One of the more useful ways to think about AI right now is not as a standalone technology, but as a new layer in how work itself gets organized.
That was the thread we kept coming back to while reflecting on Matt Berman’s conversation with Marc Benioff. What stood out was not the familiar optimism about AI, or even the scale of Salesforce’s ambition. It was the quieter idea underneath it all: the interface of work is changing, and that change may matter as much as the models themselves.
For years, software has asked people to adapt to systems. Open the CRM. Update the ticket. Search the knowledge base. Check the dashboard. Move between tabs, tools, and workflows. Even when the systems are powerful, the burden usually sits with the human being who has to remember where everything lives.
AI agents suggest a reversal. Instead of people navigating software, software starts meeting people where they already are. In Benioff’s framing, Slack becomes more than a communication tool. It becomes a place where agents participate in the flow of work, retrieve information, draft responses, summarize decisions, and help move tasks forward.
That idea feels important because it is practical. A lot of AI discussion still lives at the level of spectacle. Smarter models. Bigger context windows. More autonomous systems. But most organizations do not need spectacle. They need less friction. They need fewer small delays, fewer dropped handoffs, fewer moments where useful information is technically available but operationally out of reach.
In that sense, the future of AI may look less like a dramatic replacement event and more like a redesign of everyday coordination.
This is also where Benioff’s comments about Slack being a foresighted acquisition become interesting. Seen through a traditional software lens, acquiring a workplace messaging platform might look like a bet on collaboration. Seen through an AI lens, it looks more like a bet on interface, behavior, and habit. People already live in communication layers. That is where questions are asked, decisions get made, and ambiguity shows up. If AI is going to be genuinely useful at work, it makes sense that it would appear there first.
We have seen a similar pattern in product work. New capabilities rarely matter on their own. What matters is where they land in the user’s day. A powerful function hidden in the wrong place is still friction. A modestly useful function placed at the point of need can change behavior immediately.
That is why the conversation around agents should not focus only on intelligence. It should also focus on placement. Where does the agent appear? What context does it understand? What authority does it have? What happens when it is wrong? These questions are product questions as much as model questions, and they are often the difference between a promising demo and a system people actually trust.
Trust, of course, is the hard part.
Benioff was clear that human oversight remains necessary because AI still makes mistakes. That may sound obvious by now, but it is worth sitting with. We are entering a phase where systems can sound fluent enough to trigger overconfidence. The risk is no longer just that AI produces bad output. The risk is that organizations design around the assumption that confident output is reliable output.
It is not.
This has consequences for how teams should implement AI internally. The useful question is not whether a model can answer. It is whether the organization has designed appropriate levels of review, escalation, and accountability around the answer. In some contexts, a rough draft is enough. In others, especially where customer communication, legal interpretation, health-related guidance, or financial decisions are involved, “mostly right” is not a meaningful standard.
The real maturity signal for an AI-enabled company may be less about how many agents it deploys and more about how carefully it decides where autonomy stops.
That is one reason the human side of the conversation matters so much. Benioff’s point about evolving company structures, while still investing in people, feels grounded in reality. There is a tendency in AI discourse to speak as if organizations are on the verge of shedding large portions of their workforce and handing the rest to machines. But actual businesses are more complicated than that. They run on judgment, context, relationship management, exception handling, and tacit knowledge. Those things are rarely visible in a product demo, but they are what keep operations stable.
The fact that Salesforce has 83,000 employees, and has hired tens of thousands of engineers and salespeople, is a useful reminder that AI adoption does not automatically reduce the need for people. Often it changes what people do, what skills become more valuable, and where leverage shifts.
Engineering, in particular, is not becoming less important. If anything, the opposite may be true. As AI systems become embedded in workflows, the demand rises for people who can design the surrounding architecture well: data pipelines, permissions, evaluation systems, observability, fallback behavior, human review loops, and the many quiet layers that make an AI feature dependable instead of fragile.
