Open Comments: S4 Ep.2 - Controls vs Speed: AI in the Enterprise with Ben Lingwood

Author: The Open Group · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Open Comments: S4 Ep.2 - Controls vs Speed: AI in the Enterprise with Ben Lingwood · Channel: The Open Group · 20:46 · Watch on YouTube

Topics: Enterprise AI applications, Software engineering with AI, AI governance and regulation, AI models

Summary

Ben Lingwood discusses how enterprises can balance the speed of AI adoption with the controls needed to manage risk. He describes continuous testing, monitoring, guardrails, secure experimentation, and choosing models that fit both the task and its costs. He also explains how Tech Catalyst builds around business problems and how AI is changing development roles and learning expectations. Looking ahead, he argues that control frameworks must evolve in step with increasingly capable AI systems.

Key points

Lingwood’s role includes developing a large technical community, tracking emerging technology, and translating it into solutions for business problems.
The control-versus-speed balance depends on an organization’s size, industry, and appetite for risk.
AI safeguards include observability, circuit breakers, operating procedures, and continuously evolving guardrails.
Good control begins during design and development, including model selection, cost awareness, security, bias, and output testing.
AI systems require ongoing testing because they can produce different answers to different prompts and can face malicious attacks.
Tech Catalyst experiments in isolated, secure labs before introducing technology into production environments.
Production systems need monitoring and operational responses, such as human review, circuit breakers, or switching models.
AI, cloud, and DevOps are converging, enabling rapid prototyping and new uses involving drones and robotics.
Developers increasingly need a mix of technical, commercial, and human skills, alongside continual learning.
Lingwood expects AI capabilities and their control frameworks to advance in lockstep, with monitoring agents watching AI agents.

Chapters

Start	Chapter	Summary
00:00	Lingwood’s role and the control-versus-speed challenge	Lingwood outlines his responsibilities at PWC and explains why adopting AI requires balancing agility against the consequences of moving too quickly.
04:08	What good AI control involves	He describes controls spanning model and cost selection, development-time testing, security, data quality, and continuous monitoring.
07:15	Tech Catalyst and safe experimentation	Lingwood explains how Tech Catalyst is organized around business needs and uses secure labs to test innovations before production.
11:27	AI, cloud, DevOps, and emerging technologies	He discusses changes in software development and highlights multimodal AI, drones, robotics, and the potential convergence of quantum processing and AI.
15:04	Builder culture and lifelong learning	Lingwood describes evolving developer roles, the value of broad and specialist skills, and the need to make learning a daily practice.
19:12	The future balance of control and speed	He predicts that control systems will need to evolve alongside AI, with monitoring and observability keeping pace with new capabilities.

Quotes

small decisions can create large irreversible consequences. — 02:36

testing is something that is you do as part of BAU. — 06:44

The blast radius is not going to touch anything that would would affect businesses. — 09:53

So it’s a daily activity. — 17:39

we equally need the same control framework to sit alongside that. — 20:14

Concepts and entities

Ben Lingwood, Ash, PWC, Tech Catalyst, The Open Group, observabilidade, circuit breaker, forward deploy engineer, T-shaped people, human-in-the-loop

Full transcript (original language)

00:00 Hello and welcome back to open comments with me Ash. Today we are joined by Ben Lingwood who [music] spoke at the Open Group EA practitioners conference back in February on control versus speed [music] AI in the enterprise. Ben is the co-lead partner and CTO for PWC’s UK technology business [music] unit within tech catalyst. Ben focuses on client-f facing emerging technology. His background is diverse, spanning lead technical [music] positions in FTSC 100 organizations, head of cloud engineering [music] at SAP. He has also ran a

00:31 cloudnative startup focused on transforming and operating large organizations into [music] cloud technologies. Thank you, Ben, for joining us today on Open Comments. >> My pleasure Ash, good to be here. Thank you. >> To start, what does your role as co-lead partner and CTO of Tech Catalyst [music] look like dayto-day? >> Thanks Ash. It’s pretty diverse role actually. Um let’s let’s start going through the list. So number one, it’s a large team. We’ve got about a thousand technicians in the team. Um so part of my job is making sure that those people [music] have an engaging interesting

