The AI market has finally come back around to the (correct) conclusion that people are important. Thank goodness. And on that note, The Forrester Wave™: Conversational AI Platforms For Employee Services, Q3 2026 (aka AI to help employees) is finally live! To put it bluntly, the market has progressed further than I thought possible since our first Forrester Wave on the topic in 2019 (under the “chatbots for IT services” umbrella).
Agentic AI agents are now (unsurprisingly) common. AI agents that autonomously decide how to resolve user requests in production. This begs the question: What’s next? And what still matters, if we’ve got these magical AI tools in production? I’ve summarized some of my thoughts on where the market still needs work, and what really matters below:
AI agent/conversational AI governance has come a long way, and has a long way to go. We’ve already come a long way in wrangling LLMs and agentic systems into doing what they’re told. Default guardrails are universal, and testing is almost everywhere. Agent scripting languages, just starting to appear from providers, create more predictable tool calls. Unfortunately, we’ve got a ways to go — some platforms lacked automatic PII redaction; some lack version control. And despite testing being fairly ubiquitous, enforced pre-live testing was absent, so we can still expect entertaining headlines for the foreseeable future. Good news, vendors are aware of this and working to close these gaps.
Implementation assistance should be top of mind for all adopters. AI is easier to adopt than I ever thought would be possible. That doesn’t mean it’s easy to deploy into enterprise production, or connect into your key back-end systems. In fact, integrations were often generously described as “tricky.” Similarly, getting security approvals (appropriately) proved prolonged. Having vendor assistance made this significantly easier for customers — and with maturing customer success motions, these are increasingly accessible for teams of all sizes. Make sure you ask your vendor how they’re going to ensure their AI works for you before committing.
Agent drafting is still generally weak (but getting better). Agents generating better agent prompts has been a massive time and headache saver for builders. Unfortunately, for most, we’re still just at the “prompt generation” phase. Most generative building experiences proved unaware of other existing agents, relevant knowledge, or tools they could employ to improve agent performance from their environment (referred to in the report as agent, tool, and knowledge aware, or ATAKA-capable builders). As agents expand, this “awareness of prior art” is going to be essential to reduce rework, orphaned assets, improving compliance, and reducing attack surfaces. This is getting better, and more providers are expected to follow suit here. Make sure you’re looking to see if in drafting, systems are suggesting existing tools, knowledge, or agents to connect to.
Value is coming quicker (but scaling remains tricky). The reported timeline for value from AI agents continues to shrink. In 2024, customer reported time to value ranged from 6 to 9 months. Now, in 2026, with agentic systems, multiple customers reported going live in under three weeks. Six-month deployments were reported, but the aforementioned internal security approvals and complex integrations were cited as the limiting factors. While usefulness is getting proven faster, scaling beyond initial deployments and premade assets remains tricky. Knowledge, process documentation/process expertise, tool development, and integrations were all cited as complicating factors.
Context graphs (and enrichment) is becoming very important. The good news, no one AI provider expects to be the sole AI platform that customers use — many are actually “decomposing” their platform into a headless system that can be called from wherever the user is. The bad news, this doesn’t inherently help environment fragmentation. Instead, a new front of competition has emerged for vendors to prove their value to organizations: context graphs, aka how well can vendors connect the dots between discrete data sources, implicitly connected by users workflows. While everyone has a knowledge graph, not all have additional enrichment layers (like behavioral annotation), and third-party data capture remains highly competitive.
Questions? Schedule an inquiry with me!