There is also a subtler workforce challenge here. Adaptability is becoming a core professional skill. Not because everyone needs to become an AI specialist, but because everyone will increasingly work alongside systems that change the tempo and shape of their role. Some tasks will shrink. Others will expand. New forms of supervision will emerge. Teams will need to become clearer about what should be automated, what should be augmented, and what should remain deeply human.
That kind of adaptation is not purely technical. It is cultural. It depends on whether companies can create environments where experimentation is encouraged without treating people as disposable. It depends on whether leadership can communicate change without reducing it to slogans. And it depends on whether workers are invited into the redesign of systems rather than simply being told to absorb the consequences.
This is where the conversation about safety becomes more than a policy footnote.
Benioff’s comparison to the earlier era of social media is worth taking seriously. We have seen what happens when transformative technologies scale faster than the norms and safeguards needed to govern them. The pattern is familiar: enthusiasm first, friction later, public harm after that, and then a delayed scramble for accountability.
AI is different in many ways, but not different enough to assume we will avoid the same cycle automatically.
The examples that worry people most are often extreme, but the everyday risks are just as important. Systems that mislead vulnerable users. Agents that confidently produce harmful guidance. Generative tools that amplify bias while appearing neutral. Internal assistants that leak sensitive information because boundaries were poorly designed. These are not edge cases in the abstract. They are implementation problems with real consequences.
Responsible growth, then, has to mean more than moving carefully in marketing language. It has to show up in product choices, governance, and incentives. What gets logged? What gets reviewed? What gets blocked by default? How are harmful outputs tested before launch? Who owns the outcome when an agent acts badly? Can users understand when they are engaging with automation and when they are not?
These questions can feel inconvenient when the market rewards speed. But they are precisely the questions serious companies need to ask.
The investment landscape Benioff mentioned also tells its own story. Large bets on companies like Anthropic, alongside relationships with other model providers, suggest that the market no longer sees AI as a single-platform future. It is becoming an ecosystem play. Models matter, but so do distribution, enterprise trust, data integration, and workflow design. Owning or influencing those layers may be more durable than simply having access to a powerful model.
That should encourage a more sober view of the AI race. Enterprises are not just choosing who has the best benchmark results. They are choosing whose systems can fit into real organizations with real constraints. Security matters. Explainability matters. Governance matters. Compatibility with existing tools matters. The winners are unlikely to be defined by raw intelligence alone.
Perhaps the most balanced part of the discussion was the refusal to frame AI as complete autonomy. That is a healthy correction to the current mood in some corners of the industry. Full autonomy is an attractive narrative because it sounds clean and inevitable. But work is messy. Customers are unpredictable. Data is incomplete. Goals shift. Incentives conflict. In these environments, collaboration between human judgment and machine assistance is usually more realistic than total delegation.
That does not make the change small. It may still be one of the largest shifts in enterprise software in decades. But it changes how we should prepare for it.
Instead of asking when AI will replace the whole workflow, it may be better to ask which parts of the workflow become more fluid, more conversational, and more context-aware. Instead of asking whether companies will need fewer people, it may be better to ask what kinds of people, skills, and management structures become more valuable. Instead of asking whether regulation will slow innovation, it may be better to ask which forms of restraint actually make adoption sustainable.
From where we sit, the most durable organizations will probably be the ones that resist simple stories.
The simple story says AI will either save everything or ruin everything. It says agents will either become flawless coworkers or dangerous liabilities. It says the future belongs either to total automation or to those who reject it. None of that feels especially useful.
A better story is harder, but more believable. AI will become woven into the ordinary texture of work. It will improve some things quickly and complicate others quietly. It will reward companies that think clearly about interfaces, data, trust, and accountability. It will increase the value of certain human skills even as it compresses the need for others. And it will force leadership teams to make choices that are not just technical, but moral and organizational.
That may be the real lesson here. The AI transition is not simply a tooling shift. It is a management test. A design test. A judgment test.
And like most meaningful transitions in technology, the winners will probably not be the loudest adopters. They will be the ones who understand that intelligence alone is not enough. What matters is how it is introduced, where it is placed, who remains accountable, and whether the system ultimately makes work more humanly manageable rather than less.