01:02 job. They’re working with the new technology. Uh they’ve got career opportunities. So tech career paths. Uh they’re learning new certifications, new skills, keeping up to date and [music] you know aspiring to to work in PWC in the future. So that’s firstly the most important part of my job is is building a a community for builders and people to flourish. Uh most important thing for me. Second half of my job is as you said earlier sort of looking at emerging technologies. So we work very closely with the hyperscalers with the language model providers understanding future

01:33 trends what’s coming out what’s emerging. Uh we get access to some of the early early release models and updates as well just to assess where they can be used. Uh and I think the then the third element of my job is really a translation one is taking everything that we’ve learned from technology everything that we understand and taking that into into industry right so hearing about business problems and business issues whether that’s transformation modernizing business process looking at where for example AI or other technologies can be used to optimize the way businesses run and

02:05 operate and then bringing the technology to the problem never the other way round so I’d say [music] they’re my my three main parts of the day job. >> Thank you. And at the open group EA [music] practitioners conference, you spoke about controls versus speed. What is the core tension there? >> Ah, it’s always always a balancing act, right? And I think especially if you if you think of PWC, right? Um, we’re we’re a regulated auditor. We work with tax. We work in in very very sort of secure business areas, a lot of controls,

02:36 liabilities to sit over the top. Um, so that is not the friend to taking risks and fast agility. So the control versus speed is always the risk of move too slowly and the business and the competition roots around you or move too quickly and small decisions can create large irreversible consequences. So it’s always this this balancing act between the two. So when we look at uh adopting new technology for me we’re always looking at how do we safely assess review new technology that comes in. So

03:07 we’re running that in labs. We’re doing previews. We’re doing extensive testing. Making sure that then we’ve got the guard rails and operating procedures around those new technologies. So in the world of AI, looking at observability. So whether that’s technical observability or bias or building in circuit breakers into the AI responses. So if it starts doing something a little dodgy, a little off script, um it’s got some trips that can go in and we can revert back to standard scripted responses or FAQs. That’s actually a live example of one we did to things like a a complaints analysis tool.

03:37 [music] Uh and then all the way through to looking at the roadmap of those solutions as well. As I’m sure everyone knows, you know, new models are coming out what feels like weekly at the moment. Uh new models introduce new risks. So those those guardrails continuously need to be evolved and modeled as well. So observability is a key part of what we do as well with technical solutions. So always that balancing act and I would just sort of sneak in. It’s a really personal area as well. Um obviously a small startup can take far more risks [music] than a large corporate and it depends which industry you’re in as well. So you know if it was

04:08 a a large regulated bank where [music] decisions can have large impact again your appetite for risk could be very different from maybe a you know a small software startup house. So very personal question when you look at control versus speed and something everyone has to form their own risk profile on >> that’s great and what does good control look like in an AI enabled organization according to you >> good control um I think it’s across the board right so for me good control starts the minute someone opens their

04:39 laptop lid right so selecting the right language model and I think everyone’s seen the headlines recently with the the latest changes on token price thing that we’ve seen kind of across the industry. Uh it’s gone from monopoly money to very real money very quickly, right? Um so selecting the right model to do the right job that’s not only got a cost impact that of course impacts the the risk to the company commercially and I’m sure you’ve all seen horror stories with AI costs spiraling, people selecting flagship models to do basic tasks or selecting the wrong model. Uh and we

05:09 know from our own testing different models are good at different things. So selecting the right model for the right job that you need to do is dead important. Then as the developer is creating the solution, having one eye on the economics of that uh and sort of things like security testing, bias testing, output testing is not something that you traditionally do now like it used to in in sort of the software world at the the end of a development cycle. It’s [music] built actually into how you code and develop the solution. So it’s actually part of the process. It starts way at the beginning. Making sure then you’ve got continual testing frameworks

05:41 as well because unlike traditional software where you would get to a testing cycle and it spits out an output and you know that’s the output it gives and it passes or fails. There’s multiple different ways of asking AI solution a question. So they can give multiple different answers. So how do you test for that? So guardrails are very important that you have as I mentioned earlier the continuous monitoring the circuit breakers in there if you start detected there’s something rogue. Um, if you’ve got potentially malicious attacks going on there as well, AI provides a a fantastically broad uh attack plane for

