expert-interviews · ai-at-work · ai-agents · workflows
AI at Work: What Has Changed? Part 3
By Mantle Chat Team · Nov 16, 2026 · 17 min read






In this article, six experts — a head of operations, a product manager, the co-founder and CTO of an AI agent company, the co-founder and CCO of an AI startup, and two software engineers — answered the same questions about how AI is changing their everyday work.
Vicky Ftochogianni, a senior software engineer at Inner Circle, was working on a feature that also touched the Android and iOS apps, while the mobile team had no capacity for it. Vicky made the changes in both apps with AI’s help, despite not knowing their code well. It worked, but the app changes ended up in one large PR, slow to review and painful for Vicky and the mobile engineers alike. It was merged after feedback, and Vicky now plans up front how to slice the work. Daniel Martínez Castillo, a software engineer with more than 12 years of experience, has hardly written any code by hand since the summer and opens around 70% more PRs than last year. Back from a few days off, Daniel realised how fast the codebase had changed and needed a way to catch up. Both now open more PRs than before. Both also look at the team around the code: for Vicky, review could become the bottleneck, and Daniel’s team, which seems to work more independently now, has to be more deliberate about staying in sync.
Sander Palm Ehrenreich, head of operations at Complir, and Hossein Karimian, a product manager and product owner, both had AI produce the drafts and kept the responsibility. In one example, from Complir’s SOC 2 Type 2 audit and its GDPR work, AI drafted policies, mapped evidence requests to the places in Complir’s systems that held the evidence, and tracked what was still open. Sander “still owned every decision and every submission, but the clerical layer, which is most of a compliance program, mostly disappeared.” At Beyond Now, Hossein generated “around 60–70 user-story drafts in two to three days,” then filled in what was missing, went through the stories with developers and took open questions back to customers.
Raphael Kalandadze, co-founder and CTO of Wandero, had coding agents work across the production of the company’s introduction film, from the script to the animation code and the rendering. “My role was to direct the work and judge the edit,” Raphael says, “while the agents handled much of the production.” J, a co-founder and CCO at an AI startup who preferred to stay anonymous, spends “a lot of hours in the car” and is “now able to actually talk to my CRM, do updates there.” More broadly, AI lets J “automate everything that is not direct contact with our prospects and clients!”
Across these answers, we see several shared themes, and we look at each of them below. Five experts also named the tools they use, and all six shared advice for someone in their field who is starting to use AI at work.
Most sections are marked What they do when they describe the experts’ own practice, or What they think when they share a view or a forecast.
What they have in common
We studied the six sets of answers side by side. Six themes came up more than once: some in what the experts do, others in what they think.
What stays with them
The judgment, every decision and submission, what is worth pursuing, the customer’s needs, the bigger picture, direct contact with prospects and clients: for each of them, something stays with them.





6 of 6Out of the chat window
AI connected to Confluence, Figma and a CMS, a CRM, ClickUp and Datadog, notes and calendar, or reading documents and updating systems: for each of them, it does work, not just answers.





6 of 6Who checks the work
More PRs, dozens of drafts in a few days, a codebase that grows faster than ever, longer agent runs: someone still has to check what comes back.



4 of 6Beyond their own role
One person who can now cover what used to be several full-time functions, an engineer working outside the usual area, a CTO making creative choices with engineering tools, an engineer thinking more about the product.



4 of 6Working together, differently
A conscious effort to stay aligned, shared skills and spec files, agents that inspect the agents a team ships, AI drafts reviewed with developers.



4 of 6There is a catch
Prospects who wonder if they could build it themselves, too much work in progress, less context that doesn’t necessarily make the whole job cheaper, one PR that took a long time to review.



4 of 6What they hand over, and what they keep
What they doVicky · Sander · Raphael · Hossein · Daniel · J
We asked each expert about a recent task with AI and about how their work has changed. Here is what went to AI, and what stayed with them.

