Talk to us
Most enterprises have AI pilots, but few have agents in production. Jonas Dahlberg, Head of Growth at Redeploy, explains why the technology is the easy part and the organizational change is the hard one.

00:00 Emelie FågelstedtHello and welcome to this talk on what it means for enterprises to be agent-ready, but also why most still aren’t. In this conversation, we’re going to explore what it actually means to start working with AI agents at scale as an enterprise. We’ll go through some of the technical parts, you know, the infrastructure, the data security, accountability, but also focus a lot on the people, because at the end of the day, maybe the organizational shift is actually what makes all the difference here. My name is Emelie Fågelstedt, and I’m an independent AI and innovation consultant. And with me today I have Jonas Dahlberg, who is the head of growth at Redeploy. Welcome, Jonas.
00:48 Jonas DahlbergThank you. Really glad to be here.
00:49 Emelie FågelstedtSo am I, that you are here. So, all this talk about AI agents, what are they?
00:56 Jonas DahlbergAI agents is AI that can act. That’s the easiest way to explain it. You know, when you use an AI system like ChatGPT, Copilot, etc., you ask it to do something, right? This text summarizes. And what an agent can do is that it can automate a flow or a process, meaning it can not only read the email you get, it can summarize it, compare that to the notes you have in the CRM and automatically reply.
01:27 Emelie FågelstedtSo I guess that’s the difference, because a year ago we were still talking about generative AI and these AI bots. But you’re saying that AI agents can do more.
01:37 Jonas DahlbergI tend to think of them as a colleague. Basically, you can make an agent that, you know, scans through your documents of the day, summarizes the most important things you missed, and then the next morning you get an email summarizing that. But you can also use it to automate business processes. And that’s where it gets really interesting.
02:00 Emelie FågelstedtYou sound so confident now. So I have to ask, I mean, is everyone already using these agents? Am I behind, or what is actually the state of the agentic AI shift?
02:11 Jonas DahlbergI mean, obviously, as always with these kinds of things, there’s a lot of talk and very little action. So in general terms, very few have done something significant. There are some that are doing real things, really addressing core processes, automating work, etc. But the difference between those that are leading and being part of this and the ones that are not is just increasing every day.
02:43 Emelie FågelstedtWhen it comes to AI agents, and this whole AI era that we’re now entering into, how many companies, of all the organizations you meet at Redeploy, have actually started using this at scale?
02:59 Jonas DahlbergMaybe ten, 15, 20% of our customers are really pursuing real core business processes. Say another 50% are thinking about it. And a couple have not started thinking about it yet.
03:16 Emelie FågelstedtAnd I mean, honestly speaking, is this just a buzzword? Are agents really as great as they seem to be, when you read all the headlines and the big tech CEOs talking about it? Or is this really as good as it sounds?
03:32 Jonas DahlbergI think it’s actually a lot better. It requires a different way of thinking. I mean, you need to move away from the AI assistant and think more like: if I had an extra colleague that worked in parallel with me and did its own things, what would I ask it to do? You can automate things using AI, and it’s the reasoning that makes it all possible, reasoning and memory, context, etc. So you can do pretty advanced things. You know, if you look at a developer these days using Claude Code, you can run a couple of agents in parallel. You provide it with what you want the application to do. You then have an agent that creates the code. You have another agent that writes the test cases. You have a third agent that creates the user acceptance criteria, and then you have a fourth agent that orchestrates all of these and makes sure that whenever there’s a change in the interface that you are prompting for, it changes the test cases and acceptance criteria. And then you can have a fifth agent that basically just makes sure it’s pulled together at the end.
04:48 Emelie FågelstedtSo you have a whole team of agents working together with you.