06:13 for hackers. And again, I’m sure you’ve seen all the headlines on that as well, from poison prompts to Trojan horse prompts and all the other other fun things that uh can be done with an AI model from a security perspective. So that’s the other element. And then even going more broadly as well, and we’ve seen from our own test and our own solutions, AI models are only as good as the data that they ingress, right? So making sure that you’ve got a trusted clean data source as as that goes in and then making sure that your uh your AI solution or your agents are not then [music] providing biased outputs as they

06:44 go up. So long-waited answer but I would say look key things will be a it’s continuous b it starts at design time and c it keeps changing. [music] So testing is something that is you do as part of BAU. Every single prompt needs to be reviewed as it comes out going to an end user. Now >> Mhm. And moving on to Tech Catalyst. Tech Catalyst is described as a home of technical innovation at PWC. What makes it different from a traditional enterprise tech function? >> I think because we started this just

07:15 over a year ago. It’s just had its first birthday actually. Um it was kind of kind of fortuitous timing. um wasn’t I’d love to it was planned but it it really wasn’t being open. Um this was being formed as the AI sort of upward hype curve was starting. So we had an opportunity to build an AI first organization which was great. So I didn’t have lots of legacy developers and you know offshore development centers uh and legacy software developers to retrain. I could build this as an AI enabled innovation team

07:47 from from the root. So that that was a that was a bit of luck to be honest but that does make it a little bit different that it is purposebuilt and designed and operated for that reason. So I’d say that’s the first one. The second one um is I think that the team focuses around business issues. So sort of the the team never goes to market directly. We never go and position or present or or meet customers unless we’re going through one of our lines of service. So we go through our consulting team, our tax, our audit, uh our deals teams. So our

08:19 conversations always start after one of our line of service uh enterprise architects has gone in, subject matter experts has gone in and is having a conversation around a real problem, a real painoint, a real opportunity. So everything starts around a business painoint or an optimization versus going in showing lots of shiny AI tech uh trying to find a business solution uh that it [music] could potentially be glued into. So for me, tech is designed to do this the right way around. We’re bringing technology into business problems rather than going the other way

08:51 round. So that’s something that’s a little bit different, I think, in the industry when you look at where maybe gsis have come from or maybe sort of niche players who had that that heritage that they needed to retrain, readapt, re-engineer, reorganize. Uh we didn’t have that. Um so that that makes what we do a little bit different and also means the speed at which the team can adapt, evolve, move and learn uh is in the right place to be AI compatible. >> And how do you balance experimentation

09:22 with the need for production grade reliability? >> Uh the honest answer is with great caution. Um so it’s important that we have speed to and access to innovation. As I said earlier, we get we get early access to a lot of models and new technology. So, we’ve got the ability to test that. So, what what we’ve created um and then everyone’s sort of familiar with the full deployed engineering team concept. So, that’s something we’ve also deployed in PWC. So, I have safe labs. So, we can bring in new technology, new

09:53 solution into secure labs and test that out with real life scenarios uh as soon [music] as soon as it’s released and quite often we do it before it’s formally released. So that allows us to experiment safely because it’s in a clean lab room. If that explodes and does something nasty, doesn’t matter. Uh it’s in a contained area. The blast radius is not going to touch anything that would would affect businesses. It’s not connected to any sensitive data. So that that allows us to do the onboarding and get up to speed far quicker and obviously give give feedback then to either model providers or hyperscalers

10:24 or whoever we’re working with [music] and let’s take that technology through a rigorous testing cycle before we start applying it in into customers. So I think having that that biodal approach so the experimental phase learning things getting feedback failing fast [music] doing that in a secure isolated area before that’s then applied in any production grade solution is kind of a the secret that we’ve had to crack to make the team effective and then importantly exactly to your question

10:54 there once that solution is live and maybe say we’ve got a contract on that with an SLA and we’ve got to be sure that the the outputs are are production grade We’ve got our observability layer on the top of that then watching the outputs and learning. Uh and that also includes sort of rapid responses. So if we detect something has gone wrong, maybe there’s a road answer or something’s going wrong with the generation. Um then we’re making sure we’ve got the right operations. So whether that’s human in the loop, whether that’s circuit breakers [music] or or whether that’s uh adapting the tool to maybe use a a different model and doing rapid retraining. So long