Vicky Ftochogianni
Senior Software Engineer, Inner Circle
Android and iOS apps
Two codebases Vicky doesn’t know well
Handed to AI
Kept

Sander Palm Ehrenreich
Head of Operations, Complir
SOC 2 Type 2 audit, GDPR
The clerical layer, mostly gone
Handed to AI
Kept

Raphael Kalandadze
Co-founder & CTO, Wandero
Wandero’s introduction film
A finished film, shaped by feedback
Handed to AI
Kept

Hossein Karimian
Product Manager & Product Owner
User stories at Beyond Now
Around 60–70 drafts in two to three days
Handed to AI
Kept

Daniel Martínez Castillo
Senior Software Engineer
A feature across the whole stack
Possible in a couple of hours
Handed to AI
Kept

Anonymous expert
Co-founder & CCO, AI startup
Sales and distribution
Talking to the CRM
Handed to AI
Kept
AI speeds up the doing, but it doesn’t replace owning the outcome.
Vicky FtochogianniVicky had AI explain how the screens and data flow fit together and made the changes in line with the existing patterns: “I tested them myself rather than trusting the output blindly.” Daniel’s feature ran “from detailed Figma designs involving thousands of design tokens, to our design system, the headless CMS that manages the copy on our pages, and finally the backend layer.” Using MCP, Daniel had the agent “connect all the pieces together.” In the past, a feature of this scope “would have taken a couple of weeks to deliver.” Now it is possible “in just a couple of hours.”
Raphael and Hossein both name the responsibility that stayed the same.

What has stayed the same is the responsibility. I still need to understand the customer, choose what matters and have the important conversations. AI can investigate an idea or build something for me. I still have to decide whether it is useful, correct and worth pursuing.

What stayed the same was my responsibility for understanding the customer’s needs and making sure the stories were accurate and clear enough for developers.
Out of the chat window
What they doHossein · Daniel · J · Vicky · Raphael · Sander
All six experts describe AI that does work with their systems, not only answers in a chat window. Sander draws the line most clearly:
Not the early version that threw walls of text at you in a chat window, but the agentic version that actually does work: reads the documents, drafts the entry, updates the system.
Sander Palm EhrenreichHossein and Daniel have used MCP: at Beyond Now, Hossein “created an AI skills agent and a workflow using MCP connections to Confluence” and a product documentation hub, and Daniel uses MCP servers for Figma and Storyblok, among others, “to give my agents access to the tools and context they need.” Vicky has connected Claude Code to ClickUp, Datadog, Slack, Figma and Miro. For J, it is the CRM:

Through AI automation, I’m now able to actually talk to my CRM, do updates there, and make sure that our single source of truth is actually true. Insane if you ask me!
Raphael and Vicky also have work that runs in the background or on a schedule.

The biggest change is how much work can move forward between conversations with me. … Other jobs continue in the background, including system monitoring, research and growth work. My personal assistant connects my notes, tasks and calendar, then brings back decisions that need me.

I’ve moved from using AI only in the moment to setting up routines that run daily or weekly. They work on things we never had time for: resolving vulnerabilities, fixing the most frequent errors showing up in Datadog, preparing tasks so they’re ready for refinement, and reviewing PRs. … I still review what comes out of these routines before anything is merged.
Who checks the work
What they thinkRaphael · Vicky · Hossein · Daniel
Four experts talk about checking what AI produces. Vicky’s large PR left a lesson.

The main takeaway is that I now plan the slicing up front, especially in code I don’t own, because AI makes it easy to produce a lot of code quickly and that makes the review harder for the people who have to read it.

Another lesson is to check the work itself. A run can have a successful-looking status without producing the result I wanted. Now I ask for something I can inspect, such as working code, a report with sources or a rendered video.

