Communication tooling is one of the easiest places to see what generative AI can and cannot do today. Transcription, summarisation, tone analysis, and translation work well enough to deploy. Live “read the room” capabilities — microexpression analysis, real-time sentiment interpretation, cross-cultural coaching — sit in a different category. They demo well; they don’t yet hold up as decision-grade signals. That gap matters because most failed generative AI investments in the communication space were not engineered badly. They were scoped to capabilities the underlying models do not reliably have. In our experience, the difference between a project that ships and a project that burns budget is rarely the team — it is whether someone classified the use cases honestly before the build started. This article walks through where AI genuinely transforms communication today, where the technology is still speculative, and how to tell the two apart. For a structured way to apply that distinction to your own pipeline, our hub on how to evaluate whether a generative AI use case is technically feasible covers the framework end-to-end. What can generative AI actually do for communication right now? Three capability tiers are useful when looking at any communication-focused proposal. They map directly onto the automatable / speculative / research classification used in a structured feasibility assessment. Capability tier Examples Feasibility today Automatable Transcription, summarisation, translation, draft generation, tone classification of written text High — production-ready with current LLMs and ASR stacks Bounded research Sentiment trend tracking, meeting analytics, multilingual content adaptation, soft-skill coaching prompts Medium — works in narrow scope, needs evaluation and guardrails Speculative Microexpression-based deception detection, “true intent” inference, cross-cultural automatic coaching at scale Low — observed-pattern across our engagements is that accuracy claims do not survive controlled testing The first row is where most return shows up. Transcription paired with an LLM summary is reliable enough that you can build operations around it. Translation between widely-resourced languages is at a quality where it removes friction in distributed teams without needing a human in the loop for every message. The second row needs a bounded research phase before you commit to a build. The models behave well in narrow conditions and degrade in ways that are not obvious from a demo. The third row is where buyers most often get hurt — the capability sounds plausible, vendors will sell it, and the published evidence for it is thin. Real-time transcription and summarisation This is the most defensible bet in communication-focused AI today. Modern ASR systems handle accented speech, overlapping speakers, and domain vocabulary well enough that the transcript is the source of truth for most downstream work. LLMs on top extract action items, decisions, and risks with high enough accuracy that meeting summaries are now an operational artifact rather than a curiosity. Two technical components matter here. Whisper-class models or commercial equivalents (AssemblyAI, Deepgram, Azure Speech) handle the audio. A retrieval layer plus a constrained-output LLM call handles the summarisation. Both are reproducible benchmarks: word error rate is published, summary quality can be measured against human-written reference summaries on your own corpus. The point is that you can write go/no-go criteria for this before you build. “WER under 8% on a held-out set of 200 of our actual calls” is a defensible acceptance threshold. We use exactly this kind of measurable outcome on engagements where the buyer needs a record of why the spend is justified. Translation and language access Machine translation moved from “helpful but check it” to “publish-grade for most internal communication” over the last few years. The mechanism is unchanged — transformer-based sequence models trained on parallel corpora — but the quality crossed a usability threshold for high-resource language pairs. Where this stays interesting is the long tail. Low-resource languages still degrade noticeably. Domain-specific terminology (legal, medical, technical) needs either fine-tuning or retrieval-augmented translation to stay accurate. If your use case sits in the long tail, it is a research question, not a procurement question. For most global teams, though, the operational reality is that meeting transcripts, chat messages, and product documentation can be translated automatically with acceptable quality. This is where measurable impact tends to show up first in deployment metrics: time-to-decision in distributed teams drops, and translation cost as a line item collapses. Customer support: where the ROI is real and where the failures cluster Generative AI in customer support is the most-pitched use case in the communication space, and also the one where we see the widest gap between vendor claims and what survives in production. Routine query handling — order status, password resets, basic billing questions — works. The model can deflect a real percentage of tickets, and that deflection is measurable. What does not work reliably is the next claim up: handling complex emotional support interactions, recognising when to escalate based on customer sentiment, or maintaining brand-consistent tone across thousands of edge-case conversations. These are observed-pattern failure modes across our generative AI engagements, not benchmarked rates — the specifics depend on your support taxonomy and tooling, but the pattern is consistent. The right scoping question is “what percentage of our ticket volume falls into categories the model can handle?” before “what tooling do we buy?” If 60% of your tickets are routine and the model handles 80% of those, the math justifies the project. If 80% of your tickets are complex multi-turn cases, the project is going to disappoint regardless of the platform. Our companion piece on the power of generative AI in customer service walks through how to classify the ticket distribution before deciding what to build. Body language, microexpressions, and the speculative tier This is the part of the communication-AI conversation where buyers get burned most often. Computer vision models can detect faces, track eye gaze direction, and classify a small set of facial action units. That is real and reproducible. The leap from those primitives to “this person is being deceptive” or “this candidate is showing signs of low engagement” is where the evidence collapses. The published literature on automated affect recognition has well-documented reliability problems — class imbalance, cross-cultural variance, and confounding from lighting and camera angle all degrade accuracy in ways that controlled studies expose and product demos do not. If you are being pitched a tool that promises to read microexpressions for negotiation or hiring decisions, the right question is “show me the held-out test set accuracy on a demographic distribution matching our user base.” The answer usually ends the conversation. This is not an argument against the underlying computer vision. It is an argument for honest classification: gesture and gaze tracking is a research question with bounded use cases (presentation coaching, accessibility tooling). Inferring intent or character from those signals at scale is speculative and should not be funded as a delivery commitment. How to decide what to fund The pattern that holds across the use cases above is the same one our feasibility assessment framework operationalises: Classify each candidate use case as automatable, speculative, or research. For automatable use cases, write measurable acceptance criteria before the build (WER thresholds, deflection rate targets, translation quality on your domain corpus). For research questions, scope a bounded investigation phase with a clear go/no-go decision, not an open-ended build. For speculative use cases, do not commit budget to delivery — fund a literature review or a small pilot that tests the underlying claim, then re-classify. The point of the assessment is not to kill ambition. It is to make sure the budget lands on the work that can ship and that the speculative items get tested cheaply before they consume real engineering capacity. We have seen organisations recover six- and seven-figure GenAI commitments by doing this classification three weeks into a project rather than six months in. FAQ Where this leaves the buyer Most of the value AI delivers in communication today comes from a small number of well-understood capabilities applied to high-volume, repetitive work. The capabilities further up the stack — emotion, intent, cross-cultural nuance — are interesting research, not procurement targets. The defensible move for a buyer who has just approved a GenAI budget is to insist on a feasibility assessment before any build commitment, and to make the assessment itself an artifact: classified use cases, measurable acceptance criteria, named go/no-go thresholds. If the project ships, the assessment justifies the spend. If it does not, the assessment is what kept the loss small. That is the lift this category of work is doing. It is not glamorous. It is what separates GenAI projects that close with a defensible outcome from ones that close with a write-off.