11:27 answer, but hopefully that explains the the evolution of how we approach new innovation. Thank [music] you. And switching gears to AI, cloud, and emerging tech. With AI, cloud platforms, and DevOps all converging, what excites you most about all three right now? >> Oh, criy. Um, you’re absolutely right. They are converging at a a frightening speed. I mean, I would say in the last two years, the world I grew up, which is sort of traditional software development, has radically spun on its head. you know, the way we can develop

11:58 now with smaller teams, the way we can do rapid prototyping often in the same day. Um, the world of sort of classical DevOps has sort of had to had to level up into that space, especially with modern AI enabled development environments, idees, uh, really coming of age now. And I said earlier, having those security monitoring agents happening at the same time that we’re we’re cutting code and agents watching agents, helping agents is is a is a thing now, right? So, that that’s radical. One of the areas that sort of really excites me as we look forward and this

12:30 is probably a little bit more horizon one horizon 2 thing. So Horizon one is looking at the convergence of those technologies. So we’ve seen that with multimodal with AI, we’ve seen voice, we’ve seen video that that’s now blending its way into industry. Um even within my team, we have a drones team, right? Uh we’ve seen that convergence with starting to come into sort of the world of surveillance and robotics coming in. So, for example, the the drones team doing everything from looking at building site progression to health and safety monitoring. Um, all sorts of things in the world of commercial drones. We’ve seen that we’ve

13:01 seen that come to life in the last few [music] months. Uh, that’s with visual AI. Uh, we’re seeing now the move into robotics. So, if you’re now mixing language models with robotics and you’re able to uh literally give an AI arms and legs to start performing tasks, um, [music] that is incredibly exciting. We’re seeing that move at pace. And then I’d say Horizon 2 as we’re all looking forward uh we’ve seen great advancements in the world of of quantum uh and I think you know as we all sort of hold our breath and wonder where wonder when the point is going to happen where we’ve

13:31 got general intelligence uh takes place I think quantum is going to be one of those key points where the ability then for AI to perform really complex tasks um you know deep research areas to the point we couldn’t think about today whether that’s protein synthesis or genomics or aerodynamics next thermodynamics, you know, huge computational errors. Um, quantum processing mixed with AI reasoning models is going to add a a whole new level of breakthrough from that perspective. So, I’m I’m really looking forward to see how that one progresses.

14:02 It’s kind of equal measures of uh petrified and excited at the same time because obviously the the guard rails and the safeguards that we’re going to need around that level of intelligence is something I don’t think we’ve quite cracked yet but need to. >> Do you think with that as well it goes through a lot of testing? like there’s a lot of trial and error involved in that as well. >> Yeah. So, I think we’ll see increases in the amount of testing. We’re obviously going to have to do that. But I think back to my earlier point, testing isn’t something that you will do as a as a wave or a phase. It’s something that is

14:33 a BA process. It is always running. I think we will have control agents watching live agents, watching reasoning agents. So I think there’ll be a command and command and control team going through where outputs are being monitored and assessed live. So we we do that today for example with AI solutions in some of our banking insurance solutions before the prompt is echoed back to the user screen or the output is echoed back to the user screen. It’s already gone through a guardrail where it’s checked [music] what was the citation for that source. Where was the document that made that decision? Is

15:04 that trusted? Is the is the output correct? Is it has it detected any biases? that is happening today. I just see that approach scaling more and more as the intelligence gets higher. >> And now to pivot to innovation, culture and builder communities. [music] How do you foster a culture of builders within a large organization like PWC? [music] >> Yeah. And I look I I think as I said earlier, we’re seeing the the the roles shifting and changing. So where where

15:35 traditionally organizations have had to have sort of teams of software developers and sprints of those going through uh that’s changing methodology, architecture and design hugely. So what what I’m seeing now is the the emergence of the the forward deploy engineer and and the product engineer really coming to life. Um so the ability as I said earlier to do [music] the rapid testing and rapid development live with the customer. So you rather than using a whiteboard now, we’re often going in and just creating the solution and noodling that through with with a business owner

16:05 in in in the particular organization that we’re working with live. So that that’s radically changed the role of the profile that we need. So we we don’t so much need people squirreled away coding for days and days on end. We now need people that have got sort of the the trinity of development skills and able to drive IDE AI enabled tools and know how to use those correctly. We need people that have the commercial skills to be able to design that correctly. So back to language models, back to platform, back to the the cost for that.

