“I don’t think we’re ready for them.” That is Neil Thompson, Director of the FutureTech Research Project and Principal Investigator at MIT’s Initiative on the Digital Economy, talking about AI agents.
The episode focused on a single question many companies are asking right now: what does it take to make agentic AI real?
Only some tasks within a role can be handed to AI, so the real shift is job redefinition, not job loss.
Looking further ahead, Thompson pointed to a risk most companies have not yet begun to weigh seriously: what happens when agentic systems move from screens into the physical world.
The economic questions remain open, the strategy is still being explained through Marvel analogies, and the basics of engineering discipline have not changed nearly as much as the hype suggests.
“I don’t think we’re ready for them.” That is Neil Thompson, Director of the FutureTech Research Project and Principal Investigator at MIT’s Initiative on the Digital Economy, talking about AI agents. Not the agents themselves but us: organisations, workforces, and systems built for slower change.
Thompson was one of three experts featured in a special episode of All About AI, hosted by Abhishek Singh, Associate Editor at LiveMint, in partnership with Salesforce. The episode focused on a single question many companies are asking right now: what does it take to make agentic AI real?
Why the numbers don’t add up yet Thompson’s research has tracked how AI is changing work faster than previous technology waves, comparable to the shift computers brought in the 1980s but faster. That speed makes the maths difficult to plan around.
“We would hope that the extra value created would give people new places,” he said, “but right now we’re not ready for it, and that churn will be a real challenge for the people in those jobs.”
Thompson’s research found that most jobs are not fully automatable. Only some tasks within a role can be handed to AI, so the real shift is job redefinition, not job loss. He has also seen the opposite in boardrooms: companies pouring money into pilots without a clear sense of success. “If I was a CFO, I would be very nervous,” he said, describing organisations that could not explain what their AI spend was meant to achieve.
Looking further ahead, Thompson pointed to a risk most companies have not yet begun to weigh seriously: what happens when agentic systems move from screens into the physical world. A flawed model output today might mean discarding a bad piece of marketing copy. A robot that malfunctions, even slightly, could cause injury or a traffic jam. That, he said, is a much harder teething process than anything the industry has faced so far.
His clearest warning concerned a shift most companies have not yet begun to consider. As AI adoption matures, the real differentiator will not be which model a company uses, but its data strategy and whether its data advantage is worth protecting or trading away.
Trust, and the Wolverine problem If Thompson’s segment was about economics, Shibani Ahuja’s was about translation, turning a technical shift into something a chief executive can act on. Ahuja, SVP of Enterprise IT Strategy at Salesforce and a board member of Salesforce Canada, met with over 587 C-suite leaders last year and says the same question comes up almost every time: what exactly is agentic AI, and can it be trusted? Her explanation used comic books rather than technical language. Generative AI, she said, is like swapping in a new Iron Man arm, a tool made for a specific job. Agentic AI is like rebuilding Wolverine’s ‘adamantium’ skeleton: the full structure a company has built over the years, its workflows, applications, and technology stack, is now open to an autonomous agent to traverse.
The most common mistake she sees is treating this change as a technology project. “It’s beyond just the tech stack,” she said, explaining how CIOs now work closely with CHROs to answer a practical question: how does digital labour fit alongside human work, and how should its performance be judged?
She also flagged a quiet shift in how companies measure return on investment. Where AI projects were once justified purely by cost savings, Salesforce started building an agent around leads that would otherwise be discarded, ones too low-scoring to follow up on manually. Nurturing those forgotten leads through an agent generated several million dollars in closed deals and a healthy pipeline, turning what had been a cost centre into a revenue one.
Software is still software The final conversation, with Yakov Salomon, EVP of Software Engineering at Salesforce, brought the discussion back down to earth. His warning was simple: “Everybody likes vibe coding. Nobody likes vibe maintaining.”
Salomon argued that agents, no matter how advanced, are still software, and everything engineering teams have learned about testing, monitoring, and alerting still applies. What changes at scale are latency, since agentic reasoning cycles are not yet as fast as systems built for recommendation engines in commerce, and data quality, since agents inherit every flaw in the information they are trained and run on.
He was just as direct on privacy. Every time an agent is handed a user’s credentials to act on their behalf, a boundary is crossed. The industry is still working out how agents carry identity and authorisation as they move between systems and hand off tasks to other agents. On ethics, Salomon suggested thinking about training an agent the way one would train a person, through enough examples of what to do and what not to do in a given scenario to handle the grey areas no policy document can fully predict in advance.
Three lenses, one conclusion. The economic questions remain open, the strategy is still being explained through Marvel analogies, and the basics of engineering discipline have not changed nearly as much as the hype suggests. What has changed is how quickly organisations are expected to work all three out at once.
