Dreamforce 2026: AIforce. New Value or New Packaging?

AIforce needs to justify its incremental value

If an enterprise has already paid to put its customer data, workflows and business rules into Salesforce, how much additional value does AIforce create by giving it another way to reach them? That is the question I would take into a budget discussion after Dreamforce 2026.

Salesforce's AIforce announcement positions Claudeforce, Slackforce and Agentforce Coworker as ways to use its underlying capabilities from different interfaces. There may be less navigation and less custom assembly for supported tasks. The customer still needs to establish what becomes cheaper, faster or more reliable once setup, review and ongoing operation are included.

Making Salesforce accessible through another interface may improve a task. It does not, by itself, improve the underlying process or justify another layer of cost. The release combines product development, distribution and commercial positioning. Those interests overlap in places, but they are not identical.

A sales briefing assembled from scattered information is different from changing a field. Resolving an order exception across three systems is different from summarizing a case. Treating all of them as evidence of an AI transformation obscures the actual investment decision. Each needs a baseline, a defined scope and a measurable improvement.

This assessment covers the major horizontal offerings announced around Dreamforce 2026. The product availability snapshot is dated September 18, 2026; the analysis was expanded on September 21 with additional competitive context. The workflow and financial examples are illustrative analysis, not customer implementation results or a hands-on product benchmark.

What Salesforce is bringing together

The launch narrative brings together interfaces, execution engines, models and controls with different release dates and purchasing requirements. A demonstration can make them look like one finished product. An implementation team has to establish which pieces are available, how they connect and what must be configured or built.

• Interfaces and access: AIforce connects the portfolio story across Claude, Slack and Salesforce. The Headless Toolkit exposes capabilities through mechanisms including APIs, MCP, skills and command-line tooling. Salesforce also explicitly says its Lightning and browser experiences will continue.
• Task execution: Agentforce's named agents cover customer help through Casey, employee support through Paige, shopping through Carter, operations through Marshall, inbound pipeline through Piper, customer experience through Fin and outbound sales through Hunter. Hunter is in pilot, with November general availability planned. Its new long-horizon runtime adds persistent memory, durable execution and steering across sessions. It is initially scoped to Hunter; broader coverage is future work. Multi-Agent Orchestration is generally available, while Coworker AI Skills and Agent Optimizer have October availability targets.
• Reasoning: Koa is Salesforce's CRM-focused post-training of NVIDIA Nemotron 3 Super. It is in selected Agentforce pilots, with US general availability expected in winter 2026. Treat it as a model option to evaluate, rather than an immediate capability available to every customer.
• Context: Data 360 is the new name for Data Cloud. It sits alongside the established Customer 360 foundation. These components matter to the agent story, but they were not all invented at Dreamforce 2026.
• Controls: Salesforce's Trusted Enterprise AI Harness combines capabilities across Data 360, Informatica, MuleSoft, Tableau, Agentforce, Guardian and the platform. The proposed shared control plane covers identity, policy, evaluation, observability and cost. Existing components are available, while new capabilities and the unified experience are planned to start rolling out in early fiscal FY28.

Some packaging builds on earlier releases. Help Agent was announced in June, and Agentforce Operations became generally available in April. Counting every branded agent as a new September capability overstates the change. If integration and packaging are the main improvement, measure the assembly, testing and support effort they actually remove.

AIforce is specifically the access and experience layer in this wider portfolio. Koa, Agentforce execution and the proposed Harness address different parts of the architecture. Buying into the AIforce direction does not establish that every model, orchestration feature or control-plane capability is available through every interface. Scope and release status still have to be matched to the proposed workflow.

AIforce is part of a wider platform shift

In AIForce. Cool. Now What?, Will Coffey raises a fair challenge: how much of the announcement is new, and how much makes established capabilities easier to access? Writing from a ServiceNow implementation perspective, he argues that existing workflow context deserves more attention than the location of the conversational interface.