16:36 So when you’re using I models, these are expensive things, right? So that makes a huge difference to actually is it cheaper to do with AI or am I just adding more cost over the top? So there’s always a business case eye on that. And then the third one is >> [music] >> um the the human skills, right? You’ve got to be able to to talk through business units and noodle those things and ask the task questions. And quite often because this is moving so fast help educate people that you’re working with at the same time they’re trying to create a solution. So we’re really seeing the role of the the developer now moving much more into needing those uh

17:08 reskills. So T-shaped people I think would be my my summary answer. People with a a broad tea. >> Mhm. And moving on to lifelong learning. What does lifelong learning look like [music] for yourself? >> Oh near and dear topic actually. I’m sitting some exams at the moment. So I I think we’ve we’ve we’ve seen a big change in this over the last couple of years as well. You know, we used to sit in exam and and do something. The speed things are moving now. Um I’d say there’s not and I’m speaking personally on this, right? There’s not a day that

17:39 goes by where I don’t end up spending some time for reading. I actually block a bit of my calendar every morning and every evening to to sit down and catch up on things and read. [music] So that’s that’s really important for me. So learning isn’t something you just do once and move on or sit a certification that lasts 48 months to move on because as we all know that’s typically redundant within six months the way the pace is moving. So it’s a daily activity. So people need to certainly in our industry now need to move into that is something that you just have to do every single day. It’s also too broad

18:10 that you can’t know it all. So back to my T-shaped thing, I think you need to be a really good generalist. So whether I’m talking cloud platform or compute or you know scripting models or AI development or language model selection, I need to know a little bit about everything to a reasonable depth. But then I need my specialtity underneath. Right? So it’s the deep area that maybe you’ve specialized with historical or you’ve come into or there’s particular interest for you. And if I look at my own team, we’ve got people that specialize in UI through to robotics through to red teaming on language

18:42 models through to obviously developers um through to sort of specialist managers that that run particular sprint cycles on on development cycles. So you need your specialtity. So I think that’s where we need specialist teams curling together to to [music] start forming these solutions as well. And maybe what strange answer on this as well is not all of the people on the projects now are human. You know, we’re we’re often now working with agentic colleagues side by side who do some of these jobs as well. So we all have to learn from each

19:12 other as well as uh building learning into every part of our day job. >> Thank you. And as we wrap up this episode, looking ahead 3 to 5 years time, how do you see the balance between [music] control and speed evolving? Oh, that’s a real crystal ball question, Ash. That’s a that’s a tough one. I I think they have to go in lock step is the answer. So, the more intelligence that is built into reasoning models, uh the more these models evolving. So, whether that’s a large language model or

19:43 increasingly now focusing on smaller language models that are deep deep specialists in one areas. And we do like a a command and control and swarm models coming in now where there’s multiple specialist agents swarming around um a particular business problem or a prompt. We we’ve seen that with some of the the the open claw claw trail technologies coming of age now co-work coming in on on desktops now with the Microsoft or the anthropic desktop co-work solutions. So we’re seeing this emergence of all this technologies swarming together. We equally need the same control framework

20:14 to sit alongside that. So I would say in log spec. So wherever we’re identifying something, we’re driving an AI co-worker to to do a particular task or building an AI solution, there is equally a a sort of a an antimatter version, a a control control plane that has an agent that is watching and evolving at the same level. So we’re trying to now within the company wherever we develop a new solution, we’re also having to upscale our observability layer to make sure that those two are moving at equal speeds. Thank you, Ben, for joining us

20:46 today on Open Comments. It’s been really great learning more about you and also about your presentation at the Open Group EA Practitioners Conference back in February on controls versus speed, [music] AI in the enterprise, and also AI at large. We hope our listeners enjoyed this episode as much as we have, and we look forward to bringing more topics into the fold just like this one to you very soon. Thank you, Ben, and thank you to our listeners. Stay safe and happy listening.