By Tom Bexx · Addbox
Salesforce making itself accessible through ChatGPT and Anthropic’s Claude sounds, at first, like another sensible AI integration.
Customers can work where they prefer. Less switching between applications. Fewer screens to navigate. More work completed through a simple request.
But the direction being championed by Marc Benioff raises a question that I think deserves more attention:
If someone can use your product without opening it, which part of your product are they actually paying for?
Salesforce’s announcement with OpenAI described bringing CRM capabilities into ChatGPT. Its Headless 360 offering goes further, positioning APIs, MCP tools and other interfaces as ways to access Salesforce outside its traditional browser experience. These are supported capabilities and integrations, not a promise that every customer can perform every Salesforce task through any chatbot without configuration.
For smaller software businesses, and for companies deciding what to build next, this is much more than a Salesforce story.
It changes the question from “How do we add AI to our application?” to “What role should our application play when AI becomes the place where work starts?”
The screen used to be the product
For years, software companies have competed partly on the experience of using their application.
The dashboard. The navigation. The carefully designed customer record. The reporting screen that makes a complicated problem manageable.
Businesses have invested heavily in those experiences because that is where customers spend their time. A better interface can reduce training, improve adoption and make a product harder to replace.
Now consider a sales manager asking an assistant:
“Show me the opportunities likely to slip this month, explain why, and draft follow-ups for the account owners.”
The manager might not care which CRM screen normally answers that question. They might not want to learn how the report builder works. They want a reliable answer and a sensible next action.
If an assistant can perform the task using connected systems, the familiar application becomes less visible.
That does not mean the work underneath has disappeared. It means the customer’s attention has moved.
And attention is part of how software companies establish their value.
Salesforce is not just a user interface
There is an important distinction here.
The application layer is more than the screen. It also includes the rules that determine how the business operates: who can access a record, which changes are valid, what needs approval and what happens after an action.
Salesforce describes its headless approach as retaining business logic and governance while making capabilities available across different interfaces. In other words, the intended model is not simply to hand a chatbot an unrestricted database connection.
That makes commercial sense.
If Salesforce remains the trusted place where customer records, permissions and workflows live, it can still be valuable when the user never opens its conventional interface.
But I do not think we should pretend nothing has changed.
There is a difference between being the application a customer actively chooses to use and being the system their preferred assistant happens to call.
The latter can be a strong business. It is also a different competitive position.
The uncomfortable question is who owns the relationship
Imagine you run a specialist software company serving recruitment agencies, property managers or maintenance businesses.
Today, your customer logs into your product every morning. They see your brand, discover your new features and build working habits around your interface.
Tomorrow, they might begin in a general-purpose assistant that coordinates several products.
Your software handles one part of the job. Another supplier handles another. The assistant presents the result.
Who gets the credit when it works?
Who gets blamed when it fails?
Who introduces the customer to an alternative supplier?
And if the assistant eventually makes switching between equivalent services easier, how much of your current advantage survives?
These are commercial questions before they are engineering questions.
My concern is not that every application will disappear. It is that some businesses will make themselves accessible to agents without thinking through what they are giving up, what they are strengthening and how they will charge for the value that remains.
An integration can be a distribution opportunity. It can also weaken the relationship between you and your customer.
Even the largest companies are making bets
It would be unfair to say Salesforce has no AI strategy. It clearly has one.
What I would question is whether any major software company has a complete, proven answer to how the market will work once customers routinely delegate tasks across products.
Building your own assistant, making your application available through someone else’s assistant and preserving a traditional interface are different bets. A company can pursue all three while the balance remains uncertain.
That is not necessarily confusion. It may be rational preparation for several possible futures.
But smaller businesses should not mistake the confidence of a keynote presentation for evidence that the commercial outcome has already been settled.
My reading is that the industry is still negotiating who owns the conversation, who performs the work and who captures the margin.
A small or medium-sized business does not need to predict all of that correctly. It does need to avoid spending its entire development budget on an assumption it has never tested.
What does an MCP server actually do?
MCP stands for Model Context Protocol. It is a standard through which AI applications can connect to external information and tools. An MCP server exposes selected capabilities for a compatible AI client to use.
Think of it as providing a defined set of things an assistant can ask your product to do.
For a booking business, those might include checking availability and preparing a reservation. For a CRM, they might include retrieving an account summary or creating a follow-up task.
But “we should build an MCP server” is not, by itself, a product strategy.
It is a technical answer that comes after several business decisions:
- Which customers would use it, and through which assistants?
