Microsoft’s 2025 Work Trend Index introduced the “Frontier Firm”: an organization that uses AI throughout its work, with people delegating tasks and workflows to agents. I find the idea useful because it asks what changes around the technology: who does the work, who checks it and how people learn to use it.
The report captured considerable ambition. In Microsoft’s survey, 82% of leaders saw 2025 as a year to rethink strategy and operations, and 81% expected agents to become moderately or extensively integrated into their company’s AI strategy within the following 12 to 18 months. Twenty-four percent reported organization-wide AI deployment; 12% said their companies remained in pilot mode. Those are respondents’ descriptions and expectations, rather than a measure of the value already delivered.
Investment was growing too. Stanford’s 2025 AI Index put corporate AI investment in 2024 at $252.3 billion. Within that, private investment grew 44.5% from the previous year.
I had led cloud migrations, security programs and early AI deployments, and the pace felt unusually fast to me. My own experiments had mixed results. They made me interested in what this framing would require of an IT organization in practice.
Buying and integrating AI capability
Cloud APIs make it possible to try a model without building the underlying infrastructure. Someone still has to assess the supplier, understand the contract, monitor cost and design an application that can cope with changes in the service.
Our earlier move to the cloud was a substantial help here. When new AI services appeared, we could test their value in weeks rather than spending months preparing the environment. The architecture work had given us options that became useful later.
Giving agents a place in the workflow
In one transformation project, we mapped processes with the operations managers and treated the AI as a participant in the work. We asked which steps it could take on, such as gathering data or drafting analysis, and which decisions needed a person.
That brought practical questions into the conversation early. What access should the agent have? How would someone know whether it had done the job well? When should it stop and ask for help? Describing it as a teammate helped us discuss the handovers, provided we stayed clear about who remained accountable.
Helping people delegate
Microsoft also asked about emerging responsibilities. In its survey, 28% of managers were considering hiring people to manage a workforce of humans and agents, and 32% planned to hire AI agent specialists within 12 to 18 months.
I was just as interested in what the people already in an organization could do. An internal program that taught analysts to delegate data preparation to a generative AI tool surprised me: some of the least technical colleagues became its most frequent users once it was accessible to them. We had underestimated how much they could do with a useful tool and permission to try it.
When the software works and people still hesitate
We once rolled out an AI analytics tool for sales teams. The software worked, but adoption was very low. People did not trust its insights, and some worried that it would undermine their role.
What helped was explaining how it could take on data preparation and leave salespeople more time for the work they valued. Respected early users also helped colleagues try it. That experience changed how I planned later work: I made more room for training, communication and listening alongside the technical build.
It also made me cautious about assuming that enthusiasm in the leadership team reflected how everyone else felt. Microsoft’s survey found leaders ahead of employees in familiarity with agents and expectations of their value. An organization has to make space for that difference.
What earlier governance work made possible
Our ISO 27001 work gave us established ways to handle supplier risk, access management and compliance evidence. When we piloted AI workflows involving data, we could use that work rather than start every discussion from scratch.
Security and compliance colleagues needed to be involved while we could still change the design. A concern raised at that stage was something we could work through together. Discovering it at the end could mean revisiting decisions that had already shaped the system.
Results took iteration
An AI scheduling optimizer we deployed in a logistics function had underwhelming results in its first quarter. We needed to improve the data inputs and how people interpreted the suggestions before the efficiency gains appeared.
That experience is why I want to see how people use a system and what they struggle with while waiting for the broader results. Adoption, cycle time and user feedback help explain what an early financial measure cannot yet show.
McKinsey’s 2025 Superagency in the workplace report put numbers around the gap between expectations and reported returns: 92% of companies in its survey planned to increase AI investment over three years, while 19% reported revenue increases of more than 5% attributable to AI. For an individual program, I would still want a clear account of its costs, progress and remaining assumptions.
What I would carry into another program
I would begin with a small number of business problems and a view of where the organization needed to be in three to five years. That would guide the experiments, with priorities reviewed as we learned.
I would also check the foundations for each use case: the data, access, capacity and people required to operate it. Then I would start with one or two pilots that had an owner in the business, measure them and plan further rollout with the departments that had learned how to use them.
The best transformation I worked on had HR shaping the retraining, finance helping quantify the benefits and operations choosing the problems to address. The strongest AI champion I worked with was a non-technical operations manager. She saw a way to solve a scheduling problem she had carried for years and co-led the rollout. Her team’s experience gave colleagues a much more concrete reason to try it than a presentation could.
Sources
- Microsoft Work Trend Index 2025: 2025: The Year the Frontier Firm Is Born.
- Stanford HAI: AI Index Report 2025, economy chapter.
- McKinsey: Superagency in the workplace: Empowering people to unlock AI’s full potential (2025).
