PredictionIO implementation cost: pricing Salesforce predictions now

A data scientist who joined from a startup mentions that Salesforce once bought PredictionIO, and asks whether the org could run it. It is a fair question about a real piece of history, and it has a short answer. PredictionIO was never a Salesforce product, so there is nothing to implement in a Salesforce org under that name, and no Salesforce price for it.

The question underneath is better: what does it cost to get good predictions into Salesforce now? That one has published prices, and the rest of this page works through them.

#What PredictionIO was, and why there is no licence to buy

PredictionIO was an open source machine learning server that developers ran on their own infrastructure. When Salesforce bought the company in February 2016, TechCrunch reported the condition that matters for cost: "PredictionIO’s tech will remain open source and free to use, while the team works on evolving the project into a self-sustaining community."

So the software did not become a Salesforce product. It stayed free, under an open source licence, for anyone to run and maintain. The team went to work on Salesforce's own AI, which at the time was being built under the Einstein name. A Salesforce statement of work that includes a PredictionIO licence or implementation line has confused the open source project with the platform features that came later.

If someone in your organisation still runs a PredictionIO server, that is a self-hosted system with no Salesforce support. The cost of keeping it is your own infrastructure and engineering time. The decision about whether to keep it, replace it or retire it belongs on should I migrate from PredictionIO.

#Where prediction lives in Salesforce now

Three Salesforce products carry the predictive work that a buyer might once have looked to PredictionIO for.

Data 360. Salesforce's data platform, which the help documentation describes plainly: "As of October 14, 2025, Data Cloud has been rebranded to Data 360." Data 360 unifies records from Salesforce and outside sources into profiles, and includes a model builder for predictive models, as well as the option to connect a model trained on another platform. Scores can then be written back to records and used in Flows, segments and agents.

The bring-your-own-model route deserves a note for anyone with a PredictionIO background. Salesforce's developer material describes connecting models hosted on platforms such as Amazon SageMaker, Google Vertex AI and Databricks, so a team that already trains models elsewhere does not have to rebuild them inside Salesforce. That changes the cost shape. You keep paying for the outside platform where the model is trained and hosted, and you pay Data 360 consumption for the data it reads and the scores it returns. For teams with mature data science, that is often cheaper than retraining from scratch. For teams without it, the built-in model builder avoids running a second platform at all, and the trade is less control over how the model is built.

The Agentforce add-on. Salesforce sells add-ons for Sales, Service and Field Service and Operations that bundle what it calls its full suite of predictive, generative and agentic AI, with unmetered Agentforce use for licensed employees.

CRM Analytics and Tableau Next. Analytics products where predictions are explored, explained and put in front of managers, alongside reports and dashboards.

Which of these you need depends on the job. If the job is a single score on a record, such as the chance an opportunity closes or a case escalates, the platform's built-in predictive features may be enough. If the job needs data from outside Salesforce, such as product usage, payments or website behaviour, Data 360 becomes the centre of the project.

#What you pay Salesforce

These are list prices from Salesforce's own pricing pages, as at October 2026, in US dollars with the Australian dollar figure in brackets where shown. See the Data 360 pricing page, the Agentforce pricing page and the analytics pricing page. All three say their figures are for information and subject to change.

Two details on these pages change the arithmetic. First, the analytics page states that "The full power of the Tableau Next platform is now included in Core, Advanced, and Max editions for Sales, Service, and Industry Clouds." If you already hold one of those editions, you may not need to buy Tableau Next separately. Second, Data 360 has a free entry point. Salesforce's FAQ says that "Salesforce customers can provision Data 360 for free with a limited amount of storage and consumption credits." That is enough to prove a model on a sample before anyone signs a consumption commitment.

#How the Data 360 meter decides your bill

Data 360 pricing is consumption-based, and Salesforce is explicit about what is free and what is not. Bringing data in through batch ingestion and zero-copy integrations costs nothing. You pay when you use the data: preparing and cleaning it, matching and merging records into unified profiles, building segments, running queries, streaming in real time, and processing unstructured content such as documents and transcripts.

That has two consequences for a predictive project. The first is that the cost scales with how often you score, not with how much you store. A model that rescores every customer nightly will use more credits than one that scores at the moment a case is opened. The second is that the choice between Flex Credits and profile pricing matters. Profiles give a flat yearly cost per thousand unified customers with the profile-building work included. Credits suit organisations with fewer customers but heavier querying, or with use cases that are not about profiles at all.

Existing customers on the older Data Services Credits do not have to move. Salesforce's pricing FAQ says there is no forced migration, but that switching to Flex Credits or profile pricing means swapping all existing credits, so it is a contract conversation, not a configuration change.

