SumatoSoft

Voice AI just had its plumbing year 🎙️

Something happened to voice AI in the last twelve months that made it production-grade for enterprise deployments. 🏭

The core capability of speech-to-text plus LLM plus text-to-speech has existed for several years. What changed is the orchestration layer between them. Latency dropped from 3 to 4 seconds to under 800 milliseconds for a full conversational turn. Interruption handling, the thing that made voice agents sound robotic, became standard. Tone and emotion in synthesised speech crossed the uncanny valley for most use cases. Voice infrastructure platforms such as Vapi, Retell, and LiveKit shipped the production tooling that turns these capabilities into a deployable service. ⚙️📉

The deployment patterns we see this year: customer service IVR replacement (the largest category, with carrier-to-carrier check calls being a logistics-specific example), outbound qualifying calls for sales operations, appointment scheduling and confirmation, and field-service follow-up calls. None of these require frontier reasoning. They require reliable conversational flow under cost and latency budgets that match human-call economics. 📞✅

The question worth answering before scoping a voice project: which calls in your operation follow a script with predictable variations? Those calls are now automatable. The calls that don't follow scripts (relationship management, complex troubleshooting, anything where tone and judgment carry the conversation) remain human work for now. The economics of voice AI deployments only land when you sort calls correctly into the two buckets. 🧠🗂️

When plain RAG breaks ⚙️

A pattern across enterprise AI projects this year: vanilla RAG works for the first 80 percent of queries and breaks on the next 15. The break is structural, and a class of fixes called "graph RAG" has emerged to handle it. 📉

Plain retrieval-augmented generation works by chunking your documents, embedding the chunks, and retrieving the closest chunks to the query. This works when the answer to a query lives in a single document or a handful of contiguous chunks. It breaks when the answer requires connecting facts across documents that don't sit close together in embedding space. 🔗❌

The classic example: "Which of our suppliers have had compliance incidents and also serve customers in regulated industries?" That answer requires a multi-hop traversal. Suppliers, then incidents, then suppliers again, then customers, then customers' industries. No chunk contains the full answer. Embedding similarity won't find it. 🧩

Graph RAG approaches solve this by building a knowledge graph from the source documents (entities, relationships, attributes), then querying the graph alongside or instead of the embedding store. Microsoft published a notable graph RAG paper last year, and several open-source frameworks now implement the pattern. Production deployments are landing across legal, healthcare, supply chain, and financial services in 2026. 📊✅

Where AI vendor lock-in lives 🔒

A Docker survey published this month found that 76-81% of enterprises are concerned about vendor lock-in in their agentic AI deployments. The number is striking. The placement of the lock-in is even more striking, and most procurement teams are looking in the wrong place. 🎯

When companies evaluate AI vendors, attention concentrates on the model: GPT versus Claude versus Gemini, capabilities, and pricing per token. This is the most visible layer and also the easiest to swap. Models are increasingly commodity-like, with Stanford's 2026 AI Index showing the top frontier models separated by razor-thin margins. 🤖🔄

Lock-in sits one layer below, at the orchestration and memory layer, where vendors build their proprietary surface. How agents persist state between sessions. How they pass context to one another in multi-agent workflows. How retrieval and grounding are configured against your data. How permissions and audit trails are stored. Migrate the model, and these survive. Migrate the orchestration platform, and most of this gets rebuilt from scratch. Orchestration vendors are happy to be flexible about which model you use, precisely because the model is not where they have you. 🧠⚙️

The architectural choice that determines your lock-in exposure five years from now is not which LLM you sign with this quarter. It is whether your orchestration, memory, and retrieval layers run on open standards such as MCP, A2A, and vector stores you control, or on a vendor's proprietary stack. 🏗️

We have started explicitly writing this question into our discovery process. Most clients have not thought about it because no one in the sales conversation is incentivized to raise it. If you are signing an AI platform contract this quarter, the most important paragraph in the document is the one describing data and state portability at exit. Read it twice. 📄👀

What MCP means for your enterprise AI stack 🔌

Lucidworks released an MCP server three weeks ago, claiming reductions of up to 10x in enterprise AI integration timelines and savings of over $150,000 per integration. Vendor-reported numbers deserve scrutiny, but the underlying shift is worth understanding regardless of how generous Lucidworks's calculator is. 📊

Model Context Protocol is an open standard, originally developed by Anthropic and now stewarded by the Linux Foundation, that defines how AI models discover and call external tools. It is running on more than 10,000 public servers as of this month, with adoption from OpenAI, Google, Microsoft, and AWS. The protocol is on track to become the universal interface between AI agents and enterprise systems. 🔧
For most companies, several things follow from this. The "let's build a custom integration so our chatbot can read from Salesforce" project that you scoped six months ago has become substantially smaller, and a vendor that is not using MCP should be asked why. ❓ The economics of multi-system agents shift in the same direction. Most agentic deployments stall because connecting an agent to five enterprise systems used to require five custom adapters with five different security models, but MCP collapses that into a single pattern. 🔄

New attack surface is a counterweight worth taking seriously. Asana had an MCP-related tenant-isolation flaw earlier this year that affected up to 1,000 enterprises, and WordPress plugins exposed more than 100,000 sites. The standard arrived faster than the security tooling around it, so MCP rollouts in the next two quarters need security reviews scoped accordingly. 🔒

We have moved our AI integration practice to MCP-first as the default architecture, with custom adapters reserved for legacy systems that cannot expose what MCP needs. 🏗️ If you are scoping AI agent work this quarter, this is the architectural decision that will look obvious in eighteen months and expensive to undo. ⏳