Your Next Program Team Member May Be an AI Agent
During my time at BNY, I saw the firm invest in more than AI tools. It was also building the governance, employee training and internal capability needed for people to experiment with AI responsibly.
That foundation mattered. It gave employees room to explore practical uses of AI without treating innovation as an uncontrolled activity. People could develop ideas, test them within defined boundaries and gradually build confidence in how the technology should be used inside a large financial institution.
BNY has since taken that approach much further. The bank now describes having more than 100 “digital employees” operating alongside human colleagues. These agents can be assigned specific tasks, given login credentials and nicknames, and placed under the supervision of a human manager.
I find this development particularly interesting because it does not feel disconnected from the foundations I observed. Creating an agent that can perform real work inside an enterprise requires more than access to a capable model. The organisation needs people who understand the technology, governance that allows responsible experimentation, and systems that can provide an agent with controlled access to the information and applications it needs.
Other organisations are likely to follow a similar path as their AI capability matures. Some will develop their own agents using internal platforms and business knowledge. Others may purchase specialist agents from technology vendors and integrate them into existing teams and workflows.
Either way, the management question will be the same. What happens when an AI agent is no longer simply a tool used by the program manager, but a member of the program team with defined responsibilities, system access and recurring work?
That may sound like a small distinction, but it has significant implications for how work is allocated, reviewed and governed. If an agent produces work on behalf of a program, someone must manage it. Someone must decide what it is allowed to do, determine whether its output is reliable and accept responsibility when it gets something wrong.
For me, that is why “digital employee” may be the right language. It establishes that the agent is not
operating somewhere outside the delivery model. It sits within a line of accountability.
The risk is that the same language could cause us to place too much trust in it.
What would I assign to my first agent?
There is no shortage of work within a large program that could be supported by an agent.
My first program-management agent could collect updates from workstream leads, review delivery plans and RAID information, identify inconsistencies, draft the weekly program report, follow up overdue actions, track dependencies, prepare steering committee materials and maintain decision records.
Much of this work is important, but it is also time-consuming. Program managers often spend a considerable part of the week gathering information from different teams, reconciling contradictory updates and restructuring the same material for different audiences.
The work is rarely difficult because of one complex calculation. It is difficult because the information is fragmented.
A milestone may be marked green in the program plan while the risk register describes an unresolved issue that makes the date unlikely. A dependency may appear in several reports but have no confirmed owner. A decision may be discussed in a meeting but never make its way into the decision register. A status report may claim that an activity is substantially complete even though testing has barely started.
A well-designed agent should be capable of searching across these records and identifying where the story does not add up.
That would be useful. It would not, however, make the agent the program manager.
The move from draft to final still matters
My initial rule would be straightforward: an agent can prepare, analyse, compare and recommend, but anything moving from draft to final requires human review.
I would apply this in much the same way that I would support a junior project manager. I would expect them to do meaningful work and develop increasing independence, but I would still review important material before it was presented as an official program position.
That review becomes particularly important for steering committee papers, executive reporting, risk descriptions, financial information, stakeholder commitments and decision records. These documents do more than summarise activity. They influence funding, priorities, escalations and executive decisions.
The review cannot be a quick approval at the end of the process. A program manager must remain close enough to the delivery detail to recognise when an important issue has been missed or when a technically accurate summary creates a misleading impression.
There is a danger that agents will make poor information look more authoritative. They can produce clear language, polished formatting and plausible explanations even when the underlying evidence is incomplete.
The output may look ready before it is reliable.
That is where the concept of a digital employee becomes both useful and dangerous. It encourages us to allocate responsibility to the agent, but it may also encourage us to assume a level of competence and judgement that has not been earned.
Would you hold a weekly performance review with an agent?
BNY’s model includes human managers training agents, reviewing their performance and
conducting quality control. KPMG is approaching the same shift through workforce design, encouraging organisations to examine work end to end and decide which responsibilities should remain human, become hybrid or be assigned to digital workers.
This raises a practical management question: should an agent have something resembling a weekly performance review? I think it is worth trialling.
Most people already provide feedback to AI through corrections, revised prompts or simple approval and rejection mechanisms. A more deliberate weekly review could examine the work the agent performed, where it was accurate, where it needed substantial correction and what it failed to identify.
A program manager could review what the agent produced during the week, which outputs required rework, whether it identified inconsistencies the program team had missed and whether its conclusions were supported by the source material.
The review could also identify repeated errors and determine whether the agent needs different instructions, additional context or changes to its system access. This would not be a theatrical one-to-one meeting in which we pretend the agent is human. It would be a structured review of digital capability.
We are still in new territory, so I would be reluctant to prescribe a mature performance-management framework before organisations have enough practical experience. But I would rather see managers actively reviewing an agent’s performance than assuming it will improve simply because people occasionally correct its output.
Speed is not the same as performance
An agent may produce a program report in minutes, but that does not mean it has performed well.
The quality of the work would need to be measured through accuracy, completeness, traceability and the amount of human rework required. I would also want to know whether the agent was identifying meaningful problems rather than merely restating information already visible to the team.