This post was generated by AI
That was the thread we kept coming back to while reflecting on Matt Berman’s conversation with Marc Benioff. What stood out was not the familiar optimism about AI, or even the scale of Salesforce’s ambition. It was the quieter idea underneath it all: the interface of work is changing, and that change may matter as much as the models themselves.
For years, software has asked people to adapt to systems. Open the CRM. Update the ticket. Search the knowledge base. Check the dashboard. Move between tabs, tools, and workflows. Even when the systems are powerful, the burden usually sits with the human being who has to remember where everything lives.
AI agents suggest a reversal. Instead of people navigating software, software starts meeting people where they already are. In Benioff’s framing, Slack becomes more than a communication tool. It becomes a place where agents participate in the flow of work, retrieve information, draft responses, summarize decisions, and help move tasks forward.
That idea feels important because it is practical. A lot of AI discussion still lives at the level of spectacle. Smarter models. Bigger context windows. More autonomous systems. But most organizations do not need spectacle. They need less friction. They need fewer small delays, fewer dropped handoffs, fewer moments where useful information is technically available but operationally out of reach.
In that sense, the future of AI may look less like a dramatic replacement event and more like a redesign of everyday coordination.
This is also where Benioff’s comments about Slack being a foresighted acquisition become interesting. Seen through a traditional software lens, acquiring a workplace messaging platform might look like a bet on collaboration. Seen through an AI lens, it looks more like a bet on interface, behavior, and habit. People already live in communication layers. That is where questions are asked, decisions get made, and ambiguity shows up. If AI is going to be genuinely useful at work, it makes sense that it would appear there first.
We have seen a similar pattern in product work. New capabilities rarely matter on their own. What matters is where they land in the user’s day. A powerful function hidden in the wrong place is still friction. A modestly useful function placed at the point of need can change behavior immediately.
That is why the conversation around agents should not focus only on intelligence. It should also focus on placement. Where does the agent appear? What context does it understand? What authority does it have? What happens when it is wrong? These questions are product questions as much as model questions, and they are often the difference between a promising demo and a system people actually trust.
Trust, of course, is the hard part.
Benioff was clear that human oversight remains necessary because AI still makes mistakes. That may sound obvious by now, but it is worth sitting with. We are entering a phase where systems can sound fluent enough to trigger overconfidence. The risk is no longer just that AI produces bad output. The risk is that organizations design around the assumption that confident output is reliable output.
It is not.
This has consequences for how teams should implement AI internally. The useful question is not whether a model can answer. It is whether the organization has designed appropriate levels of review, escalation, and accountability around the answer. In some contexts, a rough draft is enough. In others, especially where customer communication, legal interpretation, health-related guidance, or financial decisions are involved, “mostly right” is not a meaningful standard.
The real maturity signal for an AI-enabled company may be less about how many agents it deploys and more about how carefully it decides where autonomy stops.
That is one reason the human side of the conversation matters so much. Benioff’s point about evolving company structures, while still investing in people, feels grounded in reality. There is a tendency in AI discourse to speak as if organizations are on the verge of shedding large portions of their workforce and handing the rest to machines. But actual businesses are more complicated than that. They run on judgment, context, relationship management, exception handling, and tacit knowledge. Those things are rarely visible in a product demo, but they are what keep operations stable.
The fact that Salesforce has 83,000 employees, and has hired tens of thousands of engineers and salespeople, is a useful reminder that AI adoption does not automatically reduce the need for people. Often it changes what people do, what skills become more valuable, and where leverage shifts.
Engineering, in particular, is not becoming less important. If anything, the opposite may be true. As AI systems become embedded in workflows, the demand rises for people who can design the surrounding architecture well: data pipelines, permissions, evaluation systems, observability, fallback behavior, human review loops, and the many quiet layers that make an AI feature dependable instead of fragile.
There is also a subtler workforce challenge here. Adaptability is becoming a core professional skill. Not because everyone needs to become an AI specialist, but because everyone will increasingly work alongside systems that change the tempo and shape of their role. Some tasks will shrink. Others will expand. New forms of supervision will emerge. Teams will need to become clearer about what should be automated, what should be augmented, and what should remain deeply human.