My takeaway was to build review and validation into the workflow and judge the output by its usefulness for development, rather than by the number of stories generated.
Daniel, who has been “trying to lean on AI as much as possible, including for code reviews,” thinks it is important to develop “a good intuition early on” about the limits of LLMs and “when to trust their output versus when to question it.”
Some of it happens before AI starts. Vicky’s team now writes a spec file before implementing, which gives reviewers “something to check the code against.” Raphael explains the goal, gives the agent the relevant context and defines what it can do alone: “Writing a useful brief and asking the right questions have become a bigger part of my job.” Afterwards, Raphael reviews what the agent produced and uses the feedback to improve the next run.
Beyond their own role
What they thinkSander · Vicky · Raphael · Daniel
Sander’s career has mostly been in startups, where “you were always bottlenecked by your own capacity.” With agentic AI, “as one person I can now cover what used to be several full-time functions.” What hasn’t changed: “you still need to wear many hats. If anything, more hats, because more output is now expected of everyone.” Sander also didn’t expect a shift toward generalists:
Three other experts describe going beyond their usual work: working in other codebases, making creative choices, thinking more about the product.

I expected AI to be a faster way to type code that I’d have to check line by line. Over time it became reliable enough that I spend much less attention on implementation details and more on the problem itself and whether the solution is right. It also unblocked me from waiting on other teams, as in the mobile work on the events feature.

I found it useful to work on creative choices through the same tools I use for engineering. A change in pacing or the way a feature was shown could become a direct change in the animation code.

I feel like I can now let my creative side take over more often, be bolder, and explore solutions I wouldn’t have dared to try in the past. I also find myself thinking more about how to improve the product itself, which is something I’ve discovered I quite enjoy.
Working together, differently
What they doDaniel · Vicky · Raphael · Hossein
Four experts describe what changed with colleagues and clients. Daniel has “the impression that communication among team members has somewhat decreased”: the team seems to “work more independently now, which means we need to make a more conscious effort to stay aligned.”

Our codebase is growing faster than ever, and with agents increasingly involved in code reviews, it’s becoming even more important to stay aligned and establish clear standards for our agents to follow, so the codebase can continue to grow in a healthy way.
Besides this, I feel that conversations within the team are now more focused on architectural alignment and product decisions, rather than low-level implementation details.

At first, everyone on the team used AI in their own way, with their own flow, so PRs varied a lot in how they were produced and what they looked like. Recently we started sharing skills and workflows, and the process is much more coherent. … When I open a colleague’s PR, I roughly know what to expect: there’s a spec I can check the code against, and the structure is familiar.
At Wandero, agents now help the team look at its own product.

With my team, a big change is that agents help us inspect the agents we ship. They review customer sessions and system logs, then bring back cases we can investigate. …
The interpretation still takes judgement. An agent might refuse a request because it followed an existing rule correctly. Before changing anything, we need to understand what the customer wanted and why that rule exists.
With clients, Raphael’s team spends time “deciding which work the agent should own,” and for Raphael, “AI changes how the work gets done, while making clear instructions and shared expectations more important.” For Hossein, the AI-generated stories gave the team “concrete drafts to review with developers,” while “collaboration with developers and customers remained essential to making the stories ready for development.”
There is a catch
What they thinkJ · Daniel · Raphael · Vicky
Four experts describe a catch. For J, it shows up in sales conversations.

One thing to watch out for, though, is how easy it is to become overwhelmed if I don’t keep my work in progress under control. It’s tempting to multitask and have several things running in parallel, but keeping all that context in my head can become quite exhausting.

One surprise was that giving an agent less context does not necessarily make the whole job cheaper. When its history is compressed, useful details can disappear. It may need to search again, reopen files or recover information it already had. Looking only at the size of an individual model call can hide that extra work.
For Vicky, the catch is the one in the large PR: AI “makes it easy to produce a lot of code quickly,” but “a big change is still hard for a human to review.”
What they expect next
What they thinkHossein · Vicky · Raphael · Daniel · J
NoteThese are expectations, not facts, and some of them are questions.
Hossein and Vicky both look further up the process: Hossein would like more AI support early in the requirements process, and Vicky would like it to help “in alignment and design, not just implementation.”

I would like to explore how AI can support more of the early requirements process, including identifying missing information and helping prepare clarification questions before development starts.
I am particularly interested in how we can evaluate quality alongside speed: producing more drafts is useful only if they help the team reach clear, agreed requirements.

