Deepgram Buys Daily and Pipecat in Voice AI Push

Direct answer: Deepgram, a voice AI company that became a unicorn earlier this year, is acquiring Daily, the company behind the open-source Pipecat framework for building voice and video AI agents. The price has not been disclosed, but the direction of travel is clear: Pipecat stays open source, Daily’s team joins Deepgram, and the combined company wants to offer developers a complete, production-ready stack for voice agents that is not locked to a single speech model.

What happened

The announcement arrived on 29 October 2026, at Deepgram’s first Deepgram Speak event. Sources: TechCrunch’s report on the acquisition and the BusinessWire announcement. Deepgram confirmed it is acquiring Daily, and Daily confirmed that its team is joining Deepgram. The financial terms were not disclosed, and neither company has announced a closing date, so the sensible reading is that this is an announced acquisition whose detailed integration will unfold over the months ahead.

What makes this deal different from a routine acqui-hire is what Daily brings with it. Daily is the maker and maintainer of Pipecat, an open-source framework that developers use to build voice and video AI applications. Pipecat has become a familiar name in the voice AI world precisely because it sits in the middle of the stack: it does not try to be the speech recogniser, the language model, or the voice synthesiser. Instead, it connects those pieces together and manages the live audio and video streams that flow between them.

Deepgram, for its part, is best known for its speech models — the technology that turns spoken words into text and, in the other direction, turns text into natural-sounding speech. Earlier this year Deepgram raised a 130 million dollar Series C round and reached unicorn status, meaning a private valuation above one billion dollars. Daily, meanwhile, has raised more than 62 million dollars from investors including Tiger Global and Highfield Capital, alongside individual backers such as Nat Friedman, Lachy Groom and Kevin Weil, according to PitchBook figures reported by TechCrunch.

There is also a shared history that helps explain why this pairing feels natural to the people involved. Both companies were in the same Y Combinator batch in 2016. That is a decade of overlapping networks, shared investors and, by the sound of it, shared engineering values. Daily co-founder Kwindla Kramer has said that experiments following the release of GPT-4 led to the creation of Pipecat, and that he was drawn to Deepgram chief executive Scott Stephenson’s commitment to keeping Pipecat open source and vendor-neutral, as well as to a matching engineering culture. Kramer is joining Deepgram as Head of Open Source, which is a signal — though not a guarantee — that the open-source commitment is meant to be structural rather than cosmetic.

Alongside the acquisition, Deepgram announced a new Voice Agent Platform that combines Pipecat with Daily’s real-time network and Deepgram’s own speech models. The pitch is a complete, open, production-grade stack: the framework, the network that carries the calls, and the speech models, available together rather than assembled piecemeal by each developer. Deepgram also announced the Real-Time Context Standard, or RCS, an open protocol co-developed with the Pipecat team, which we will come back to below because it tells you a lot about where voice AI is heading.

It is worth pausing on what has not been announced, because in acquisition news the gaps matter as much as the headlines. No price, no closing timetable, no customer numbers and no performance benchmarks have been shared. That is normal at announcement stage, but it means nobody outside the two companies can yet judge whether this was a large bet or a modest one, or how quickly the combined product will actually ship. The analysis in this article is therefore based on what the companies have said and what the structure of the deal implies, not on figures that have not been published.

How Pipecat and the voice AI stack actually work

To understand why this acquisition matters, it helps to understand what a voice AI agent actually does under the bonnet, and why building one well is harder than it looks.

At its simplest, a voice agent is a loop. A person speaks. The system has to hear that speech, understand it, decide what to say back, and then say it — all fast enough that the conversation feels natural rather than stilted. In practice that loop breaks down into several distinct stages, each handled by a different kind of model or service.

The first stage is speech-to-text, often shortened to STT. This is the technology that listens to audio and produces a written transcript. Good speech-to-text has to cope with accents, background noise, overlapping speech, mumbling, and people who trail off mid-sentence. It also has to do this in real time, word by word, rather than waiting until someone has finished a whole paragraph. Deepgram’s core business sits here, and it is a genuinely hard problem: a transcript that is 95 percent accurate can still be useless if the missing 5 percent includes the date, the amount, or the name that the whole conversation depends on.