There are concrete precedents. ServiceNow's release notes identify MCP Server Console as a new Zurich application in 2025. Its AI Control Tower launched in May 2025 with AI inventory, lifecycle and governance capabilities, including third-party AI. However, the prebuilt ITSM, ITOM, CMDB and SPM MCP servers cited in Coffey's comparison became generally available on September 10, 2026. The chronology is more mixed than a claim that everything was already available in 2025.

Those precedents place AIforce in an established industry direction: business platforms exposing context and actions through more interfaces. They do not establish functional equivalence, equal implementation effort or superior reliability. A connection to the same AI client says little about which business operations it supports, what happens when they fail or what the customer must configure.

Salesforce's Agentforce Operations offering already addresses coordinated work across systems. Describing Salesforce as only an interface would be inaccurate, while describing that work as newly enabled by the AIforce brand would also mislead. The buyer needs evidence that the proposed configuration can complete its process, recover from failure and outperform the existing approach.

For a company already invested in either platform, compare the installed capabilities and operating team before adding another layer. Assess where the process, relationships and approvals are maintained, then estimate the additional work required on each option. Neither an earlier release date nor a longer feature list settles whether the customer needs another purchase.

Availability changes the business case

The Salesforce in Claude release note describes the September beta as read-only. It cannot create, update or delete Salesforce records. It requires Claude Enterprise, administrative setup and individual Salesforce authentication. That is a material gap between the broad action-oriented launch narrative and the documented beta. It changes what a customer can responsibly scope and fund.

A team can test account research or preparation with that beta. It cannot treat the same test as proof of record updates from Claude. When the business case depends on an action the current release does not support, the missing capability belongs in the decision, with a dependency and a fallback. A roadmap date is not an operating result.

Agentforce Coworker became generally available in August, but activation has prerequisites. Its setup documentation specifies eligible editions, activated Data Cloud or Data 360, entitlements and user permissions. The record-action documentation requires administrators to enable Agent Actions and users to confirm changes. Modifications are currently limited to Salesforce CRM records and up to 15 records per command, with the user's permissions checked at confirmation.

Slackforce Surfaces introduces generated shared dashboards, reports and other interactive experiences. The availability section says live-data features begin rolling out in October. A September demonstration should not become an assumption that every deployed Surface already refreshes continuously from live sources.

The launch announcement also presents the proposition as requiring no new permissions model or custom integration work. For a supported connection, prebuilt components may remove assembly work. Extending a task into an ERP, applying company-specific authority or recovering an incomplete transaction still requires a design. Those obligations do not disappear because the front-end connection is packaged.

Before committing to a pilot, document the interface, supported action, release status, region, entitlement and required configuration. Ask the vendor to confirm that exact combination. General availability for a product family is insufficient evidence for the workflow in the business case. Make deployment obligations explicit before procurement turns a demonstration into a commitment.

Whose problem does AIforce solve?

An account leader preparing for a difficult customer conversation may search opportunities, cases, meeting notes and messages before forming a view. That is a plausible task for assistance. Measure how much searching, interpretation and correction remains after the answer arrives. If the employee still opens every source to verify it, the saved navigation may be smaller than the demonstration suggests.

For a stable, repetitive task, conversation may introduce overhead. A trained employee can sometimes complete a well-designed form faster than they can explain an instruction, wait for an answer and verify the interpretation. Natural language is an additional interaction method. Its suitability depends on the task, the user and the consequences of ambiguity.

There are three distinct ways AIforce could create value: reducing custom connector and interface work, making an established process easier to adopt, and improving the business outcome itself. A company can benefit from the first two without claiming autonomous transformation. Its business case should specify which benefit it is buying and measure that benefit directly.

Low CRM adoption also has several causes. Employees may dislike navigation, but they may equally distrust the data, disagree with the required process or see no personal benefit in maintaining records. Moving the same obligation into Slack does little for an incentive problem. An easier way to retrieve conflicting numbers does little for an ownership problem.

