Conversational AI Stopped Talking and Started Doing the Work

For roughly a decade, the customer-service chatbot had one honest job: to say "I'm sorry, I didn't quite get that," and route you to a human. In 2026, its job description changed. It issues the refund. It reschedules the flight. It updates the address on file and hands back a log of what it did. That single shift - from answering questions to completing tasks - is the whole story of this year's conversational AI market, and Forrester just put numbers on it.

In June, the research firm published The Forrester Wave: Conversational AI Platforms For Customer Service, Q2 2026, an evaluation of 14 vendors judged not on how clever their language models sound but on something far less glamorous: how well they integrate, how safely they operate, and how confidently they can evolve as the underlying AI keeps changing. The result reads less like a beauty contest and more like a stress test. Plenty of platforms can talk. Fewer can be trusted to act.

Conversational AI platforms are shifting from answering questions to completing tasks.Forrester, Q2 2026 Wave framing

01 / THE SCOREBOARDWho Forrester put in the top-right

Three names landed in the Leaders band: NiCE Cognigy, praised for strength of strategy; Kore.ai, which posted the field's highest current-offering score at 4.14 with top marks across ten criteria; and Omilia, which earned its place on OEM relationships and edge-case voice work. One rung down, the Strong Performers: Sierra, Uniphore, Intercom and Rasa.

The interesting part isn't the ranking. It's the shape of the field once you plot it. Raw AI capability clusters everyone toward the top; what pulls them apart, left to right, is whether the thing actually plugs into the systems an enterprise already runs.

That framing is a quiet rebuke to the last three years of AI marketing. Every vendor can point at a benchmark and a demo where the model dazzles. Almost none of that shows up in the scores. What Forrester rewards instead is governance, model management, back-end integration and the ability to keep an agent from doing something expensive and wrong at 2 a.m. Kore.ai earned its top spot partly on exactly those unglamorous strengths. In a category built on conversation, the differentiator turned out to be control.

02 / THE MISSING NAMESThe report you notice for who isn't in it

Read the vendor list twice and the omissions start to shout. Google - which owns frontier models and a Contact Center AI product inside Google Cloud - was not evaluated. Neither was Botpress, the developer-favorite that lets teams build agents in code, nor Norway's boost.ai, which built a real European business on safe, controllable conversational AI. A Wave covers vendors that fit the cycle's inclusion criteria, so absence isn't a grade. But when the biggest name in AI sits out the scorecard that defines a category, the silence is worth a paragraph of its own.

Sometimes the most interesting name in a report is the one that isn't in it.

It's a useful reminder for any buyer reading an analyst grid: the map is not the territory. Capacity, the support-automation company that has spent years rolling up smaller tools, and Observe.AI, which turns 100% of recorded calls into coaching and QA, both run large books of business without needing a spot in this particular quadrant. Being excluded from a scorecard and being irrelevant are very different things.

03 / THE SIERRA QUESTIONA two-year-old worth $15.8 billion

No company embodies the moment better than Sierra. Founded in 2023 by Bret Taylor - former co-CEO of Salesforce and chairman of OpenAI - along with former Google executive Clay Bavor, Sierra skipped the chatbot era entirely and sold autonomous chat and voice agents built to complete work, with cross-session memory so they remember who you are. Customers followed. By early 2026 the company reported roughly $150 million in annual recurring revenue, up from $100 million a few months earlier, with more than 40% of the Fortune 50 on the books. A round led by Tiger Global and GV pushed its valuation to about $15.8 billion.

And yet Forrester made Sierra a Strong Performer, not a Leader - and handed it a 1 out of 5 on legacy system integration. That is the whole tension of 2026 in a single data point. Valuation is one scoreboard. Enterprise-readiness is another. A company can be the most exciting thing venture capital has funded this year and still trip over the twenty-year-old order-management system your refund actually lives inside.

04 / THE HARD PARTSEveryone shipped the robot before the ruler

Forrester's takeaways land on the unglamorous problems, and they rhyme. Legacy integration remains the common blocker for newer platforms. Analytics lag capability - providers pushed out agentic frameworks faster than the measurement tools to prove those agents work. And pricing is a mess: only one vendor, Rasa, scored above 3 out of 5 on pricing flexibility and transparency. In a market projected to reach $50 billion, the product is nearly ready and the invoice is still a riddle.

There's a personalization gap too. For all the talk of agents that "know you," only one vendor was singled out for genuinely strong customer-data management. Cross-session memory - the feature Sierra leans on hardest - is still the exception, not the norm. Most agents meet you as a stranger every time, which is a strange thing to charge enterprise money for in the year of the personalized everything.

What actually differentiated the front-runners was speed-to-value: prebuilt components, AI-assisted development, and forward-deployed engineers who sit with the customer until the thing works. In the enterprise, boring-and-installed beats brilliant-and-blocked. The winning agent is rarely the one with the cleverest answer. It's the one that finishes the task and shows its work.

05 / THE CONSOLIDATIONThe land-grab is ending; the plumbing is starting

Maturity has a tell, and it's M&A. NiCE didn't build its way to the top-right - it bought Cognigy and stitched it into CXone. Elsewhere in the adjacent voice-AI world, SoundHound moved to absorb LivePerson. The names that spent years turning contact-center conversations into data - Uniphore, Observe.AI, Capacity - now have to turn that data back into action, automatically. The measurement companies are quietly becoming the automation companies.

The winning agent isn't the one with the cleverest answer. It's the one that finishes the task and shows its work.

For a support leader holding a budget, the buying question has flipped. It used to be "how smart is the bot?" Now it's "how much of my stack does it touch, can I trust what it does there, and can I see the receipts?" The vendors that answer those three questions cleanly are the ones sitting in Forrester's top-right. The rest are betting that a bigger model will paper over an integration they haven't built yet. It won't.

06 / WHAT IT MEANS FOR YOUReading the map without buying the hype

If you're evaluating any of these platforms, treat the Wave as a starting filter, not a verdict. A few practical takeaways from this cycle: ask every vendor to demo an agent completing a task inside your own systems, not a scripted one in theirs. Push hard on analytics - if they can't measure containment, resolution and error rates out of the box, you're buying a black box. And get pricing in writing, because the market has not solved that yet. The technology to replace deflection with resolution is real. Whether a given vendor can deploy it against your particular mess of legacy systems is the only question that matters.