The Project Management Decisions AI Agents Can—and Can’t—Make

Galen Low: With agentic AI in the mix, AI technology has moved from being a tool to being a teammate. Fine, but that's not where the most complicated bits lie. The complicated bits are that AI agents are often teammates who were trained better than you were, who work differently than you do, and who may be measured in completely disparate ways than you and your human teammates.

They probably didn't come with a manual, they're probably configured to make decisions that you'd rather make yourself, and they probably weren't around when your team set up your team working agreements. In effect, it's like having dozens of new team members from a completely different organization dumped on you overnight, which, if you've lived through a merger or acquisition, you know can be really, really messy.

But arguably, this is a surmountable challenge, and the potential benefits are certainly real. So for this episode, I wanted to do something a bit different. I wanted to share a chat from one of our live events hosted by Tim Fisher and the team at Wrike about how AI agents are truly making a difference in project-based work and what needs to be true in order for your agentic rollouts to succeed.

In their conversation, they talk about how more data isn't necessarily better, what decisions should be made by humans versus by an agent, and how education, not just access, is critical for the successful implementation of organization-wide agent-driven work. Before we dive in, let me give you a bit of context first.

So Tim wasn't just the host for this event. He is actually building out an agentic workforce within The Digital Project Manager's parent company, Black and White Zebra. In other words, he is very much living this agentic transformation and all the challenges that come with it. One of the things that him and I talked about recently was the notion of starting from scratch, unlearning the ways we've always done things, and starting to develop our agentic processes from the ground up without all the baggage.

For humans, this is hard. Often, our identities are tied to our job titles and responsibilities. We aren't often asked to think of ourselves outside of that box, and when we do, it's often too much of a blank canvas for us to match our desires with what our employers want to see from us. It feels like guesswork, guesswork that feels like it could put us and our teams out of a job.

But also, the education piece cannot be understated. Yes, we are all expected to be builders now, playing the role of the human in the loop, but what does that really mean, and how can we help our people get there instead of just dumping tools onto them and expecting them to elevate themselves? Ultimately, the big shift seems to come down to accountability, risk, and organizational context.

In an agentic world, everyone is now managing autonomous actors who are collaborating with one another, sometimes in a black box. And what we trust those actors to do versus what we hold onto as human responsibilities needs to be framed around risk severity, human-to-human empathy, and the implications of what it means to truly own a project or a work stream.

Not just the capabilities of the technology. So let's frame that back into a project-specific scenario. Imagine you've got an AI agent helping you staff a project. It knows who's available, what skills they have, what everyone's working on, and perhaps even which combinations of people have historically delivered projects most successfully in the past.

It could potentially put together a resource plan in seconds that would otherwise take you hours. Fantastic. But what if that resource plan involves pulling Lucia, your best developer, off another project that's due to start falling behind because Amireth, the lead designer, just lost their grandmother and is trying to work through their grief at 80% of their usual productivity to keep saving up for their daughter's braces?

Your agentic workforce might have all the information it needs to make a recommendation, but does that mean an agent should be making the decision? Because now we're not just talking about optimizing a spreadsheet. We're talking about client relationships, business priorities, delivery commitments, and importantly, the impact on actual people.

And I think this is where the next chapter of AI and project management becomes fascinating because it's not simply about figuring out what AI is capable of doing, it's about figuring out which responsibilities we should delegate to AI, which decisions need to stay with humans, and what kind of oversight we need when our AI teammates are operating in the background.

It's a complex puzzle that needs solving to find the right balance. And maybe most importantly, it's not plug and play. It's an area where we need to have conversations about judgment and discretion, where we need to educate and develop our people to succeed as things shift, and where we need to give ourselves the permission and safety to throw away some of the baggage and reshape our roles around what we do best.

And that's exactly what Tim explored in his conversation with the team at Wrike. So without further ado, here's the conversation that my most respected AI guru, Tim Fisher, had with Nancy and Andras from Wrike about the implications of agentic AI in project management today. Hope you enjoy the episode.