I would first identify where time or value is being lost: finding information, interpreting it, entering data, obtaining approval, waiting on another team or correcting mistakes. That diagnosis determines whether the answer is a better interface, cleaner data, a process change, conventional automation or an agent.

My reading of the strategy is that keeping Salesforce central as employees move into other interfaces serves Salesforce's commercial position. Customers should establish what they gain in exchange: retired work, avoided expense or an outcome they could not economically achieve before. A successful distribution strategy for a software vendor is not evidence of a successful operating change for its customer.

Put the promise through a retail exception

Consider a hypothetical retailer whose wholesale customer reports a late delivery and asks for compensation. The account manager needs the commercial relationship from CRM, the order from the order management system, shipment evidence from the warehouse and carrier, and the credit position from finance. Several teams may hold pieces of the answer.

An assistant could assemble that context, flag conflicting dates and propose a response. Test whether it retrieves the right evidence, identifies what it cannot establish and reduces total handling time after verification. A read-only experience may be sufficient for preparation; adding autonomous execution needs a separate justification.

Issuing a credit creates a different requirement. Someone must determine eligibility, value, authority and the destination system. Salesforce can remain authoritative for the account and case while the ERP remains authoritative for the credit. A platform connection does not automatically settle which system owns each business fact.

The proposed action should carry the order identifier, evidence timestamps, policy version, amount and required approver. Authorization should be checked when the action executes. A permission that existed when the conversation began may have changed, or the customer may already have received compensation through another channel.

For implementation, expose a bounded business operation such as requesting an eligible credit, with a unique request identifier that prevents duplicate execution. Salesforce's own agentic architecture guidance describes authoritative transactional systems, policy checks and duplicate suppression. These are design responsibilities even when a prebuilt connector removes some integration work.

Then test partial completion. If finance records the credit but the CRM update fails, retrying the whole instruction must not issue another credit. The workflow needs a recorded state, reconciliation and an owner who can resolve the exception. A recovery action may be a compensating transaction or manual intervention; a universal rollback across independent systems cannot be assumed.

The strongest outcome is a correct resolution with evidence, a known cost and less total effort. A persuasive chat transcript is only one part of that result. This distinction becomes more consequential as agents handle longer work: a paused request must be able to expire, accept cancellation and recheck facts before resuming.

Delegating between agents adds another boundary. A service agent handing work to an operations agent should pass the evidence, permitted scope and current state of the request. The receiving agent needs to know what has already happened and when to return control. Without that discipline, automated coordination can create duplicate work and make accountability harder to trace. More agents should earn their place through a better outcome, not become an architectural objective.

How much does Koa change the decision?

Salesforce's Koa technical paper is more qualified than the strongest launch language. It reports improvements over its Nemotron base model on evaluated tasks while stating that Koa remains below the strongest frontier models. These results warrant task-specific evaluation. They do not establish universal superiority, a production error rate or a financial return for a customer's workflow.

Salesforce says Koa inference runs within its own infrastructure. That identifies the hosting boundary. It does not settle questions about model changes, evaluation access, regional availability or switching costs, and it does not establish that customers can download the model or operate it independently. Hosting inside a familiar platform is a deployment characteristic that still needs to satisfy the customer's requirements.

An executive does not need one model to win every benchmark. The company needs an acceptable combination of task quality, speed, cost and operational control. A specialized model could be valuable if it handles the relevant cases reliably at a better cost, or makes the deployment easier to govern. Those are hypotheses to test against the actual task distribution.

Include normal cases, ambiguous instructions, unavailable tools, stale records and requests the agent should decline. Score whether the action was appropriate, not merely whether the answer sounded convincing. A model that correctly stops when a required fact is missing can be more useful than one that completes every demonstration.