The second stage is the language model, the LLM. This is the brain of the operation. It takes the transcript, along with conversation history and any instructions it has been given, and decides what should happen next. That might be a spoken reply, a database lookup, a booking action, or a decision to transfer the call to a human. The language model is where the reasoning happens, but it works in text, not in sound. It does not hear tone of voice, hesitation, or background noise unless someone explicitly passes that information along — and that, as we will see, is exactly the gap the new Real-Time Context Standard is trying to close.

The third stage is text-to-speech, or TTS. This turns the language model’s written reply back into spoken audio that sounds like a person rather than a robot. Modern text-to-speech has improved dramatically: voices can now convey warmth, urgency, or a light touch, and can be tuned to a brand. But the voice still has to be generated quickly, in small chunks, so that the reply starts playing before the whole sentence has been composed. Otherwise every exchange gains an awkward pause.

Between and around those three stages sits the orchestration layer, and this is where Pipecat lives. Orchestration is the unglamorous but essential work of keeping everything in sync. It manages the live audio and video streams, decides when the system should listen and when it should speak, handles interruptions — the moment a caller says “actually, wait” halfway through an answer — and routes each piece of the pipeline to the right model at the right moment. Without good orchestration, you get the classic failure modes of early voice bots: talking over the caller, long silences, answers that arrive too late to be useful, or a system that cannot recover when the caller changes topic mid-sentence.

Latency is the word engineers use for delay, and in voice it is everything. In a text chat, a two-second wait is barely noticed. In a spoken conversation, a two-second gap feels like the other person has stopped listening. A gap of half a second can already feel slightly off. So the whole pipeline — speech-to-text, language model, text-to-speech, and the network carrying the audio — has to be tuned as one system, not as separate parts that each add their own delay. Pipecat’s appeal is that it reduces that latency by managing the streams and the pipeline pieces together, rather than leaving each developer to stitch them together from scratch.

Daily’s other contribution is the real-time network itself. When a voice agent answers a phone call or joins a video call, the audio has to travel reliably between the caller, the agent, and the various models doing the work. Dropped packets, jitter, and unstable connections all show up immediately as garbled speech or frozen audio. Daily provides hosted infrastructure for voice and video calls — the pipes, in effect — built for exactly this kind of low-delay, always-on traffic. Pair that with Pipecat’s orchestration and Deepgram’s speech models, and you can see the shape of the combined offer: the framework, the network, and the models, designed to work together.

That brings us to the Real-Time Context Standard. Today, when different models in a voice pipeline hand work to each other, they mostly hand over words. The transcript goes from the speech recogniser to the language model, and the reply goes from the language model to the voice synthesiser. What gets lost in that handover is everything that is not a word: the tone of the caller’s voice, whether they sound frustrated or relaxed, who is speaking when more than one person is on the call, background sounds that might matter, and guidance about how the response should sound. Deepgram says RCS is designed to carry that metadata alongside the words — context drawn from the audio itself, such as a whisper, instructions for the language model on how to use those signals, and guidance on how the response should sound. If it works as described, and if other companies adopt it, it could make voice agents noticeably more responsive to how something is said, not just what is said. That is a meaningful step towards conversations that feel human, rather than conversations that merely transcribe correctly.

It is also worth being clear about what RCS is not. It is not a new speech model, and it is not a finished industry standard on day one. It is an open protocol announced by one company, co-developed with the Pipecat team. Its value will depend on whether other model providers, framework builders and platform companies choose to support it. Open protocols succeed when they solve a shared problem and when enough parties adopt them that supporting them becomes the path of least resistance. That adoption question is one of the key things to watch, and we return to it below.

Why Deepgram is buying rather than building

Every acquisition invites the same question: why buy this instead of building it yourself? In this case there are several plausible answers, and they reinforce each other.

The first is open source as strategy. Deepgram could have built its own orchestration framework, and it could have kept it proprietary. But Pipecat already exists, already has a community of developers building on it, and already carries the credibility that comes from being open and vendor-neutral. Buying Daily gives Deepgram stewardship of that community and that credibility in one move. It is much harder, and much slower, to build a developer community from zero than to join one that already trusts the project. The appointment of Kramer as Head of Open Source, and the establishment of an independent Pipecat Foundation with contributing organisations, are both designed to reassure that community that the project will not simply become a Deepgram sales channel. Whether that reassurance holds will be judged by governance decisions and by what happens to contributors who build competing products on Pipecat — but the intent, at least, is visible.