04:50 Jonas DahlbergYes. And if you compare that to a common business process, let’s say you’re a claims handler at an insurance company, for example. Typically you open up a claim, you look at the kind of text. What case is this? Do you know who this is? Is it a long-time customer? You look for signs of fraud. Obviously these days you check whether, you know, the geolocation of the photograph, for example, matches the location that the person gave. You do a number of basic checks. Now imagine all these manual steps, and imagine that you had agents do them instead. So basically when you open the claims case, it has already checked the geolocation, the case categorization is correct, the most relevant pictures are included and the others removed, and probably a recommendation saying: we think it’s a valid claim, however we find a discrepancy, the license plate does not match the brand of the car, so it’s probably fraud. And then you as a human start working. Interesting, I wonder why. And you can really start digging into it.
06:07 Emelie FågelstedtWhich really means that you as a human can let go of a lot of those manual administrative tasks that are just checking data and comparing things, and get straight into what you are really good at. And would it be okay, to make the comparison, that generative AI and AI bots were more like assistants, and when we talk about agents, they are more like colleagues?
06:29 Jonas DahlbergDefinitely, that’s one model that works for me when I think about them. But then AI assistants these days are pretty capable too, meaning, you know, they do remember who you are. They remember what you have been telling them in the past. So it’s a blurry thing. But when you start to automate business processes using agents, that is different from an AI supporting you when you’re working.
07:00 Emelie FågelstedtShould all organizations start looking at agentic AI today?
07:05 Jonas DahlbergI think you need to build AI proficiency from the ground up. I’ve been working as a consultant for a couple of months, helping customers, you know, get started with AI. And one of the first things I did was prompt trainings for leaders. You know, C-level: who has used it? Everyone raised their hand. How many actually know how to use it? Very few hands. And you need to learn. You need to understand how it works and what the limitations are, because only then can you start thinking: okay, I’m in marketing, this is what I do, how can it help me?
07:43 Emelie FågelstedtDoes the whole AI shift in an organization really require the leadership team to have high AI competence?
07:51 Jonas DahlbergYes. It has the potential to fundamentally change the way you operate, but then you as a leader also have to be able to actually do what is necessary. You need to get the right type of people in and give them the tools they need to get the work done. You need to be willing to change what you’re doing organizationally. You know, maybe you shouldn’t have 300 of this role. Maybe you should have 150 of them and 50 of these and 50 of these and 100 of those. You need to be ready to actually change how you operate. And that starts at the top.
08:36 Emelie FågelstedtBecause AI isn’t really something that just the IT department should be looking at. It’s a leadership decision.
08:44 Jonas DahlbergYes, it is. And it starts there. It starts at the leadership level, you know, in terms of priorities. What do you want to address? What kind of business process would you like to change? And then start working with that. Then you need IT for all the basic things: security, infrastructure. There’s a lot of stuff you need.
09:04 Emelie FågelstedtOf course.
09:05 Jonas DahlbergBut I still think that’s easier than the actual implications when you start putting agentic systems into production.
09:16 Emelie FågelstedtBecause it’s not just about building a good agentic platform. This is really about rethinking what value agentic AI could bring to the whole business model.
09:26 Jonas DahlbergYes. The companies that have come somewhere have started to really look into automating business processes, not necessarily supporting processes: to fetch information, to summarize information, to relate that information to other data sources, etc. At the top end of these agentic things, you get some kind of recommendation for a decision. I still haven’t seen a fully automated process where the agent actually makes the final decision. There’s still very much a human in the loop, and I think that makes a lot of sense.
10:07 Emelie FågelstedtYeah, I was going to ask about that. Where do you think we’re heading? To a future where the agents actually make the decisions?
10:13 Jonas DahlbergYes. Yes, we will have human oversight.
10:17 Emelie FågelstedtYeah.
10:17 Jonas DahlbergAnd that will eventually evolve into humans checking the logs, maybe once a week.
10:22 Emelie FågelstedtYeah. I heard someone say that instead of human in the loop, we’ll have human over the loop.
10:27 Jonas DahlbergYes. And I think that’s exactly what’s going to happen. The thing is, when we started working with AI, we couldn’t trust these systems. You had something like 70% correctness. In some cases it was like 10% and in some cases 90. You couldn’t trust them because there were so many errors. If you look at simple tasks these days, summarize a text, read out the sentences of this type of document, check this image for signs of counterfeiting, all these kinds of things, the precision is really high. And in many ways that removes a reason to even question the result. And that is really dangerous. It means that we need to be over the loop, or at least check these systems for coherence, so that you don’t get into problems like data drift, for example, where the underlying data changes, the agentic behavior changes and you need someone to catch it.