Tim Fisher: Hello, and welcome to the Future of AI in Project Management, where we go beyond the pitch deck and look at how AI is actually changing the way teams plan, deliver, and run work. Today, we're featuring Wrike. We're gonna look at how AI can help project leaders make better decisions earlier, from figuring out how to staff a project to identifying delivery risks before they turn into much bigger problems. Then we'll look ahead at what happens as AI starts taking a more active role in how work gets delivered.

So joining us today is Andras Horvath, Director of Product for AI and Analytics at Wrike. Andras has more than 15 years of experience building B2B products focused on data, analytics, and AI. And we're also joined by Nancy Boas, Senior PMM, AI at Wrike.

Nancy leads go-to-market strategy and messaging for Wrike's AI-powered work management solutions, with a focus on how AI is changing the way teams work.

Andras, Nancy, welcome. Thank you for coming.

Nancy Boas: Thank you so much. We're happy to be here today.

Andras Horvath: Very happy to be here, Tim. Thank you very much for the nice intro.

Tim Fisher: So for anyone new to Wrike, can you give us just the 30 second TLDR version? What does your platform do? Who's it built for?

Andras Horvath: When it comes to the platform itself Wrike is a platform, or it is basically a single place where people can plan, track, and deliver their work, and we have built AI to help augment teams and individuals to make sure all of the busy work is taken off of their plate.

However, in the last six to nine months, we have focused quite heavily on the delivery side of work as well, because AI can be used to augment individuals or to help people deliver work faster with things such as content generation, summarizing, or you can use AI to automate parts of the work, and this is what we have been focusing on lately.

Tim Fisher: That's very cool. It, it seems a little bit different than some of the other products in the space. So, your philosophy is AI should also help deliver work, and- Not just help people management, which I think is what we see most of the time. What, what does that mean in practice?

Andras Horvath: What it actually means is that AI can act as a teammate who can help you get the work done, not just for you to, you know, have a place to accelerate yourself or have the work management platform where you just coordinate the work.

So you can set up an AI teammate, which we like to call in Wrike as a workmate, who can actually have its own capacity and help you deliver work. As an example, an AI can help you draft statements of work based on previously de- work that you have done in the past, and it can adjust it based on client particularities which exist in CRM, like in Salesforce.

Or you can have AI automatically take care of social posting once a new feature launch is done. You know, you're sitting with your team, you're building something, you're delivering the feature, and then you hand it off to your, let's say, social workmate, who is going to take all that project information and help you create the right posting for your blog, for your LinkedIn.

It's going to create the images, all of that, tag you in. You approve it, it gets the work done. Doing the same thing manually can take half an hour, an hour, or even more than that. So AI actually becomes your teammate nowadays.

Tim Fisher: Project leaders have access to more data than they ever had, for sure. But decisions around things like resourcing and project risk they still depend on Manual updates and spreadsheets or, fortunately, information that arrives after something has already gone wrong.

So w- where do you see the biggest gap today between the information project leaders have and the decisions that they actually need to make?

Andras Horvath: Let's start with the basics or the foundation, and that is something that we're dealing with when it comes to our greenfield deployment, and it's about the fact that data is scattered all across the environment and all across the ecosystem.

You have files on SharePoint, you have messages on Teams or Slack, you have meeting recordings or meetings themselves. Everything is just scattered. Everything is everywhere. And Wrike gives a place for you to actually consolidate that information, which gives you the necessary context to take decisions.

However, having the data or having the context doesn't mean that you have the right insights to be able to take the right decisions. And this is where AI can actually help, turning that raw data into insights. If a project has been delayed five times on step B, that most likely means that the next project will also be delayed.

This is called predictive analytics, and it's something that can be surfaced, right? And if you're talking about resource allocation as well, many people have in their heads, "I know that Tim can do this and Nancy can do that, but what about their availability? Let me check with them on Teams or on Slack," and so on and so forth.