The second answer is vendor neutrality as a selling point. Stephenson has argued that Deepgram wants developers to use the best models for their use case, not necessarily Deepgram’s own models. That sounds generous, but it is also commercially sensible. Developers are wary of being locked into a single provider’s models, pricing, and roadmap. A framework that lets you plug in different models — swapping a speech recogniser here, a language model there, a voice there — is more attractive than a closed stack, because it preserves choice. If Deepgram owns the neutral layer that everyone builds on, it benefits even when a developer chooses a rival’s model for one part of the pipeline, because the developer is still inside Deepgram’s orbit, and still likely to use Deepgram models where they are genuinely the best fit. It is the classic platform play: own the layer that connects everything, and the individual components matter less.

The third answer is consolidation. Voice AI is going through a period of rapid investment and rapid bundling. TechCrunch’s reporting notes that investment is hot on both sides of the market: on the model side, with companies such as ElevenLabs, Gradium and Fish Audio, and on the infrastructure side, with companies such as LiveKit and Vapi raising multi-million rounds this year. When a market moves that fast, the companies that can offer a complete, working stack — rather than a single component that a customer must integrate themselves — have an advantage, especially with larger customers who want reliability and support rather than a do-it-yourself project. Deepgram’s Voice Agent Platform is a direct response to that pressure: framework plus network plus models, presented as one production-grade product.

The fourth answer is the shared Y Combinator history. Both companies came through the same 2016 batch, which means their founders have known each other and each other’s work for a decade. Acquisitions between companies with that kind of shared background often move faster and integrate more smoothly, because there is less cultural translation to do. Kramer has explicitly pointed to matching engineering culture as part of the attraction, alongside the open-source commitment. Culture is easy to dismiss as PR language, but in an acquisition where the asset being bought is largely a team and a community, culture is the asset. If the Daily team leaves, Pipecat’s momentum leaves with them. Keeping the team, and keeping them motivated, is therefore not a soft concern — it is the deal.

There is a fifth, quieter answer: expressivity and end-to-end speech. Stephenson has described the acquisition as letting developers build agents with expressivity and end-to-end speech systems, plugging in different models, with a context bus connecting the parts. In plain English, that means voice agents that do not just get the words right but get the delivery right — pace, tone, emphasis, and the ability to respond to the emotional texture of a conversation. That is where the next round of differentiation in voice AI is likely to come from. Transcription accuracy is improving across the board and becoming less of a differentiator on its own. What callers notice is whether the agent sounds like it is listening. Deepgram is betting that owning the orchestration layer and the context protocol is the way to win that next round.

None of this means the strategy is risk-free. Buying a community is not the same as keeping it. Developers who chose Pipecat partly because it was independent may watch closely for signs that Deepgram’s interests start to override the project’s neutrality. Rival model providers may be less enthusiastic about plugging into a framework owned by a competitor. And integrating a real-time network, a framework, and a set of speech models into a single platform is genuinely hard engineering work, not a rebrand. But as a statement of intent, the deal is coherent: Deepgram is choosing to compete on the whole stack, on openness, and on developer trust, rather than on a single model alone.

What it means for UK small businesses

For a small business owner in the UK, an acquisition between two voice AI infrastructure companies might sound remote. It is not. It sits directly upstream of the phone-answering services, booking assistants, and customer-service agents that small firms are increasingly being offered — and increasingly being asked to pay for. Understanding this deal helps you judge those offers more clearly, and ask better questions before signing anything.

Start with the most immediate effect: better voice agents should get cheaper and easier to supply. Right now, building a voice agent that handles real phone calls reliably is a specialist job. The supplier has to assemble speech recognition, a language model, voice synthesis, call handling, and orchestration, and then keep all of it running without awkward pauses or dropped calls. When more of that plumbing is available as a shared, open framework, and when the network and models are designed to work with it, the cost of building a decent agent falls. Suppliers can spend less time reinventing infrastructure and more time on the part that actually matters to your business: understanding your services, your booking rules, your tone of voice, and the questions your customers actually ask.

In practical terms, that should show up in a few places. First, phone answering. A voice agent that can answer out-of-hours calls, take a message accurately, or handle a straightforward booking without putting the caller on hold is already valuable for trades, clinics, salons, restaurants, and professional services. The current generation is usable but uneven: fine on simple queries, shaky on anything with a twist. A more mature orchestration layer and better handling of audio context — tone, interruptions, background noise — should make these agents noticeably more reliable on real calls, not just in demos. That matters because a missed call is often a lost customer, and a bad automated experience can be worse than no answer at all.

