) with attribution to The State of Conversational AI, Q3 2026 and a link back to the source page.

Data availability. Every figure traces to a dated source line in the provenance log, and every table links to a full-data page. Collection methods, tools, dates, and limitations are documented in the methodology. The processed datasets and the collection scripts are published in the open-data repository at github.com/dappros/state-of-conversational-ai, so any figure can be independently reproduced.

Authorship and contact. Compiled by Ethora Research (Dappros Ltd). Corrections and source suggestions are welcome via ethora.com/contact.