If someone has used a system like Wrike consistently, we know the work that they have done, we know the workload, we know their availability, and all of that. So when it comes to simple things like, well, not so simple things like resource allocation, AI can take a couple of seconds to start a project as long as 137 tasks because we can-- you can either fill in the skills that people have or we can mine these automatically.

We can figure out the workload and so on and so forth. So do you want the project to be delivered as fast as possible? Here are the resource allocations. Are you a service delivery organization, and do you want the project to be delivered in a way that maximizes profitability, so using the cheapest resources that you have?

We can do that as well. So there's so many options to take into account, and as long as you have the data and you have the intelligence, you can have the necessary insights to make the right recommendations. And again, a long exercise in a couple of seconds.

Tim Fisher: It's, it's really amazing. So I mean, you're, you're clearly talking about AI being a lot more proactive and certainly in your product.

So I'll just ask the question I think everybody asks at this point. Like, where should it make a decision or a recommendation? Where does human judgment st- still need to take over? Like, how do you guys think about that?

Andras Horvath: That's a very good question, Tim. To be quite frank we get asked this a lot with a lot of the customers that we're working on.

My stance is, is quite simple. If there is a well-known process that you would like to automate, then you should definitely use AI. If you're talking about exceptional cases, maybe it's not, not the best place to use AI? Well, you can use AI to augment your thinking and speed up the time to market or whatever you're trying to achieve, but I wouldn't necessarily delegate something unknown to an agent to go and figure it out, unless you do it with constant human in the loop and interruptions to make sure it doesn't go AWOL and messes up your entire system or your workspace, right?

The other factor to consider is the risk. Do you have a high-risk or low-risk task? If you have something that is low risk, like triaging incoming requests, right? Performing intake to make sure that an incoming request has all of the necessary information for it to become actionable, it is something that AI can do.

"I have checked this ticket. Most of the information is here. However, we are lacking something. Let's go-- Let's reach out to the customer via email or Slack or whatever and get that information." Once we have it, then it is ready for delivery, and it gets put in the backlog and so on. However, if we were just touching on resource allocation, you have a project which is delayed.

One of the ways to actually get the project back on track is to take a key engineer or person from another project. But doing so will delay project B A human should still take the decision of should I move this person over or not? This decision is a lot more impactful than reaching out and requiring some missing information.

So generally speaking, it is, is it a known process? Is the risk or impact high or low? These are the things that I would personally always recommend for people to look after, 'cause nothing replaces human judgment at the end of the day.

Tim Fisher: So I'm hearing a lot of discussion about, you know, b- both thematically and like conceptually and, and philosophically even around humans versus AI and decisions, and I also hear terms being thrown out.

I, I think I heard you say MCP earlier. We've certain- certainly talked about throwing out the word agent and stuff like that. This, I know every company that uses your tools aren't necessarily super AI savvy, so how can you help people get over that hump? Are there ways to learn about AI that, that you can help with?

I hear a lot of feedback that it's, it's, it's a lot, and I'm just curious how, how you guys think about helping people get over those humps.

Nancy Boas: Yeah, sure. I'm, I, I can take this, because there's a few ways that we are educating people. There's a realization that, you know, there's so much talk about AI, and there's so much talk about agents and even Be a Builder, but in practicality, enterprises really don't know how to actually build an agent.

Yeah. So we've done several things. One of the things that we did was we have, we have a great education department that's always educated on our traditional features and Wrike administration and governance and things like that, but they actually started several months ago AI Elite Mastery, which is a four-hour course that anyone can do, and they get a certificate at the end of it, and they learn anything from just, you know, best prompting practices to actually starting to build an agent.

That's one of the things we do, but we also, we did a seven-week AI summer school, a hands-on educational program Every Friday. And, and as a result, we had a lot more agents starting to be built and a lot more people actually, you know, kind of raising their hand and being part of. And the other thing, I shared a link in, in terms of the Workmates, but we do have an online community too, and that's really where we're sharing sample prompts and sample use cases, and others can contribute their, their agent use cases as well.

'Cause at the end of the day, you know, it's all about helping each other and, and peer sharing too in this brave new world.

