AI Platforms

UiPath is a robotic process automation company. ServiceNow sells IT service management. Salesforce sells CRM. All three were evaluated in Forrester's most recent AI platform assessment, alongside Amazon, Google, and Microsoft.

Four years earlier, the same evaluation covered RStudio, RapidMiner, H2O.ai, Cloudera, and TIBCO. None of them appear now.

Nothing about that turnover reflects those companies failing. It reflects Forrester deciding that an AI platform is a fundamentally different thing than it was, and the vendor list is where that decision becomes visible.

What the category used to mean

For most of the last decade, an AI platform was a data science workbench. An environment where you prepared data, engineered features, trained models, evaluated them, and pushed the good ones into production. The user was a data scientist. The output was a model. The success metric was predictive accuracy against a holdout set.

Forrester's Q3 2022 evaluation, published under the name AI/ML Platforms, scored fifteen vendors against twenty five criteria: Amazon Web Services, C3 AI, Cloudera, Databricks, Dataiku, DataRobot, Google, H2O.ai, IBM, Microsoft, Palantir, RapidMiner, RStudio, SAS, and TIBCO Software.

The buying advice attached to it was consistent with that framing. Look for breadth of tooling for data scientists and the extended AI team of data engineers, ML engineers, and architects. Look for industry solution accelerators. Look for extensibility, on the reasoning that serious AI projects do not rest on a single model.

Sensible advice for building models. Almost entirely beside the point for what enterprises are buying now.

What it means in 2026

The Forrester Wave: AI Platforms, Q3 2026 published in August 2026, scoring fifteen vendors against nineteen criteria across current offering and strategy. Forrester's own framing is that agentic AI has redrawn what an AI platform is, what it must do, and which vendors compete to provide it, and that anyone familiar with previous evaluations should expect to recalibrate.

The evaluated field: Amazon Web Services, C3 AI, Databricks, Dataiku, DataRobot, Google, IBM, Microsoft, Oracle, Palantir, Pegasystems, Salesforce, SAS, ServiceNow, and UiPath. Forrester describes it as among the most heterogeneous fields it has ever assessed in this market.

Set the two lists side by side and the substitution is precise. Out go the pure data science and analytics tooling vendors. In come workflow automation, RPA, IT service management, CRM, and enterprise applications.

That is not a market that grew. It is a market whose defining problem changed. Building a model is a data science problem. Getting an agent to take an action inside a business process, with permissions, against real systems, with an audit trail, is a workflow problem. The vendors who already owned workflow turned out to be closer to the new requirement than the vendors who owned model training.

The criteria confirm it. Reporting on the evaluation identifies governance control, runtime architecture, and agent development among the criteria assessed. None of those three would have appeared in a data science workbench evaluation in 2022.

What the results show

Five Leaders: Google, Amazon Web Services, C3 AI, Palantir, and Pegasystems.

C3 AI took the highest current offering score of all fifteen providers, with the highest possible score in eight criteria. Forrester assessed its agentic platform, built around modelling the enterprise as a single unified ontology graph so that every application and agent works from one representation of the business. That architectural choice, one governed model of the enterprise rather than per-application data access, is a real differentiator and a real commitment.

Google took the highest score in the strategy category, on an integrated argument spanning a front-door application layer for business users and an agent platform underneath for technical teams to build and deploy production agents.

Pegasystems placed as a Leader and came second on strategy, with Forrester describing a comprehensive platform whose roots are in workflow automation, machine-learning-based customer offer personalisation, and decisioning. Read that description next to the category shift and it explains itself. Pega did not pivot into this market. The market moved toward what Pega already was.

Note also what is absent from the scoring dimensions. This Wave scores current offering and strategy only. There is no third axis, neither market presence nor customer feedback, which is worth knowing when comparing this evaluation against others in the same period.

Two firms, one name, two different markets

The most useful thing a buyer can know about this category right now is that Forrester and Gartner are not evaluating the same market, despite nearly identical category names.

Gartner published its Magic Quadrant for AI Platforms for Data Science and Machine Learning in June 2026, covering eighteen vendors: Alibaba Cloud, Amazon Web Services, Cloudera, Databricks, Dataiku, DataRobot, Domino Data Lab, Google, H2O.ai, IBM, MathWorks, Microsoft, Posit, Red Hat, SAS, Siemens with Altair, Snowflake, and Teradata.

Eight names appear on both lists: AWS, Databricks, Dataiku, DataRobot, Google, IBM, Microsoft, and SAS.