- Which tasks would become meaningfully easier?
- What information should be available?
- Which actions require explicit approval?
- How will access, usage and charges be controlled?
- Why would someone continue paying for your service?
MCP standardises part of the connection. It does not establish demand, make every client behave identically or decide what your product should become.
A well-built connection to something customers do not need is still wasted development.
There are several sensible directions, not one
For many smaller companies, the right approach will be a combination of three options.
Put AI inside the existing product
This makes sense when customers benefit from your specialist interface and the surrounding context.
A quotation tool might suggest a scope of work, highlight missing information and explain an unusual price. The customer still reviews the quote in a purpose-built screen.
There is no requirement to replace that screen with a conversation. Sometimes the best AI feature is a useful suggestion at the right moment.
Let external assistants use selected capabilities
This makes sense when customers already work in an AI assistant and would benefit from accessing your product there.
A specialist data provider could make a reliable research service available to agents. A scheduling product could expose availability checks. A business system could allow authorised users to retrieve a status summary.
An existing API integration may be sufficient for a narrow workflow. MCP may be useful when supporting compatible clients is part of the plan.
The choice should follow the customer’s task, not whichever acronym is currently attracting attention.
Let an agent coordinate work across systems
For a business buying software rather than selling it, this may be the most valuable opportunity.
Consider a regional maintenance company. Preparing a quote might involve a customer record, an inspection report, a price list and a scheduling system.
An agent could gather the relevant information and prepare a draft for a person to approve.
The benefit is not “having an AI agent”. It is reducing preparation time, avoiding repeated data entry and making omissions easier to spot.
That is something a smaller business can measure.
Do not throw away the interface too quickly
I do not believe everything should become a chat window.
Try comparing twenty quotations, reviewing a complex schedule or making several precise changes to a financial model through conversation alone.
A well-designed table, calendar or visual editor can be much better.
My expectation is that strong products will combine conversation, direct controls and automation. A user might ask an assistant to prepare something, review it in a structured interface and approve the final action.
The interface becomes less about making people navigate every step and more about helping them understand, judge and control the result.
That is still valuable design work.
For some products, it may make interface design more important, because the consequences of a poorly understood automated action are greater than those of an inconvenient menu.
The business rules must survive the move
One practical danger is exposing capabilities that were previously safe only because a person had to move through a controlled sequence of screens.
A normal interface might prevent an invalid discount, warn about a duplicate booking or require approval before a message is sent.
If those checks exist only in the front end, an alternative access route could bypass them.
My recommendation is to treat agent access as a new operational channel, with its own failure cases.
Permissions, validation and approval requirements should be enforced by the underlying service. Actions should leave a clear audit trail. Repeated requests should not accidentally create duplicate orders. Users should be able to see what happened and correct mistakes.
This is where experienced software development and integration work matters. Connecting a tool is only the beginning.
What I would advise a smaller business to do now
Start with one workflow that is expensive, repetitive or frustrating.
Watch how it actually happens. Identify the information involved, the decisions people make and the exceptions that require judgement.
Then test the most appropriate experience. That could be an improvement inside your existing application, a controlled external connection or an agent that prepares work across several systems.
Measure the complete result:
- Time saved after checking and corrections
- Accuracy and exception rates
- Customer or employee adoption
- Cost per completed task
- The consequences when something goes wrong
For software vendors, add another question: does the change strengthen the reason customers pay you, or merely move their attention elsewhere?
You do not need to rebuild the entire product to learn the answer.
Where Addbox can help
This is the kind of decision that belongs in an AI strategy discussion with Addbox, not just a development ticket.
At Addbox, the starting point is understanding your product, customers and business processes. From there, we can help decide where AI belongs, whether MCP or another integration approach makes sense, and which parts of the experience should remain firmly under your control.
That might mean exposing a small set of capabilities to external assistants. It might mean improving your existing product with AI. It might mean modernising the underlying systems before giving an agent permission to use them.
For a small or medium-sized business, the objective should be a practical advantage you can explain and measure, with room to change direction as the market develops.
Salesforce making its interface optional is not proof that software has become worthless.
It is a warning that the place where customers experience value may no longer be the place where that value is created.
If you build software, what would customers still pay you for if they never opened your application?
And if you run a business, how much of your team’s day is spent operating software rather than doing the work the software was supposed to make easier?
Those are the questions I would start with.
Sources
Need help deciding where AI belongs in your product?
At Addbox, we help software companies and operating businesses decide where AI creates measurable advantage, which capabilities to expose, and which parts of the experience should remain under your control.