11:40 Emelie FågelstedtYeah. And I guess that also depends on what kind of relationship you have with these agents. You were saying they’re kind of like colleagues, and then they sound super confident. They say that they’re going to do something, they say they know this is the right decision, and then it turns out that, wait, no. We all know from just talking with a normal chatbot like ChatGPT or Claude or Gemini that sometimes they do make mistakes, and that’s fine, because we’ve learned to be a bit more critical when we talk with them. But when it comes to making business-critical decisions, I guess you can’t have that room for hallucinations and confident answers that turn out to be wrong.
12:18 Jonas DahlbergWell, we do not demand 100% accuracy from humans. You don’t fact-check what I’m saying, for example. But if you asked an AI the same question, you would fact-check it. So I’m kind of questioning whether it’s reasonable to trust humans 100% and agents maybe only 96. I’m not sure.
12:45 Emelie FågelstedtYeah, that says something about relationships in the future. A lot of people are talking about actually seeing these agents as colleagues, or as headcount in teams. Before we come into that: the whole world is changing, and I think we’re all aware of the shift that’s happening. But as you were saying in the beginning, a lot of organizations haven’t really gotten that far yet. Of course, most have already started experimenting. They have AI pilots on the go, maybe agentic AI pilots. But what does it actually take to go from a pilot into production and start working with this at scale?
13:29 Jonas DahlbergThere are, of course, a number of technical things that you need in place. You need a proper data platform. You need a proper AI platform, somewhere to host it, monitor it and keep track of it, and to make sure you can get least privileged access, for example, so you don’t get agents running around nosing in the wrong kind of data sets. That is typically done with an agentic AI landing zone or something like that, where you put the guardrails, structures, policies, monitoring, etc. in place. That’s actually kind of the easy part. The difficult part is the organizational change. I think the main conclusion I’ve drawn over the last couple of months, working with a number of different customers, is that you need to address the things that are really hard.
14:23 Emelie FågelstedtWhat are the things that are really hard?
14:25 Jonas DahlbergThe really hard things, the ones that really matter. If you’re an insurance company, look at claims. If you’re a bank, look at anti-money laundering or sanctions or PEP lists. If you’re in manufacturing, start looking into your manufacturing data, your core processes, whatever that is. What type of business case do you have to automate it? And then be ready to actually change things. But we also see many customers that don’t necessarily look at cost as the key driver. For example, we are looking at one big company that works with a lot of contracting on a global basis. They sell all over the world, and they have delivery contracts on products and services. It turns out that whenever a contract is proposed, there are some 2,800 people involved all over the world, constantly reviewing these contracts. For a company that size, 3,000 people working on it is not a big issue. The issue is the time it takes. The lag from when the customer wants to buy something and we create a contract, until that contract is finally signed off, is maybe three months. Cutting it probably won’t have much effect at all on the cost side, but on the revenue side it’ll be humongous. If you can get the average lead time from contract proposal to contract approval down to a week instead of three months, that will make a big difference to their top line. This is what I mean by AI moving on from the personal assistant. Yes, we will have those, they have really good value and we’re all going to use them. But the real difference comes when you start building it specifically for your key business processes. That’s what we’re seeing. But it requires some guts.
16:43 Emelie FågelstedtYeah. Guts, knowledge and even a whole new way of thinking. This whole AI shift really requires a different type of mindset. As you’re saying, maybe it’s not just about adding a new function. It’s rethinking your processes and how you actually operate.
17:02 Jonas DahlbergYeah. And that’s a really good point, because that’s the other thing. Don’t try to take your current process and make it agentic. Think of the outcome. What is the input and what is the outcome? And then think about how to make that agentic.
17:20 Emelie FågelstedtWe see the benefits. We understand that this is happening. I think most C-suites are also saying, okay, we need to do something with agentic AI. But you were also saying in the beginning that most companies are probably not really there yet. So what is it that’s actually blocking them from taking this step fully?