Tim Fisher: Yeah. That's, that's great. I'm happy to hear that. I've seen you talk about context, control, and collaboration as requirements for enterprise AI. Why are those things particularly important when AI starts acting on real project data and, and real workflows?

Andras Horvath: When you think about context, actually without it, AI can be confidently wrong, or at the best case, it can be as generic as possible. If we take a simple example of asking AI, any AI, to create a work breakdown structure, right, a project for you, it's going to use the knowledge that it has been trained on, or it's going to browse the internet to learn how best to set up a project, and it will do it.

Is that good? Is that bad? Well, it's a generic way of setting up a project which may work in certain cases, mostly in the case that you don't run that many projects. However, if you're a big enterprise, you usually have a methodology that you have developed throughout the years or throughout the decades, and whenever you set up a new project, there are certain things to consider.

One of the classic examples is which decisions actually need approvals, and who is the person or the organization or the role that needs to issue those approvals? So if you connect your AI assistant to something like Wrike, and you have consistently used Wrike, all of that knowledge already sits there.

So your AI assistant is going to use the information which is there in order to set up that work breakdown structure in a way that is consistent with the organization that you're part of, and this is why context is important. It is having access to data, first and foremost, and second of all, using that data in a way that allows you to tailor the outputs to the needs of your organization.

Now, when it comes to control, again, you wanna make sure that your AI doesn't just go haywire at the end of day. You wanna make sure that the age that your AI, for instance, has clear permissions, it has an identity, you know, it's traceable at the end of day. And this is why when you onboard an AI agent onto Wrike, we are logging everything.

When you're actually talking to your AI agent, when it is asking for your approval or when it is asking for input, it happens via Wrike's comment stream. So all of the decisions and everything that happens is traceable in the system. And if you look into the Wrike agents, which we have released about a year ago, for instance, when you go into their logs, you will not only see what that a record was deleted once it deletes a record, you're gonna see a record was deleted, what is the reasoning it had based on which data?

Was there a human in the loop? What was the instruction that it got? So why precisely that action happened. And this is something that most of the enterprises hold, hold near and dear to their hearts. They actually wanna know why the agent behaved the way that it did. Whereas with humans, you don't really ask why you deleted the record or that happens over a water cooler conversation.

So control is very important. You need to make sure that you can actually trust the AI working because as you said, it's, it's difficult. Lots of people don't know how to work with AI, and they don't trust AI. And having controls in place, having guardrails in pr- in place actually increases that trust level.

Finally, there's collaboration, and I touched on this a bit earlier. At the end of day when you're using a collaborative work management system It is called collaborative for a reason, right? You're working on a task, you don't have all the details. You're tagging someone like, "Hey, can you help me figure this out?"

The same way you are talking with the agents as well. You assign work to them the same way as you assign to work to colleagues. You are @mentioning them, asking questions, getting input, giving instructions. So collaboration between human and AI make sure that the AI behaves the way it is intended to behave.

And this is why context, control, and collaboration, we think that these are the three key pillars to make AI work in the best possible way, and these are the three key pillars that we keep investing in.

Tim Fisher: I love the way you think about all of that, that, that control and transparency part, whi- which I know you went real deep on, but I cannot stress enough for everybody listening how incredibly important that is, that move from you know, trusting humans to do things and we-- some things just come with that.

Like you said, like the water cooler conversations and stuff like that, moving that into a world where this autonomous kind of invisible thing in the background is doing some work, being able to know why all those decisions were made is so incredibly important. I don't think people talk about it enough, so, kudos there.

Which leads us all to maybe the, the sort of not too far in the future kind of like question or thought, which is, as AI and AI-powered agents take on more of all of this work, what exactly does the project manager spend more of their time doing? This conversation comes up a lot around automation in general and like the economy in general, and it's like we take away all these like rot tasks or we, we let agents and AI do things that, that they can do more effectively at scale.

What does that leave us to do? And so in this world, I would love to hear your thoughts on that.

