Forrester's AI Platforms Wave, Q3 2026: What ITSM Buyers Should Take Away

Forrester published The Forrester Wave™: AI Platforms, Q3 2026 this month, and the most important thing about it is not who leads. It is what the category now means.

In its public commentary on the evaluation, Forrester says agentic AI has redrawn the boundaries of what an AI platform is and what vendors compete to provide. For most of the past decade, an "AI platform" meant a data science workbench: an environment where teams readied data, trained models, and produced analytical insight. The new evaluation judges platforms on something harder: whether they can grasp the context of a task, move through enterprise workflows, and finish the job.

The 15 vendors assessed make the point on their own. Alongside hyperscalers such as AWS, Google, and Microsoft sit data platforms like Databricks, enterprise application vendors like Salesforce, ServiceNow, and Oracle, automation specialists like UiPath and Pegasystems, and analytics-rooted players like IBM, Palantir, and C3 AI. Forrester describes the field as among the most varied it has ever evaluated in this category, and treats that variety as a strength rather than a flaw, because each vendor has gravitated toward a distinct area of depth.

I run revenue at an ITSM company, so I read analyst research the way buyers do: what does this change about how you should evaluate? In this case, quite a lot. Three takeaways stand out, and they lead to a very practical checklist.

Takeaway 1: The Model Is Not the Moat

The most revealing structural decision in this Wave is who is not in it. Forrester deliberately excluded frontier labs such as OpenAI and Anthropic. Not because they are irrelevant, but because those labs are evolving into full platforms themselves and will get their own coverage: Forrester has scheduled a Frontier AI Model Platforms Landscape for Q4 2026 and a full Wave evaluation for Q1 2027.

The platforms that were evaluated are, with few exceptions, model-agnostic. Forrester locates their value in what surrounds the model: data, process context, agent development, governance, and deployment. It frames that flexibility as an advantage, since enterprises can replace models with better ones over time without rebuilding the stack.

For buyers, the implication is blunt. Raw model capability is becoming a shared input that every serious vendor can access, which means it cannot be the basis of your decision. The durable differences live in the layer around the model: whether the system is grounded in your data, whether it carries your process context, whether agents can be built and governed by your teams, and whether any of it can be deployed and audited in your environment.

This is also the definitive answer to the "LLM wrapper" question. The accusation was always aimed at the wrong layer. A thin chat interface on a frontier model is a wrapper. A system where eight specialized agents reason over one shared conversation, grounded in enterprise knowledge and wired into enterprise actions, is a platform, and Forrester has now built the category's evaluation around exactly that distinction.

Takeaway 2: Your AI Strategy Will Be a Portfolio

Forrester's second argument breaks with a decade of procurement instinct. Its guidance to buyers is that an AI platform strategy will be "a portfolio, not a monolith."

The reasoning, paraphrased: every platform in the evaluation is general-purpose, but each has an affinity, whether that is data science depth, workflow specialization, industry solutions, or application development. Push every use case onto one platform and most of them will sit outside what that platform is genuinely good at. Enterprises get the most value by putting each use case on the platform whose strengths actually fit it.

For years, the safe enterprise answer was consolidation: fewer vendors, one suite, one throat to choke. Forrester is saying that for AI, that instinct now destroys value. Heterogeneity is rational because the work is heterogeneous.

We have been making a version of this argument for a while. Our CEO wrote recently about why a mile deep beats a mile wide in agentic AI for employee service. The portfolio thesis is the analyst-grade version of the same idea. Employee service delivery is a use case with unusually demanding requirements: identity and permissions, actions across dozens of systems, compliance and audit obligations, high volume, and outcomes that are measured daily. Use cases like that reward depth. Under a portfolio strategy, the question stops being "which mega-suite do we standardize on" and becomes "which system is genuinely built for this workload."

Takeaway 3: Execution Is the Bar, and Execution Is Measurable

The quiet consequence of redefining the category around finished work is that success becomes verifiable. Insight is debatable; execution is binary. A password got reset or it did not. Access was provisioned or it was not. The ticket never needed to exist, or an agent spent twenty minutes on it.

That is a much less forgiving standard than the one AI programs have been grading themselves on. Our CEO made a related point this week off the back of Gartner's audit survey, where 93% of teams use AI but only 38% have a strategy and usage pools in low-stakes drafting work. Forrester's category redesign attacks the same disease from the vendor side: if platforms are now evaluated on context, workflow navigation, and completed processes, then buyers should be too. Adoption dashboards do not survive contact with an execution standard. Resolution numbers do.

The ITSM Buyer's Checklist for the New Category

A necessary disclosure first: Rezolve.ai was not part of this evaluation. The Wave assesses horizontal AI platforms, not domain-specific service management systems. I am writing about it because the criteria shift describes, almost line for line, what buyers should now demand from any system claiming to bring AI to employee service. Five questions cover it.

Does it understand service context? Not "can it chat," but is it grounded in your knowledge base, your tickets, your policies, and your entitlements, with cited answers you can check? That is the difference between an assistant that answers and executes and one that speculates.

Does it execute end to end? The new bar is completed work: diagnose the issue, take the action, confirm the fix, and close the record, on the channels employees already use, including Teams, Slack, email, and voice.

Is autonomy governed? Execution without governance is a liability. Every autonomous action should be permission-aware, approval-gated where risk warrants it, and fully audited and explainable after the fact. Certifications matter here too: SOC 2 Type II, ISO 27001, GDPR, and HIPAA-readiness are table stakes for a system acting inside your environment.

Does it integrate with the enterprise you actually run? If value lives in what surrounds the model, then your surroundings are your systems. Look for governed agents you can build in plain language, with MCP-based tool integrations, running on a true system of record, either standalone or augmenting the ticketing platform you already own.

Can it prove completed outcomes? Ask for numbers that describe finished work. Across Rezolve.ai deployments, roughly 70% of requests are resolved before they become tickets. MyEyeDr resolves issues in about ten minutes, with around 30% never becoming tickets at all. Black Angus Steakhouse cut after-hours IT dependency from 90% to 10%. JLL runs its Ask Ethics program on Rezolve.ai for roughly 100,000 employees across more than 80 countries. Those are execution metrics, and any vendor operating at the new bar should be able to show you their own.

Choose Like It Matters

Forrester ends its commentary with an unusually stark warning about how much rides on platform choice, and it is right. Strip away the drama and the operational advice is simple: the category has finally caught up to what good buyers already knew. Judge AI systems by the work they complete under governance, match deep use cases to deep platforms, and treat the model as the one component you should always be able to replace.

If you want to apply that standard somewhere concrete, bring us your hardest service workload and we will show you what executed looks like.