17:40 Jonas DahlbergIn many cases it’s simply the business model and the way you run the company. Who is responsible for the claims process, and what’s their incentive to actually change it, for example?
17:54 Emelie FågelstedtYeah, you’re right. But as you were saying in the beginning, the technical part is probably the easy part and organizational change is the difficult part. The technical part is of course also important, because without this foundation, the infrastructure, the right data and so on, nothing will really work either. So why is it important to ensure that these AI agents get access to the right data, that it’s structured the right way, and that you have the whole infrastructure in place?
18:23 Jonas DahlbergOnce again, if we draw the parallel with people.
18:27 Emelie FågelstedtYeah.
18:27 Jonas DahlbergAt a company like Redeploy, for example, 100-plus people, a lot of consultants and managers, some leaders, some sellers, that’s basically it. And then the managed service component, of course. So there are a number of different roles doing different things. Imagine adding to that a highly capable, extremely smart person without any type of structure. There could be a lot of innovation, but it could also be a lot of disruption.
18:56 Emelie FågelstedtYep.
18:58 Jonas DahlbergAs long as you can keep track of what the agents access and you have some basic monitoring, you are fairly safe for many use cases. But when you have business decisions at the end of it, like a claim or a credit rating or whatever type of checks you do, in many cases you are obliged by law to be able to reconstruct the decision tree. Then you need to produce, basically: this was the data, this was the agent, this was the instruction, this is the outcome, and this is where it went wrong.
19:45 Emelie FågelstedtDo you think, then, that identity and security are sometimes overlooked when it comes to this area? Is it something enterprises should be thinking about more?
19:54 Jonas DahlbergFrom another perspective: if the agent has an identity similar to a person or a user, you can track what it’s doing. You can see what it accesses. You can set access-level security on different solutions. As long as the agent itself has an identity, you can use the security systems you already have, which is vitally important. As long as you stay within one ecosystem or one platform, it’s quite easy. But it becomes more difficult when you really want to deliver the user experience that the user wants.
20:39 Emelie FågelstedtWhat does it actually take to be agent-ready? What is the minimum that needs to be in place for an organization?
20:47 Jonas DahlbergControl of your data sets. Basically a good, solid data platform, and an AI platform, basically the landing zone, where you can enforce security, compliance, etc. And then you need a proper business plan for what to do.
21:03 Emelie FågelstedtAnd I guess you should start smaller, with a set use case, instead of trying to become agentic everywhere at the same time.
21:11 Jonas DahlbergExactly.
21:12 Emelie FågelstedtIf a CIO was listening to this right now, what are three questions that they should be asking themselves tomorrow to prepare for this shift?
21:20 Jonas DahlbergI think the number one question is where this can make the effect they want, both top line and bottom line. Really start talking to their colleagues in the leadership team. Imagine AI that acts on behalf of someone. If we were to deploy that, where in your business would it make the biggest difference, both in cost cutting and on the revenue side? And really, really listen to the answer. But you need to really understand it. You need to get the numbers. 80% of claims are easy. Address those.
22:00 Emelie FågelstedtIs there a single question you’d say would immediately reveal whether an organization is in fact agent-ready or not?
22:10 Jonas DahlbergNo, there’s not. I wouldn’t say it’s one question, but there are a couple of them. Do you have complete control of the agentic? Do you have an agent inventory? Do you know what agents are running? Do they have least privileged access? Can you reconstruct what they are doing? And if they stop working, what’s your plan? Many of the CIOs I’ve spoken to over the last couple of weeks, when you bring up these questions, they say no.
22:42 Emelie FågelstedtI think a lot of boards are also starting to realize that if agents are out there acting autonomously, performing tasks and maybe making decisions, there are of course a lot of risks involved. Maybe the CEOs and the teams are aware of this. Maybe they aren’t. But what do you actually answer the boards?