The Agentforce side is metered separately unless you buy the add-on. Salesforce states that "Flex Credits are a new unit of payment for Agentforce that align cost to the business value your AI agents create." A standard action uses 20 credits, which at list works out at ten US cents. An agent that looks up a prediction and acts on it uses an action each time.

#Three common predictive jobs, and where their cost sits

It helps to look at the kinds of prediction buyers actually ask for, because each one lands on a different part of the price list.

Scoring opportunities or leads. The data is usually already in Sales Cloud: activity history, stage changes, deal size, source. A model built on that history and shown on the opportunity page is a contained project. The licence question is mostly whether your edition and any Agentforce add-on already include the predictive features you need. Consumption is low, because scoring happens when a record changes rather than across the whole database every night.

Predicting case escalation. Service history is in Service Cloud, but the strongest signals often sit elsewhere: product telemetry, delivery tracking, previous complaints in another channel. Once outside data is involved, Data 360 enters the budget, and the meter starts running on unification and queries. The build cost moves from modelling to integration, and the run cost depends on how quickly the score must refresh. A score that must be current when the case opens needs streaming data, which is metered separately from batch.

Predicting churn across a customer base. This is the most data-hungry job. It needs billing, usage and support history joined at customer level, often from a warehouse that already holds most of it. Zero-copy access avoids copying that warehouse into Salesforce, which keeps ingestion free, but every query against it is metered, and your warehouse vendor bills its own compute on the other side. Profile pricing can make the Data 360 cost predictable here, because the number of customers is known and stable.

Writing your intended jobs down in these terms before you ask for quotes does two things. It tells bidders where the effort really is, and it shows you early which meters will dominate the second-year bill.

#What an implementation is made of

Predictive projects fail on data far more often than on models. The work packages reflect that.

Define the outcome. Write down exactly what you are predicting, for which record, and what someone will do differently when the score is high. "Churn" is not a definition. "A subscription account that does not renew within 30 days of its end date" is.

Audit the history. Check that you have enough past examples of both outcomes, recorded consistently, over a period long enough to cover seasonal swings. This is where projects stall: the field was renamed, the process changed, or half the records were closed by a bulk update.

Bring the data together. Connect the sources the model needs, through Data 360 connectors, zero-copy access to a data warehouse, or MuleSoft where the integration already exists. Map and unify the records.

Build and test the model. Use the model builder in Data 360 or connect an outside model. Test it on data it has not seen, and check that it performs fairly across customer groups.

Put the score to work. Write the score to the record, show it on the page, and wire it into a Flow, a queue, a segment or an agent. A score nobody acts on is a cost with no return.

Monitor and retrain. Models drift as behaviour changes. Plan who checks accuracy, how often, and what triggers a retrain.

#What makes a predictive project cheap or expensive

Data that already lives in Salesforce, recorded consistently for two or more years, makes for a short project. Data spread across a warehouse, a billing system and a product database makes for a long one, and most of the cost is integration and unification rather than modelling.

The second driver is how many models you need. One score on one object is a contained project. A family of models across sales, service and marketing needs a shared approach to features, testing and monitoring, and that groundwork is worth paying for once rather than reinventing per model.

The third is whether you already have data science skills in house. Teams that can build and validate their own models need a partner mainly for integration and workflow. Teams that cannot will rely on the platform's built-in tools and should keep their ambitions to problems those tools handle well.

#Getting comparable quotes, and what gets left out

Give every bidder the outcome definition, the list of data sources with approximate volumes, the number of customers who will be profiled, the scoring frequency, and the edition you hold. Ask each to split their quote into the work packages above and to forecast Data 360 and Agentforce consumption for the first year.

The items commonly missing are:

• A credit forecast tied to your scoring frequency and query volume.
• Profile counts, including duplicates before unification, if you choose profile pricing.
• Sandboxes with enough real data to test a model. Data 360 sandbox capacity is worth asking about specifically.
• Data warehouse costs on the other side of a zero-copy connection, which your warehouse vendor bills, not Salesforce.
• Shield or masking, if the model uses sensitive data.
• Ongoing monitoring time, which does not end at go-live.
• The cost of retiring any self-hosted PredictionIO or other open source server you still run.

#Where to go next

For the history of the 2016 deal, see what happened to PredictionIO. PredictionIO and Salesforce explains what became of the technology. The licence question is covered on PredictionIO Salesforce licence cost. If an old architecture diagram still shows a PredictionIO server feeding Salesforce, integrating PredictionIO with Salesforce explains what that link would involve today, and before you choose a firm, use the questions on hire a PredictionIO Salesforce partner.