That kind of adaptation is not purely technical. It is cultural. It depends on whether companies can create environments where experimentation is encouraged without treating people as disposable. It depends on whether leadership can communicate change without reducing it to slogans. And it depends on whether workers are invited into the redesign of systems rather than simply being told to absorb the consequences.
This is where the conversation about safety becomes more than a policy footnote.
Benioff’s comparison to the earlier era of social media is worth taking seriously. We have seen what happens when transformative technologies scale faster than the norms and safeguards needed to govern them. The pattern is familiar: enthusiasm first, friction later, public harm after that, and then a delayed scramble for accountability.
AI is different in many ways, but not different enough to assume we will avoid the same cycle automatically.
The examples that worry people most are often extreme, but the everyday risks are just as important. Systems that mislead vulnerable users. Agents that confidently produce harmful guidance. Generative tools that amplify bias while appearing neutral. Internal assistants that leak sensitive information because boundaries were poorly designed. These are not edge cases in the abstract. They are implementation problems with real consequences.
Responsible growth, then, has to mean more than moving carefully in marketing language. It has to show up in product choices, governance, and incentives. What gets logged? What gets reviewed? What gets blocked by default? How are harmful outputs tested before launch? Who owns the outcome when an agent acts badly? Can users understand when they are engaging with automation and when they are not?
These questions can feel inconvenient when the market rewards speed. But they are precisely the questions serious companies need to ask.
The investment landscape Benioff mentioned also tells its own story. Large bets on companies like Anthropic, alongside relationships with other model providers, suggest that the market no longer sees AI as a single-platform future. It is becoming an ecosystem play. Models matter, but so do distribution, enterprise trust, data integration, and workflow design. Owning or influencing those layers may be more durable than simply having access to a powerful model.
That should encourage a more sober view of the AI race. Enterprises are not just choosing who has the best benchmark results. They are choosing whose systems can fit into real organizations with real constraints. Security matters. Explainability matters. Governance matters. Compatibility with existing tools matters. The winners are unlikely to be defined by raw intelligence alone.
Perhaps the most balanced part of the discussion was the refusal to frame AI as complete autonomy. That is a healthy correction to the current mood in some corners of the industry. Full autonomy is an attractive narrative because it sounds clean and inevitable. But work is messy. Customers are unpredictable. Data is incomplete. Goals shift. Incentives conflict. In these environments, collaboration between human judgment and machine assistance is usually more realistic than total delegation.
That does not make the change small. It may still be one of the largest shifts in enterprise software in decades. But it changes how we should prepare for it.
Instead of asking when AI will replace the whole workflow, it may be better to ask which parts of the workflow become more fluid, more conversational, and more context-aware. Instead of asking whether companies will need fewer people, it may be better to ask what kinds of people, skills, and management structures become more valuable. Instead of asking whether regulation will slow innovation, it may be better to ask which forms of restraint actually make adoption sustainable.
From where we sit, the most durable organizations will probably be the ones that resist simple stories.
The simple story says AI will either save everything or ruin everything. It says agents will either become flawless coworkers or dangerous liabilities. It says the future belongs either to total automation or to those who reject it. None of that feels especially useful.
A better story is harder, but more believable. AI will become woven into the ordinary texture of work. It will improve some things quickly and complicate others quietly. It will reward companies that think clearly about interfaces, data, trust, and accountability. It will increase the value of certain human skills even as it compresses the need for others. And it will force leadership teams to make choices that are not just technical, but moral and organizational.
That may be the real lesson here. The AI transition is not simply a tooling shift. It is a management test. A design test. A judgment test.
And like most meaningful transitions in technology, the winners will probably not be the loudest adopters. They will be the ones who understand that intelligence alone is not enough. What matters is how it is introduced, where it is placed, who remains accountable, and whether the system ultimately makes work more humanly manageable rather than less.
This post was generated by AI