Keep predictable calculations and policy checks explicit where practical. A credit limit does not become safer because a model paraphrases it well. Use model judgment where interpretation adds value, and use defined business rules where consistency is the requirement. This also makes failures easier to diagnose: the team can distinguish a reasoning error from bad evidence, a missing permission or a broken service.

More generated interfaces need more consistent business meaning

Surfaces and other generated experiences could shorten the distance between a question and a useful shared view. The challenge is consistency. If one team asks for revenue and another asks for sales, they may receive attractive views that use different dates, currencies, channels or cancellation rules. A polished interface can make that disagreement harder to notice.

The practical response is to maintain approved metric definitions, authoritative sources and visible freshness information. A generated view should make its scope understandable. For a consequential decision, the user needs to distinguish an estimate from a booked result and a historical snapshot from current operational state.

Salesforce's Data 360 engineering description distinguishes working context, short-term traces, approved durable memory and validated learnings. It also describes assembling authorized evidence under freshness, latency and cost constraints. The implementation must demonstrate those distinctions: a conversation transcript should not automatically become a permanent business fact.

For the retail example, a customer's statement that a parcel never arrived is evidence of a complaint. It does not overwrite the carrier event or establish eligibility for a credit. The workflow must preserve both facts, recognize the conflict and apply the relevant process. More context helps only when its meaning and authority remain clear.

Generated interfaces also need ordinary product discipline. Employees should be able to inspect the proposed action, correct an interpretation and understand the result. Accessibility, permissions and usable error messages remain requirements. An interface assembled dynamically still represents the company to the person relying on it.

What does a completed outcome actually cost?

The commercial discussion should start with a unit of work that the business recognizes. For this example, use a correctly resolved eligible service case, with no unresolved downstream correction. Keep rejected cases, escalations and rework in the measurement rather than removing inconvenient outcomes from the sample.

The following figures are invented planning assumptions in Canadian dollars. They are neither Salesforce prices nor a productivity forecast. They show how a promising saving can change once the full workflow is counted.

• Current process: 10,000 eligible cases per month at eight minutes of employee effort each equals 80,000 minutes. At a fully loaded CAD 60 per hour, that represents CAD 80,000 of monthly labour capacity.
• Proposed AI process: three minutes of active handling plus one minute of review, with 10% of cases requiring 15 additional minutes of exception work. Average effort is 5.5 minutes per case, or CAD 55,000 of labour capacity for the same volume.
• Gross capacity value: CAD 25,000 per month, equivalent to about 417 hours. If incremental licensing, consumption, data and ongoing support cost CAD 14,000 per month, the potential net monthly value is CAD 11,000 before one-time implementation costs.

That is capacity value, not automatically cash saved. Finance needs to know how the recovered hours will reduce overtime, avoid hiring, absorb demand or improve service. If staffing and output remain unchanged, presenting the entire amount as a cash saving would overstate the result. One-time integration, testing and change costs also need a separate investment case.

Require a retirement plan alongside the new expenditure. Name the manual activity, custom integration, tool or subscription that will disappear, when it will disappear and who will confirm it. If everything stays and AI adds another bill, the business case must explain why the additional outcome is worth the additional cost. Previously funded data and workflows should not be counted again as benefits created by the new purchase.

Sensitivity matters more than a single attractive estimate. If exceptions rise from 10% to 20%, average effort becomes seven minutes per case. Gross capacity value falls to CAD 10,000, below the assumed CAD 14,000 operating cost. Under those assumptions the recurring case turns negative before implementation expense. Exception handling belongs in the business model from the beginning.

Commercial configuration also affects the answer. Coworker's billing documentation distinguishes seat-based and consumption access, with user permissions affecting billing. It treats eligible CRM and Slack searches differently from additional Data 360 source processing, indexing and queries. An existing Salesforce contract does not establish the incremental cost of the proposed workflow.