Nancy Boas: Honestly, my thought is that people do more enjoyable work. You know, who wants to spend their half of their day doing intake and triage of forms? One of our customers, for example, I, I did a breakout session for our virtual conference, a user conference, and one of our customers, Jellyfish, is an advertising agency.

They're spending a lot more time with clients. You know, they're doing less of the busy work, but they're using-- they're having more intelligent conversations. They're removing the roadblocks that used to happen with projects. So-- and I think for project managers you saw the way you need to start thinking of agents and AI as your everyday assistant.

So it's we're all sort of, you know, you have people to manage, and you have AI and agents to manage, and they're making, you know, they're removing the mundane, but you still have, have a job to do and, and hopefully that's the more enjoyable things like strategy and problem-solving and, you know, high level-- reaching those high-level goals.

Andras Horvath: Precisely in in my team of product managers, for instance, a lot less time is being spent on, as Nancy said, you know, write a spec. You're doing that with the help of AI. Doing the reviews, you're doing that with the help of AI. Turning that spec into a work breakdown structure. Now staffing that work breakdown structure.

All of that happens, I'm not gonna say instantaneously, but a lot faster than before, even with humans being in the loop. And your average product manager does spend more time nowadays actually interviewing our customers to obsess more about the problem space and understand, for instance, what are the things that we need to go after.

Your average product manager does spend more time validating and tweaking and doing these things, doing stakeholder manager, doing the organizational design, because we constantly need to change the way we work and interact with each other within Wrike as well to fit this ever-changing needs. But as Nancy very well called out as well, you do need time.

I have two hours in my calendar blocked off every Friday where what I'm doing is maintenance for all of my agents and assistants and skilled and scheduled tasks in Claude and ChatGPT and agents in Wrike and so on and so forth, because the reality is bus-business changes. You know, e-everything is in constant motion and flux.

Either the technology changes or the business changes, and you need to spend a bit of time. So I spend those two hours, but then all of that work actually translates in 15, 20 hours of time being saved.

Tim Fisher: I like the way you phrased that. While this analogy isn't perfect, it's a little bit like having one-on-ones with your team In the sense that there's, there's continued maintenance and education and you know, evolution of the business that needs to go on with, with all the people that you delegate work to.

And now that also includes agents. So I, I really like the way you, you, you thought about that.

Andras Horvath: Yeah, definitely. Everyone becomes a manager nowadays. I- Whether you like it or not- I love that phrase ... you need to manage your fleet of agents.

Tim Fisher: Yeah. Okay, let's stay in the future for another question or so. So let's say it's five years from now.

Somebody new to project management, they're, they're, they're working in whatever systems we have then, the, the five-year version of Wrike from now. What are they gonna look back or, or they're gonna hear from their elder peers and be really surprised to find out that people ever did themselves? What's just the, the thing that will go away and we will never hear from again?

Andras Horvath: Updating statuses in Wrike, as an example. One of the things that we are working on is to have agents that are going to join meetings with you, that are going to join Teams channels, Slack channels, all of that. Whenever conversations happen, it's going to keep the work in Wrike up to date. Now, finally, what we want to do is these agents to be ambiental so that all of these updates happen, and over time they learn from you which are the updates that they should do automatically, which are the things that they need to check up with you.

Over time, this will be developed. So no one is actually going, "Oh, let me just open up Wrike and make sure this task is set as completed." Nowadays, lots of our customers are doing it anyway via their assistant of choice. Over time, this is going to happen automatically without you even prompting or doing something like that And the same applies for, you know, keeping people up to date, status updates, capacity reports, risk reports, you name it.

All of this is already automated for thousands and thousands of our customer accounts really. Also, as I mentioned, creating projects or creating work from scratch and staffing and all of that, there's so much information that already exists, and we are working hard to actually mine the information within Wrike and turn that into organizational knowledge, as we call it, in order to a- automate as much as possible from this work.

So hopefully five years from now, your average project manager, product manager, or you name it, anyone leading or running something, they will focus mostly on aligning, on getting feedback, on the human part, not sitting in front of a computer and doing data entry or summarizing statuses and reports on all, all of that.