Seven appear only on Forrester's: C3 AI, Oracle, Palantir, Pegasystems, Salesforce, ServiceNow, and UiPath. Every one of them comes from workflow, applications, or operations.

Ten appear only on Gartner's: Alibaba Cloud, Cloudera, Domino Data Lab, H2O.ai, MathWorks, Posit, Red Hat, Siemens with Altair, Snowflake, and Teradata. Almost every one comes from data science tooling or data infrastructure.

Gartner kept the data science framing, which is stated plainly in its category name. Forrester dropped it. If you build a shortlist by combining both documents, you will produce a list of twenty five vendors solving two different problems, and the comparison will not resolve.

Pick your problem first. If the question is where your data scientists build and manage models, the Gartner list is the relevant population. If the question is where your agents run against business systems, Forrester's is.

Where the model builders went

An obvious absence in Forrester's list: none of the frontier model labs.

That is deliberate rather than an oversight. Forrester has separated them into a new category, frontier AI model platforms, defined around vendors that develop and own state-of-the-art models, carry the capital burden of advancing them, and wrap them in the tooling and infrastructure that turns model capability into business capability. The Frontier AI Model Platforms Landscape is scheduled for Q4 2026 with a full Wave evaluation in Q1 2027.

Forrester's position is that together the two evaluations map the complete AI platform decision space, which is a reasonable way to describe a genuine architectural fork.

One path is a bet on a single lab's trajectory, accepting dependence on that lab's model in exchange for being early to whatever it ships. The other is a bet that your data, your workflow context, and your ability to swap models underneath matter more than any model's current lead.

Most large enterprises will end up making both bets in different parts of the business, which is fine as long as it is a decision rather than an accident.

Where this sits in the wider framework

Forrester now describes the agentic technology market as three functional planes, and it is worth locating this category inside that structure.

The build plane is where agentic systems get created, and AI platforms are its centre of gravity. The orchestration plane, which Forrester calls adaptive process orchestration, is where agentic and deterministic components get embedded into business processes. The oversight plane, the agent control plane, is where a heterogeneous agent estate gets inventoried, governed, and constrained independently of the runtimes it runs on.

Vendors span these planes and blur the boundaries constantly, which is exactly why the framework is useful. Several of the fifteen vendors in this Wave also appear in Forrester's orchestration and oversight coverage, and a vendor telling you it covers all three is describing scope rather than depth.

The practical use of the framework is that it tells you which questions to ask in which conversation. Build plane questions are about developer experience, model access, and deployment. Orchestration questions are about process modelling and integration. Oversight questions are about independent governance, and the honest answer from a build plane vendor is that it can govern what runs on it and nothing else.

What to test

Bring an agent that has to do something consequential. Not summarise a document. Something that writes to a system of record, has a permission boundary, and would matter if it went wrong. Everything interesting about these platforms lives in that gap between demonstration and production.

Ask what happens between a working prototype and a governed deployment. Every vendor in this field can produce a working agent in a workshop. The differentiators are what a production deployment requires, who approves it, what gets logged, and how long the path takes.

Interrogate the data layer honestly. C3 AI's ontology approach and Palantir's data foundation are architectural commitments with real consequences, positive and negative. A unified enterprise model is powerful and expensive to build and difficult to leave. A lighter approach is faster to start and produces the fragmentation problem later. Neither is wrong, but they are not interchangeable and vendors will not frame the trade-off for you.

Check model portability specifically. Ask what changing your underlying model provider actually involves, in engineering time, in prompt and evaluation rework, and in commercial terms. The answer ranges from a configuration change to a rebuild, and it is rarely volunteered.

And separate the platform decision from the application decision. Several vendors here arrived from owning an application category, and their AI platform is genuinely strong inside their own estate and thinner outside it. If most of your process already runs in one of those products, that gravity is a legitimate reason to buy. It is not the same as the platform being the best available.

One note on timing

This Wave published in August 2026, which makes it the most current evaluation in this space and also means the announcement cycle is still running. Additional vendors will confirm their placements over the coming weeks, and the Strong Performer and Contender tiers are not yet fully public.

More importantly, the frontier model platforms Landscape and Wave arrive in Q4 2026 and Q1 2027. Anyone making a major platform commitment in the next two quarters is doing so with half the map published. That is an argument for a shorter initial commitment rather than for waiting, since the category is moving faster than the research cycle in either direction.

Forrester does not endorse any vendor named here, and tier placement should not be read as a recommendation to buy.