Second, bookings and routine customer service. The highest-value use of a voice agent for most small firms is not replacing humans entirely; it is handling the repetitive, well-defined work so humans can focus on the complicated cases. Confirming an appointment, rescheduling within your rules, answering opening hours and pricing questions, taking a delivery order with standard options, logging a support request with the right details — these are all tasks where a voice agent can be genuinely useful, provided it knows when to hand over. The combination of Pipecat-style orchestration with a real-time network should make handover smoother too: a call that starts with an agent and moves to a person should not feel like starting again from scratch.

Third, cost and choice. This is where the open-source and vendor-neutral parts of the story become practical rather than abstract. If the agent you buy is built on an open framework, and if its components can be swapped — a different speech model here, a different voice there — then you are less dependent on a single supplier’s pricing and roadmap. That does not mean you can switch suppliers for free: your prompts, your integrations with your booking system, and your call flows still represent real work. But it does mean the underlying technology is less of a trap. A supplier who builds on open components has less leverage to raise prices sharply, because a competitor can build on the same components. When you are evaluating suppliers, that is worth asking about explicitly, and we list the questions below.

Fourth, voice quality and expressivity. Customers judge a voice agent within seconds, mostly on how it sounds. A flat, robotic voice that barrels on regardless of what the caller just said will lose trust immediately. A voice that pauses naturally, acknowledges an interruption, and adjusts its tone when a caller sounds frustrated will keep people on the line longer. The RCS protocol, if adopted, is aimed at exactly this: giving the system the context to respond to how something is said, not just what was said. For a small business, that is the difference between an agent that feels like a phone menu with a voice attached, and one that feels like a competent receptionist. It will not arrive overnight, and it will depend on suppliers actually implementing it, but the direction is encouraging.

There are also risks, and a careful buyer should weigh them alongside the benefits. The first risk is concentration. When infrastructure consolidates — when one company owns the framework, the network, and a leading set of models — that company gains influence over pricing, features, and the direction of the ecosystem. Open source mitigates this, but it does not eliminate it. A foundation and a licence are only as strong as the governance behind them, and a small business cannot audit that governance itself. The practical protection is to avoid deep dependence on any single stack: keep your call scripts, your customer data, and your booking integrations in a form you can move, and be wary of any supplier who cannot explain what happens to your system if they change provider or go out of business.

The second risk is overpromising. Voice AI demos are famously flattering. They use quiet rooms, clear speakers, and cooperative questions. Real calls involve accents from across the UK, mobile connections from moving cars, background noise from workshops and kitchens, and customers who change their minds halfway through a sentence. Any supplier who promises that an agent will handle everything, or who cannot show you recordings of difficult calls, should be treated with caution. Ask for a trial on your actual phone line, with your actual customers’ kinds of questions, before committing to a long contract.

The third risk is data and compliance. Voice calls contain personal data, and in the UK that brings GDPR into play alongside sector rules where they apply. A voice agent supplier should be able to tell you, in plain language, where call audio is processed, how long recordings and transcripts are kept, whether data leaves the UK or the European Economic Area, and how a customer can request deletion. The fact that a framework is open source does not automatically answer those questions: the framework is code, but the hosted service that runs it is where your data actually goes. Do not let a supplier hide behind the word “open” when you ask about data handling.

The fourth risk is cost structure. Voice agents are typically priced per minute, per call, or per month with usage allowances, and the true cost depends on how long calls last, how many need human handover, and how much customisation you require. A cheaper headline price can become expensive if every call runs long because the agent is slow to understand, or if you pay separately for every integration. Ask for a total cost based on your actual call volume, not a generic tier, and ask what happens when you exceed it.

So what should you actually ask a supplier who tells you their voice agent uses Pipecat, Deepgram, or the new Voice Agent Platform? Here are questions that will quickly separate a thoughtful supplier from one selling a demo:

• Which parts of the system are open and swappable, and which are tied to a single provider? If we wanted to change speech model or voice provider later, what would that involve?
• Can you show us the agent handling a difficult call — an accent, background noise, an interruption, a change of mind — not just a clean demo?
• How does handover to a human work, and what does the human see when they pick up? Does the caller have to repeat everything?
• Where is call audio processed and stored, for how long, and how do we handle deletion requests under GDPR?
• What is the total monthly cost at our actual call volume, including handover, integrations, and any per-minute charges?
• What happens to our call flows, scripts, and data if we leave, or if you change your underlying provider?
• How do you test accuracy on the specific things our business cannot get wrong — dates, amounts, names, and booking rules?

