Quick answer: Video API integration projects often fail for the same few reasons. Teams underestimate the real-time layer behind WebRTC (NAT traversal, signalling and media servers), hit scaling and reliability limits under load, get caught out by browser and device differences and bolt security and GDPR on too late. The demo usually works with two people on one office network, and the trouble starts when real users, real devices and a compliance review arrive. You can de-risk it by deciding early whether to build the real-time layer yourself or embed a prebuilt one, and by planning for scaling, data residency and ongoing maintenance from day one.
Video API integration projects fail when a team treats real-time video as a normal web feature rather than a piece of live infrastructure. A one-to-one call between two laptops on the same network is straightforward. It gets hard once you add restrictive firewalls, mobile networks, large rooms, mixed browsers, recording and a legal requirement to know where the data is stored. Each of those is a separate engineering problem, and a project that budgets for the first call but not for the rest often stalls somewhere between the prototype and production.
Seven causes come up again and again, and all of them are easy to miss in a quick demo.
WebRTC gives the browser the building blocks for peer-to-peer audio and video, but it leaves the coordination to you. It does not define how two peers find each other: as Mozilla's documentation puts it, 'WebRTC doesn't specify a transport mechanism for the signaling information', so you have to build and run a signalling channel to exchange session details. For the basics of the technology, see our explainer on how WebRTC works.
On top of that, two users can rarely connect directly. WebRTC uses ICE, STUN and TURN to get through Network Address Translation and firewalls. STUN helps each side find its public address, ICE tries the possible routes between the two users, and TURN is a relay server that passes the media along when they cannot connect directly. You have to provision and secure that TURN server, and you pay for its bandwidth. Teams that assume a video SDK handles all of this often find the gap only when calls fail silently for users behind strict corporate networks.
A call that works for 4 people in a peer-to-peer mesh can collapse at 14. In a mesh, every participant sends their stream to every other participant, so each participant's upload bandwidth and CPU use climb sharply as the room grows. Moving beyond small calls means introducing a media server such as a Selective Forwarding Unit, which receives each stream once and forwards it on. That fixes the upload problem, but each participant still downloads every stream they display, so large rooms also need measures such as sending each stream at several quality levels (simulcast). Growing past small calls also means capacity planning, geographic distribution and failover, all of which you have to keep doing for as long as the product runs. Reliability problems often surface late, because they only appear under concurrency and network conditions that a small test never reproduces.
WebRTC is a standard, but the way browsers and devices implement it still varies. Codec support, mobile behaviour, in-app webviews and older devices all differ, and MDN's browser compatibility tables show which browsers support each feature, which is why a feature can work in one environment and fail in another. A lot of integration time goes into handling these differences and into testing across the full matrix of browsers, operating systems and hardware your users have, rather than the handful your developers use.
Compliance is where European products often stall, because it is treated as a last-minute box to tick rather than something that shapes the design. Encryption in transit and at rest, access controls and a clear answer to where media and metadata are processed all need to be decided up front. Under the GDPR, transfers of personal data outside the EU and EEA are only lawful if the conditions in Chapter V are met, such as an adequacy decision or standard contractual clauses. A stack that routes media or recordings through servers elsewhere therefore adds transfer work and can fail a data protection review late in the project. Keeping everything inside the EU and EEA is the simplest route, though not the only lawful one. The GDPR also expects data protection by design and by default, which is far cheaper to honour when it shapes the architecture than when it forces a rebuild weeks before launch.
A common cause of overruns is estimating the cost of the first working call instead of the cost of a production service. Building makes sense when real-time video is your core product and you need full control. If video is only a feature of a larger product, building it yourself takes far more effort and time than the initial API call suggests.
A video integration is never really finished, because browsers change, codecs evolve, security patches land and scaling needs shift as usage grows. A self-built stack carries all of that maintenance for the life of the product, and the people who understand the real-time internals become a dependency the business cannot easily replace. Embedding moves the infrastructure upkeep to the provider, but your own app still needs maintaining. If you budget only for the initial build, with nothing set aside for upkeep, quality drifts, edge-case bugs pile up and nobody wants to touch the integration any more.