I expect my work to shift further from writing code toward defining problems, setting direction and reviewing. I’m curious how far routines can go: more of the repetitive work, like vulnerabilities, recurring errors and refinement prep, running on its own with me only reviewing the result.
Raphael looks at longer runs, Daniel at whole teams.

I’m most curious about agents that can carry a useful project forward over a longer period. … The next challenge is continuity. That means remembering why a decision was made and noticing when an assumption changes. …
What I’m still unsure about is how to make longer runs easy to trust without reading every step.

Today, we’re often integrating agents into workflows originally designed for humans. … In the short term, I expect most of our tools to become AI-native, and some of our existing processes to gradually give way to new ones. …
I believe this technology will force us to go back to the drawing board and rethink how we structure teams, distribute responsibilities, and build software.
Raphael is also “researching how several agents should work together,” and wants “to understand when the extra complexity earns its place.” Daniel sees “a lot to figure out,” but finds the transition “particularly exciting.” J doesn’t make a forecast:
If I knew I’d tell you. The world is evolving so fast that I’m not even sure if a human, let alone me, can fathom what the next change will be. ;)
Anonymous expertThe tools they named
What they doRaphael · Vicky · Daniel · Hossein · J
Five experts named their tools. Four of them use Claude, two use Codex, and two have connected AI to other tools over MCP.

Raphael Kalandadze
Co-founder & CTO, Wandero
Coding, code review, research, analysis, creative work
Assistant from the phone, scheduled jobs

Vicky Ftochogianni
Senior Software Engineer, Inner Circle
Implementation, investigation, PRs
Connected to
With the team
Recurring work

Daniel Martínez Castillo
Senior Software Engineer
Main setup
Personal projects
MCP servers
Pull request reviews
Experiments

Hossein Karimian
Product Manager & Product Owner
Research, explanations, structuring ideas, drafting
User stories at Beyond Now

Anonymous expert
Co-founder & CCO, AI startup
Tools
Where they would start
We asked the experts what they would tell someone in their field who is starting to use AI. Their advice draws on their own fields and experience.
Put it to work
“Stop chatting with AI and put it to work instead.”
Sander Palm Ehrenreich
Start small
Know what problem you want to solve, or pick something you hate doing, and try to automate that away.
Anonymous expert
Start with a real task you can judge
Explain the result you want, give the agent the relevant material and decide how you’ll check its work. Once that works, give it responsibility for a bigger piece of the job.
Raphael Kalandadze
Slice it small, write it down
Small, shippable pieces from the start, a short spec before you build, and always review and verify what it produces.
Vicky Ftochogianni
Judge the final work
Build human review in from the beginning, and measure success by the quality of the final work, not just the volume of generated content.
Hossein Karimian
Don’t juggle too much
Experiment, connect agents to your existing tools, and keep your work in progress under control from the beginning.
Daniel Martínez Castillo
Raphael and Hossein give the same first step: begin with a task you know well enough to judge the result. Raphael adds a lesson: “I wish I’d understood earlier how much the surrounding system matters: the instructions, memory, tools and feedback all shape what the model can do.” And J adds: “Once you see the light. There is no going back! ;)”
Raphael’s advice ends with this:
The goal for me is to spend more attention on direction and important choices. Be the decision maker, not the manager.
Raphael KalandadzeThank you
This is the third article in our series of expert interviews. At a time when AI is changing how we work, we think it matters to stop and reflect on what is actually changing. If you missed them, Part 1 and Part 2 bring the answers of twelve other experts, from video and documentation to HR and marketing.
A big thank you to Sander, Hossein, Raphael, J, Vicky and Daniel for their answers and for sharing their experience.
A new part of AI at Work comes out every two weeks on Shift. If you’re an expert in your field and would like to take part, or you’d like to get the next parts by email, you’ll find both there.
And you: which task would you hand over to AI first? Tell us on LinkedIn.
Thank you for reading all the way to the end.