A supplier who answers those clearly, with examples from businesses like yours, is worth talking to. A supplier who waves them away with “the AI handles it” is not. The Deepgram-Daily deal does not change that basic discipline, but it does make it more likely that good answers will be available at a price a small firm can afford, because the underlying building blocks are becoming shared, open, and better integrated.

One more point for UK firms specifically: voice is a local problem as much as a technical one. A system trained mostly on American English will struggle with regional UK accents, place names, and the particular ways British customers phrase requests — “I was wondering if you might have anything on Thursday” is not the same as “book Thursday.” Suppliers who have tuned their agents on UK calls, and who can demonstrate that tuning, will outperform generic systems regardless of which framework they use. When you evaluate a voice agent, test it with your customers’ voices, not with a polished demo voice.

There is also a useful parallel here with how small firms have adopted other technologies. Websites, card payments, and online booking all started as specialist, expensive services and became commodities as shared platforms matured. Voice agents are on a similar path, and this acquisition is a step along it: the infrastructure is consolidating and opening up at the same time, which usually means better products at lower prices within a year or two. That does not mean every small firm should rush to adopt a voice agent today. It does mean that if you have been put off by cost or quality, it is worth revisiting the question in the coming months, and worth choosing a supplier whose system is built on components you could move if you needed to.

For agencies and freelancers who build for small firms, the implication is similar but sharper. If you are selling voice agent builds, the stack you choose now matters. Building on an open framework with a real-time network and swappable models gives you flexibility and a story you can tell clients about avoiding lock-in. It also means you need to stay close to how Pipecat evolves under Deepgram’s stewardship, and to whether RCS gains traction beyond Deepgram’s own products. An agency that can explain that landscape to a client in plain English — without hiding behind jargon — will have an edge over one that simply resells a single vendor’s black box.

Honest limits / what to watch

It is tempting to read an acquisition announcement as a finished story. It is not. The most important questions about this deal are still open, and they will be answered by execution, governance, and market response over the coming year, not by press releases.

The first thing to watch is the open-source promise, and how it is tested. Deepgram has said Pipecat remains fully open source under its BSD-2 licence, and that an independent Pipecat Foundation is being established with contributing organisations. Those are strong statements, but open-source communities judge promises by actions: who sits on the foundation’s board, how decisions are made, whether contributors outside Deepgram can still shape the roadmap, and whether the project remains genuinely usable with rival models. If Pipecat continues to welcome contributions from companies that compete with Deepgram, and if the foundation has real independence, the promise will be believed. If governance quietly centralises, or if new features start to work best only with Deepgram models, trust will erode — and developers have shown repeatedly that they will fork a project or move to an alternative when they feel that trust has been broken. The BSD-2 licence matters here: it is a permissive licence that allows the code to be used, modified, and redistributed, which gives the community a genuine fallback if stewardship disappoints. But a fork only succeeds if enough developers move with it, so the community’s response in the first six months will be telling.

The second thing to watch is integration execution. Announcing a Voice Agent Platform that combines a framework, a network, and speech models is one thing; shipping a platform where those pieces genuinely work together, with documentation, support, and reliability that production customers can depend on, is another. Integration projects of this kind often take longer than announcements suggest, and early versions can be rough around the edges. Developers evaluating the platform should look for concrete evidence — stable releases, migration guides for existing Pipecat and Daily users, and clear statements about what happens to existing Daily services — rather than assuming the combined product is ready on day one.

The third thing to watch is competition. Deepgram is not acquiring into an empty market. ElevenLabs, Gradium, and Fish Audio are pushing hard on the model side, and LiveKit and Vapi have raised significant rounds on the infrastructure side this year. LiveKit in particular offers a real-time infrastructure alternative, and Vapi offers a voice agent platform that competes directly with what Deepgram is building. That competition is good news for buyers, because it keeps pressure on pricing and features, but it also means Deepgram’s platform will be judged against credible alternatives, not in isolation. If a rival offers a better developer experience or a more convincingly neutral platform, developers will vote with their projects.