23:03 Jonas DahlbergI think the way you manage that is basically that you do some kind of risk taxonomy, and then you make sure that you have an agent inventory with a risk taxonomy on it. So you know that these three are working on really critical processes, and whether they make decisions themselves or not, for example. You simply create a good model to describe it, and then categorize every agent you have.
23:29 Emelie FågelstedtBut when it comes to risk assessment, what if these AI agents actually make mistakes, if they make the wrong decisions?
23:36 Jonas DahlbergIf you have an agent in marketing, for example, doing copy or writing articles, obviously the CMO is still accountable for what the agents and the AI are producing, which means that it will take some time before we get to completely agentic decisioning. I think there will be a human in the loop for a long time still.
24:06 Emelie FågelstedtAnd if these mistakes start happening, maybe they also start scaling, because other agents catch on. How many companies, or boards, actually have the ability to just press stop, to have a rollback plan and make sure that, okay, we can keep on operating as normal?
24:25 Jonas DahlbergYeah, I think very, very few have agents in production.
24:30 Emelie FågelstedtYeah.
24:31 Jonas DahlbergWhich means very few have even thought of the need to have a fallback plan for what they do. But similar to the kind of discussions you have around sovereign cloud these days, and you know, the risk between the US and Europe, etc., you obviously need similar continuity plans for your agentic infrastructure as well. What do you do? Can you put key things on-prem? Worst case, can you automate it? Can you rely on multiple vendors for delivery safety? There are a number of questions that need to be answered over time.
25:15 Emelie FågelstedtFor the organizations that are already starting to move forward, that are deploying these agents and making progress with it, what are they actually doing differently?
25:26 Jonas DahlbergWhat they are doing differently is that they are addressing real business process problems. The problem of consistency in background checks, for example. If you have humans checking data and reading through documents hour after hour after hour, at the end of the day you will see a significant degradation in performance. They will miss things. If you can get the AI to do that, then you can use what we are really good at: spotting anomalies, the gut feeling, these kinds of things that come with being human. We are still better at being human.
26:09 Emelie FågelstedtSo that’s comforting.
26:10 Jonas DahlbergYeah, it is. Still, for a while. We’ll see how long that stays true.
26:17 Emelie FågelstedtDo you think it’s realistic for an enterprise to become agent-ready in six months, or is this a multi-year process that you need to go through?
26:25 Jonas DahlbergYou know, if you get some basic things in place, you can basically say you’re already agent-ready. Then come the follow-up questions. What agents do you have? How do they perform? How do you monitor them? If you go through all the questions and can answer them well, I guess you’re agent-ready. What’s the value we’re getting out of it? What is it actually changing? Is it cost? Top line, bottom line? Customer effect? What it is actually impacting from a business perspective is what’s important.
26:58 Emelie FågelstedtSo if you are a CIO or a CEO and you have these stalled AI initiatives behind you, and some of them might not have performed the way you were hoping for, what should you actually be doing right now? Give up? Or how do you push forward and really get there?
27:14 Jonas DahlbergI think there’s a lot of talk around these AI pilots that fail. But, you know, this is a new toolset, and organizations need to learn it. You need to build the skills. You need a certain level of AI proficiency in the leadership teams, for example. And unless you did these pilots and these tests, you wouldn’t have learned what worked or not. Don’t look at it as a failure. Look at it as a learning experience. It did cost money, but I still think it’s useful. And if you now take that learning about what works, the state of your data, etc., and you start asking the right questions to the right people, what can we actually change in your part of the business, and you get the reply, you can apply these learnings from the failed pilots to doing something real and really make a difference. So I think it’s important to rethink that. Whenever you learn something, it should be considered a success.
28:15 Emelie FågelstedtAnd I guess you’re right. It’s like what you say within innovation: fail fast, but always fail forward. For someone listening to this now, what do you want them to take with them from this conversation?
28:29 Jonas DahlbergI think AI agents will make a huge difference, but you need to think early on about where to apply them, and relentlessly pursue that.
28:46 Emelie FågelstedtMakes sense. Thank you so much, Jonas, for sharing all these thoughts today.
28:50 Jonas DahlbergThank you.
Tell us what you are working on and we will connect you with the right person at Redeploy.