Consider a decision register. The agent may be able to review meeting transcripts and populate decisions quickly, but the register is only useful if the decisions are accurate, clearly worded and attributed to the correct owner. A mistake could result in teams acting on a decision that was never actually made.
The same applies to overdue actions. An automated follow-up may save time, but it must distinguish between a genuinely overdue commitment, an action that has been superseded and an item that was recorded incorrectly in the first place.
For steering committee materials, the standard must be higher again. The report needs to represent delivery reality, not simply combine the latest workstream updates into a professionally written document.
The best agent may not be the one that drafts the report fastest. It may be the one that finds the contradiction everyone else overlooked.
When the agent is wrong, the manager remains accountable
If an AI agent reporting into my program produces inaccurate work or misses something material, the accountability will sit with me.
That may be uncomfortable, but it is no different from the basic responsibility of leading a team. Managers cannot present flawed work to an executive committee and avoid accountability by saying that a junior team member produced it.
The same principle should apply to an agent. “The AI wrote it” will not be an acceptable explanation for an inaccurate steering committee report, a missed dependency or a decision that was recorded incorrectly. The manager will need to explain how the output was reviewed and why it was accepted.
This is another reason the employee identity is more than a novelty.
An agent needs an identity because the organisation must know what it accessed, what it changed, what it produced and who was responsible for supervising it.
The employee-style model may help establish that accountability. A distinct identity, controlled login credentials and an assigned manager make the agent visible within the organisation rather than allowing it to operate as an anonymous capability behind a process.
However, identity should not be mistaken for judgement. Giving an agent a name and a place within the team does not make it capable of accepting responsibility for its own mistakes.
The accountability remains human.
The program manager of the future may manage several agents
As enterprise AI platforms become more integrated and specialised, I can imagine a program manager supervising several agents operating in different capacities.
One may focus on plans and milestones. Another may monitor risks, issues and dependencies. A financial agent could compare forecasts with actual expenditure. A reporting agent could prepare different views of the program for executives, delivery teams and external stakeholders. A decision agent could maintain the history of important choices, including the evidence and assumptions behind them.
These agents may also interact with one another, passing information between specialised roles as part of a wider delivery process.
The way organisations obtain these agents may differ. Large enterprises with mature AI platforms may build agents around their own processes, data and control environments. Others may purchase specialist agents from vendors, much as they currently buy software platforms or managed services.
A purchased agent would still need to be integrated into the organisation’s access controls, delivery processes and accountability structures. Buying the capability would not transfer responsibility for how it is used.
The effectiveness of the program manager would then depend partly on how well they manage this digital capacity. They would need to allocate responsibilities clearly, avoid duplication, establish approval points and recognise when one agent’s output conflicts with another.
This could create a significant difference in delivery capability between peers.
A program manager who understands how to organise and supervise several specialist agents may be able to maintain deeper visibility across a complex program than someone relying largely on manual reporting and fragmented updates.
The differentiator will not necessarily be access to AI. Most organisations will eventually have access to similar capabilities.
The differentiator may be management.
More time for leadership, or simply more work?
The optimistic outcome is that agents reduce the administrative burden on program managers and create more time for the work that genuinely requires human leadership.
That includes stakeholder engagement, negotiation, decision-making, conflict resolution, coaching, intervention and benefits realisation.
These are often the areas where experienced program managers create the greatest value. They are also the areas most likely to be neglected when the program manager is consumed by reporting cycles and information collection.
There is, however, another possible outcome.
Rather than allowing program managers to spend more time on leadership, organisations may increase the number of projects assigned to each person. The productivity gained through agents could be converted into more change, wider portfolios or lower headcount.
My instinct is that we will see both.
Some organisations will use agents to improve the quality of leadership and delivery decisions. Others will primarily use them to increase capacity. The pressure to demonstrate productivity and reduce cost will be difficult to resist.
This is why leaders should be clear about the outcome they are designing for. If the only measure of success is the number of projects one program manager can carry, agents may simply create a more efficient version of an already overloaded delivery environment.
The better question is whether the program is being managed with greater insight, stronger decisions and earlier intervention.
A new management discipline
AI agents will not remove accountability from program managers. They may increase it.
Managers will need to understand what their agents are doing, where their information comes from, how their outputs are reviewed and when they should be prevented from acting independently.
They will also need to learn how to improve an agent over time. That could become as important as coaching human team members, although the method will clearly be different.
Calling an agent a digital employee does not mean pretending it is human. It means recognising that once it is performing recurring work inside the organisation, it needs an assigned role, a clear manager and measurable expectations.
It also needs scepticism. The moment an AI agent becomes part of your team, its mistakes become management failures rather than technology excuses.
The practical question for program managers is no longer whether agents will enter the delivery environment. Banks are already experimenting with identifiable agents, human managers and combinations of human, hybrid and agent-led roles.
The question is whether we are ready to manage them. If an AI agent joined your program team tomorrow, what would you allow it to own, what would you review each week and what would you never allow it to finalise without you?
Comments