The fourth thing to watch is whether the Real-Time Context Standard gets adopted beyond Deepgram. An open protocol announced by a single company is a proposal, not a standard. Its fate depends on whether other framework builders, model providers, and platform companies see value in supporting it, and whether the Pipecat Foundation helps govern it in a way that invites participation rather than control. If RCS remains a Deepgram-only feature, it will still be useful to Deepgram customers, but it will not change the industry. If it gains broad support, it could become a meaningful piece of shared infrastructure. Watch for adoption announcements, for support in other frameworks, and for whether rival model providers implement it — those signals will tell you more than the launch announcement alone.

The fifth thing to watch is the undisclosed price and what it implies. Because the price has not been shared, outsiders cannot tell how big a bet Deepgram has made, or what return it needs to justify it. A large acquisition creates pressure to monetise quickly, which can strain an open-source commitment. A smaller acquisition leaves more room for patience. Neither scenario is inherently good or bad, but the absence of a price means the financial logic of the deal is opaque, and observers should be cautious about claims — from either supporters or sceptics — that depend on knowing what was paid. The same caution applies to any future claims about customer numbers or performance that have not been published with evidence.

There are also broader questions that no single deal can answer. Voice AI raises real issues around consent, recording, and transparency: callers should know when they are speaking to an agent, and businesses should be clear about that rather than trying to pass an agent off as human. Regulation in the UK and the EU continues to evolve around AI transparency and data protection, and suppliers will need to keep pace. The technology is also only as good as its handling of edge cases — strong accents, speech impairments, distressed callers, and safeguarding situations where a human must take over immediately. Those are not footnotes; they are the cases where a voice agent’s quality matters most, and where a buyer should demand the clearest answers.

For developers, the practical advice is straightforward: keep building on Pipecat if it serves your project, watch the foundation’s governance, and avoid designing your system so that it only works with one provider’s models. For small businesses, the advice is even simpler: do not buy on the basis of an acquisition headline. Buy on the basis of a trial, on real calls, with clear answers on cost, data, handover, and portability. The Deepgram-Daily deal makes the voice AI stack more interesting and potentially more accessible. It does not remove the need for due diligence, and it does not guarantee that any particular supplier’s agent will work well for your business.

Stepping back, the deal is best understood as part of a larger shift. Voice AI is moving from a collection of impressive demos and separate components towards integrated, production-ready platforms. The companies that win that shift will be those that combine good models with good infrastructure and earn developer trust along the way. Deepgram’s bet is that openness is the way to earn that trust, and that owning the orchestration layer is the way to benefit from it even when developers choose rival models for parts of the pipeline. It is a coherent bet, and the shared history and engineering culture behind it give it a better chance than most. But bets are judged by outcomes, and the outcomes here — governance, integration, adoption, and competition — are still to come. That is what makes this a story worth following, rather than a conclusion to file away.

Key facts

For more on how voice AI is being used in practice, see our coverage of Block’s Dessa voice AI and human conversation and Nvidia’s Nemotron Bolt E4B for local voice agents.

Get a Free AI Consultation

FAQs

What is Deepgram acquiring in the Daily deal?

Deepgram is acquiring Daily, including its open-source voice AI framework Pipecat. Daily’s team joins Deepgram, and Pipecat will remain fully open source under its existing BSD-2 licence, with an independent Pipecat Foundation being established.

Will Pipecat stay open source after the acquisition?

Yes, according to both companies. Pipecat remains fully open source under its BSD-2 licence, and Daily co-founder Kwindla Kramer is joining Deepgram as Head of Open Source. The test for that promise will be how the project is governed and maintained over the coming months.

What is Deepgram’s Real-Time Context Standard?

The Real-Time Context Standard, or RCS, is a new open protocol co-developed with the Pipecat team. It is designed to carry audio context that words alone do not capture, such as tone, who is speaking, background sounds, and instructions on how a response should sound.

What does this mean for a UK small business?

In practical terms, it should make capable voice agents for phone answering, bookings and customer service cheaper and easier to build and buy, because more of the underlying plumbing is shared and open. It does not remove the need to check accuracy, costs, data handling and what happens if a supplier changes direction.

Should a small business build on Pipecat directly?

Most small firms should not build on it directly. Pipecat is a developer framework, so the realistic route is to buy a finished service from an agency or supplier that uses it, and to ask that supplier how portable the system is if you ever want to switch models or providers.