Without call-level visibility, you cannot see why sessions drop, where quality degrades or how features are used. If you do not plan for session events, quality metrics and usage data, you end up guessing at the cause and arguing about problems you cannot measure. Good session analytics have to be designed in, because they are very hard to reconstruct afterwards.
At Digital Samba, our support team relies on this kind of data every day. When a customer reports poor call quality or dropped connections, our call statistics show whether the problem is on our side or somewhere between the end user and us, such as their network, computer, internet provider or an internet exchange. More often than not the cause is outside the platform, and the data lets us point the customer to the specific things to check.
Most of these causes trace back to whether you build the real-time layer or embed a prebuilt one.
Building it yourself means taking on signalling, NAT traversal and TURN relays, media servers, scaling, recording, device testing, security and compliance as your own long-term responsibilities, along with maintaining all of it. You get full control, which matters when real-time video is the product itself, and at very high volumes running your own infrastructure can work out cheaper than paying per minute.
Embedding a prebuilt platform means the real-time layer already exists and the provider runs it, so your team integrates a tested component instead of standing up live infrastructure. The provider then handles signalling, NAT traversal and TURN, media servers, scaling and failover, and keeps that layer maintained. Device testing, compliance review and instrumentation stay with your team either way. Embedding does not make real-time video simple, but it moves the hardest infrastructure work to the provider. Embedding has its own costs too. You depend on the provider's pricing, roadmap and uptime, you have less control over the interface and features than with your own build, and per-minute pricing adds up as usage grows. It is worth checking the provider's exit and data-export terms before you commit. Even so, for most products where video is a feature rather than the whole business, embedding is the lower-risk path to production.
Here is a short checklist that heads off the common failures. If you embed, the provider runs TURN, scaling and failover, so those items become questions to ask the provider rather than work for your team.
Digital Samba is a video conferencing platform for teams that would rather not build the real-time layer themselves. You get a prebuilt SDK, and we run the signalling, the TURN relays, the media servers and the scaling, so your team integrates a tested layer.
For European products, an EU-based provider makes residency simpler. You are still the controller of your users' data, so you need a data processing agreement and a review of sub-processors, as with any provider. We are a European company (Digital Samba S.L., based in Barcelona), and the service runs on European infrastructure, with no US hyperscaler in the media path. Optional end-to-end encryption is available for calls where no server, including ours, should be able to read the media. Encryption is switched on when the room is created, not during a call.
If you want to understand the privacy side in more detail, see our guide to video API security and privacy risks in Europe and the embedded video API and SDK built in Europe. There is a free plan, called Start, to build on; the current capacities and tiers are on the Digital Samba Embedded pricing page.
They fail because teams treat real-time video as a simple API feature instead of live infrastructure. The first call is easy, but scaling, NAT traversal, device differences, security and GDPR are separate engineering problems that surface after the prototype. A project that budgets only for the demo often stalls before production.
It depends on whether real-time video is your core product or a feature. Building gives you full control but means owning signalling, media servers, scaling, testing, security and maintenance for the life of the product. With embedding, the provider handles the infrastructure side (signalling, TURN, media servers and scaling), which is usually the lower-risk route when video supports a larger product. Device testing, compliance review and instrumentation are yours whichever route you take.
WebRTC provides the browser building blocks but leaves the coordination to you. It does not define a signalling mechanism, and connecting two users usually needs ICE, STUN and TURN to get through firewalls and NAT. When a direct path is blocked the media has to be relayed, and all of those pieces have to be built, run and maintained.
Decide early where media, recordings and metadata are processed, and keep personal data inside the EU and EEA if that is your requirement. An EU-based provider simplifies the residency question, but you remain the controller, so you need a data processing agreement, a review of the provider's sub-processors and a check of any other tools in your stack that handle personal data. Treat data protection as an architecture decision rather than something to retrofit.
Small tests rarely reproduce real concurrency and network conditions. A peer-to-peer mesh that works for a few people overloads as the room grows, which is why larger calls need a media server to forward streams. Reliability issues appear under load and on varied networks and devices, and a quick two-person test exercises none of them.
This is an educational explainer for product and engineering teams evaluating how to add video to their software. It is a practical guide, not legal advice; confirm specific compliance requirements with your own data protection team.