Multimodal AI models are designed to work across several types of information — or modalities — such as text, images, audio, video and structured data. They can combine those inputs to develop a richer understanding of a situation.

Think about the last time you had to make a reasonably important decision at work. Perhaps you opened the customer record in Salesforce, read an email, checked a figure in an attached PDF and looked at a photograph the customer had sent. You assembled those different pieces of information into a picture.

People do this naturally. We read words, interpret tone of voice, understand diagrams, recognise objects in photographs and compare what we see with information we already know.

For Salesforce customers, multimodal AI means that it can begin to work with more of the information surrounding a CRM record: documents, photographs, conversations and other evidence employees already use to understand customers and make decisions. As Salesforce develops Agentforce, Data 360 and Document AI, multimodal AI is likely to become an increasingly practical part of Salesforce architecture in 2027 and beyond.

What are multimodal AI models?

A modality is simply a form of information. Text is one modality. Images, audio, video and structured data are others.

Stanford University's Institute for Human-Centered Artificial Intelligence defines multimodal AI as systems that can process, understand and generate multiple types of data simultaneously. (Source)

A familiar example is an AI assistant that can receive a photograph and answer questions about what it sees. Give it a chart and it can interpret the visual information alongside the text. Give it a PDF containing words, tables and diagrams and a multimodal AI model can potentially consider all three.

Consider an insurance claim. The written claim provides the driver's account. Photographs show the damage. Location and weather data add context. The policy establishes what is covered. Previous claims provide history. Together, those sources tell a fuller story.

Research into multimodal AI models is developing quickly. Stanford's 2026 AI Index reports that scientific publications on multimodal foundation models for biomedical discovery increased from just two in 2021 to 462 in 2025. Google DeepMind has demonstrated multimodal models working across medical images, clinical information and genomics. The applications are highly specialised, but the principle travels well: related information becomes more useful when it can be understood together.

Bringing multimodal AI models into Salesforce

Salesforce provides a structured representation of the customer and the business around them. An Opportunity has a value and stage. A Case has a status. Relationships connect Accounts, Contacts, Opportunities, Cases, Assets and custom objects. Around those records sits another layer of information: emails, signed agreements, PDFs, photographs, reports, call recordings, presentations and spreadsheets. (Article on Document AI by Salesforce)

Picture an employee assessing a financial application. Salesforce tells them who the applicant is, the amount requested, the product and the stage of the application. They then open a valuation report, inspect photographs, check supporting financial documents and read the latest correspondence.

The Salesforce record is the spine of the process. The surrounding material provides much of its context.

Multimodal AI models create the opportunity to work with both. Salesforce is already developing the building blocks. Data 360 can provide AI and automation with structured and unstructured information. Document AI can extract structured information from business documents, while its multimodal processing can interpret images, diagrams and image-rendered tables embedded in PDFs.

Salesforce estimates that around 80% of enterprise data is unstructured, much of it sitting in documents and content repositories beyond the direct reach of traditional CRM processes. Multimodal AI is also one of the developments we're watching as Salesforce moves towards Dreamforce 2026 and the next generation of Agentforce.

Multimodal AI gives more of that information a role in Salesforce.

Documents, images and Salesforce data working together

Documents are particularly interesting because they sit between human communication and structured business data. A person can look at a financial statement and understand the relationship between a paragraph, a table and a figure several pages later. We recognise signatures, diagrams, handwritten annotations and visual hierarchy.

Traditional document processing can lose some of those relationships when a document is converted into plain text. Salesforce's Document AI architecture documentation specifically identifies table hierarchies, spatial relationships and cross-references as examples of information that can be lost.

Multimodal processing can retain more of that context.

For document-heavy organisations, this creates an inbound journey: Document → Understand → Extract → Validate → Salesforce Data → Business Action. There is also an outbound journey: Salesforce Data → Generate → Review / Approve → Sign or Send → Store → Next Action.

This is where multimodal AI connects naturally with Salesforce document automation. At CloudAnalysts, PDF Butler is our preferred Salesforce document automation platform because of the flexibility it gives us to design around a client's actual process and requirements. We use it to generate documents from Salesforce data, collect information, manage review and approval processes, obtain signatures and connect documents to the next stage of a Salesforce workflow. That saves many hours on every single relationship and deal.

Multimodal AI extends that opportunity further. Information arriving in documents and images can be understood, extracted and brought into the Salesforce process; PDF Butler can then use Salesforce data and decisions to create and manage the documents travelling in the other direction. Together, this creates a two-way document journey:

Documents → Data

Data → Documents

Or, across the complete process: Receive → Understand → Structure → Decide → Act → Generate → Sign → Store → Continue.

A financial-services example

Consider a property lender receiving a new application. The Salesforce Opportunity contains the applicant, property details, requested loan, product and stage. Supporting information might include a valuation report, bank statements, identification documents, property photographs, application forms and email correspondence.

An experienced employee brings these sources together. They compare the valuation with the property data in Salesforce, check whether supporting figures match the application and identify gaps or inconsistencies.

A multimodal AI model could assist with that work by interpreting the documents and images alongside the Salesforce data. It could extract relevant information, highlight discrepancies, identify missing evidence and prepare the application for review. Agentforce or Salesforce automation could then create tasks, request further information, route the application or prepare the next document.

The same pattern can apply across claims, onboarding, compliance, service and other processes where employees currently move between Salesforce and supporting evidence before deciding what to do.

Good architecture becomes even more valuable

Giving an AI agent access to more information also gives it more opportunities to misunderstand how that information should be used. Suppose a document contains a customer's address. The AI may extract it perfectly. Salesforce still needs to know whether it is the customer's current address, the address of a financed property, an old address appearing on a bank statement or simply supporting evidence that should never overwrite the Contact record.

Consent presents a similar challenge. Finding a consent statement in a signed document does not establish how that consent should be represented, which purposes it covers or when it expires. Multimodal AI therefore needs to sit within a clear Salesforce architecture covering data ownership, hierarchy, permissions, consent, validation, retention and human approval.

It also raises specific security questions. Salesforce's current Document AI architecture documentation states that LLM interactions use the Einstein Trust Layer, including zero-data-retention controls. It also notes that PII masking is currently unavailable for multimodal data such as attached documents processed through Document AI.

More context needs appropriate controls. Data 360 can help organisations decide how different information should participate in the architecture. Some values need to become structured Salesforce data because workflows, reporting or compliance processes depend on them. A complete report may remain unstructured but become available as context when an authorised agent needs it. A 60-page report might therefore provide five governed data points to Salesforce while the full document remains available to answer more detailed questions.

Preparing Salesforce for multimodal AI

One of the easiest ways to identify opportunities is to watch what employees already do around Salesforce.

Which PDFs do they open before updating an Opportunity? Which photographs do they inspect before progressing a Case? What information do they repeatedly copy from documents into fields? Which conversations, spreadsheets or reports do they consult before making a decision?

These are processes where people are already working multimodally.

From there, organisations can decide which information AI could usefully interpret, what should become structured data, where that information should live, which actions can safely follow and where human judgement belongs.

For years, organisations have worked to create a 360-degree view of the customer. Multimodal AI models can enrich that view with information that has traditionally sat around the CRM record rather than within it.

Data 360 can provide context. Document AI can turn selected unstructured information into governed data. Agentforce can use that context to reason and take authorised actions. Document automation can carry data and decisions into the documents that move the process forward.

For Salesforce customers looking towards 2027, the first opportunities may already be sitting in plain sight: the documents, images and conversations employees open alongside Salesforce every day.