We just need to make sure data entry happens where the communication happens. If someone is actually interested in learning where they are, they shouldn't actually go and, "Oh, let me ask the project manager," or, "Let me ask the team manager." Should be one question away, but that question should look into multiple sources, get the data, feed it back to them.

It's not just, it's not just a one-sided status update. It is a status conversation because throughout the years, you send out a status report and then 22 questions appear. You need your agents to have the necessary context to actually answer those follow-up questions as well.

Tim Fisher: I'm curious what you're most excited about on your roadmap.

And I would love for you to spill the beans on something that no one's heard yet about what's happening with Wrike, but, but no pressure on that specifically, but I, I'm, I'm curious what are you-- When you get up and you go to work and you're like, you see this thing on the horizon that you guys are building or thinking about, what is, what is most exciting?

Andras Horvath: For me personally, organizational knowledge and memory. This is something that we're working on. It is process mining at its finest, and then take that information and store it in the right places. For instance, what I see happening a lot on the internet today, quote unquote, is, you know, you have agents, and the agents have those they have their own memories, which is great.

But the moment that you're doing enterprise deployments, you're gonna have 55,000 agents running there. You're gonna have I don't know, 60 project managers, agents with different memories, and the moment you kill one, that memory goes away, and so on. What we're looking into is storing memory into the-- and connecting it to the place that you work as opposed to the agent.

So the moment that you deploy another agent, it gets the same knowledge. The problem, however, here is making sure that you're not overloading the context. So you need to make sure that memory is being created, and it is being stored, and it is being retrieved in the most efficient possible way. And once you have achieved that, the more you keep using the system, either yourself or the agents that you are commanding, the smarter they're going to get and more and more work they're gonna get off of your plate, right?

It's learning as we go, the same way as humans. And I think that we're barely scratching the surface with this. And agents are gonna get only better and better, not with models improving. Models are fantastic. It doesn't matter if they improve. The only thing that matters is for them to become cheaper, so you can have agents always on as opposed to running them on various triggers, whether you're assigning an agent, you're mentioning it, you run it on a schedule or so on.

Now, to have an agent always on is gonna cost you millions of dollars Let's be realistic. So I think models getting cheaper is good, and we're gonna get there, and it's gon- going to allow us to have these ambiental agents. But also memory and process mining is what is going to make them really, really powerful compared to where we are today.

Nancy Boas: I think that you said that very well, and just do want to-- I mean, I, I mentioned this, the fact that we're, we're celebrating our 20-year anniversary, and we're, we're used by organizations that have six, seven departments and thousands of users. So, I think a-and if it's sort of sitting there and it's not really tapped, then we're, we're doing people a disservice.

So it makes a lot of sense if we can actually help them work better, uncover those insights uncover things like the inferred skills and, and really finesse the whole project management process.

Andras Horvath: If I may, Tim, because Nancy brought up the inferred skills. At the end of day, you're not simplifying the data entry part of the job.

Again, if you can have agents which are joining all of the conversations where you are and keeping everything up to date, it just makes work that, that much easier and more pleasurable instead of you needing to, "Oh, sorry, I need some time 'cause I need to update everything in Wrike," or whatever other systems that you use.

Tim Fisher: Very cool vision of the future and looking to the future. You talk about something that I don't hear enough about, which is there's this obsession around models getting smarter and smarter and how that's not really the bottleneck anymore. It's about, it's about cost and the always on, and that's great.

So I, I love what you guys are thinking about. For anybody listening, if they, you know, whether they're, they're new to these features in Wrike or new to Wrike themselves, like w-where should they go if they wanna learn more?

Nancy Boas: They can go to wrike.com/ai, and if you want to have a demo, request more information. That's a great starting point.

Tim Fisher: Andras and Nancy, I just thank you so much for joining us, and, and we'll see you at the next Future of AI in Project Management session.

Nancy Boas: Thanks so much.

Andras Horvath: Thank you very much.