Request a worked estimate using expected volume, source mix, model usage and exception rate. Compare it with measured pilot consumption. Count quality as well as effort: faster cases are not an improvement if customers reopen them, credits are incorrect or another team inherits the cleanup. Revenue claims require a credible comparison too; a generated sales message is not incremental revenue.

More choice at the interface can mean deeper platform dependence

The partner announcements have separate delivery timelines. Salesforce's AWS expansion includes currently available headless access through Amazon Quick and model choice through Bedrock, alongside future capabilities. The Google Cloud announcement identifies Gemini-powered Agentforce reasoning as generally available, while the Salesforce connector for Gemini Enterprise is in private preview and the sales agent in Gemini is in beta. The checkout and infrastructure roadmaps have their own later dates.

For customers, a wider choice of interfaces and models can reduce dependence on one interaction experience. It does not necessarily make the underlying workflow portable. Customer context, permission mapping, memory, action definitions and evaluation history may become more deeply tied to the Salesforce architecture.

Price that dependency into the decision. Ask what can be exported, which business rules live outside prompts, whether another client can invoke the same bounded actions and what would happen if a model or integration became unavailable. Include the cost of maintaining a fallback and moving the workflow later. An open protocol helps connectivity; it does not make every surrounding dependency interchangeable.

A company using Salesforce, ServiceNow and an ERP also needs to decide which component coordinates a process that crosses them. For the retail credit example, identify one authoritative record of execution state and one accountable owner, while the individual systems retain their business records. Competing coordinators must not each retry the same credit or disagree about whether the request is complete.

Reusing permissions is a useful starting point, but existing access may be broader than a delegated task requires. Define the agent's permitted operations and approval thresholds, then retain evidence linking the request, policy decision, system changes and final result. Logs in several products become useful operational evidence when they can be reconciled around the same transaction.

The operating team also changes. Administrators need to understand access and supported actions. Engineers need traces across system boundaries. Business owners need to define acceptable outcomes and exceptions. Finance needs usage attribution. Fewer browser logins can coexist with more responsibility for the shared services underneath.

Run a pilot that is allowed to choose the simpler answer

Select one workflow with meaningful volume, observable outcomes and a business owner who can change the process. Compare three approaches: the current process, an improved conventional process, and an AI-assisted or agent-led process using capabilities actually available to the organization. The conventional improvement might be cleaner data, a simpler screen or a defined automation.

In the first week, agree on case eligibility, baseline effort and outcome quality. Include representative difficult cases and record which work the pilot excludes. In the second, test retrieval and proposals against known outcomes. Investigate missing evidence and mistaken actions before widening the scope.

In the third, permit bounded execution only where the product supports it and the business has approved the control design. Test cancellation, permission changes, repeated requests and partial failures. In the fourth, review mature outcomes, total effort, consumption and the work created for support teams. Longer sales or operations cycles need a longer observation period; a four-week pilot cannot manufacture evidence for outcomes that have not happened yet.

Set minimum quality and recovery requirements before comparing cost. Then expand the scope when the AI option produces a better accepted outcome at an acceptable cost. Keep a read-only or proposal-only deployment if that is where the value is demonstrated. Defer when a required capability remains on the roadmap. Choose the conventional improvement when it performs better.

My assessment is that AIforce presents a clearer access strategy than a proven economic case. Some customers may recover enough implementation effort or employee time to justify it. Others may add interfaces, consumption charges and support obligations around a process that remains fundamentally unchanged. The evidence has to distinguish those situations.

I would fund a specific available capability against a defined problem, with an accountable owner and a measurable result. I would not approve a broader AI commitment on the assumption that simpler access will eventually produce value. Salesforce should demonstrate the incremental benefit with the customer's workflow and cost structure. The customer should be prepared to conclude that the next purchase is unnecessary.

Customers should not mistake easier access to a platform for a return on the investment in it. The next purchase should retire work or prove a better outcome, even when the demonstration is impressive.

Michel Junior Julien