Many companies have already tested generative AI for content creation, customer engagement, and internal productivity. The next challenge is more complex: moving from tools that assist individual tasks to AI systems that can carry work across multiple steps, channels, and teams. Once AI begins to orchestrate workflows rather than simply respond to prompts, questions around reliability, context, governance, and accountability become harder to ignore.
This shift is especially visible in marketing. Campaign production often involves multiple handoffs across briefs, copy, design, approvals, localization, media plans, and performance data. AI agents may help reduce the waiting time between these steps, but they also introduce new risks when errors travel through a connected workflow. For enterprises, speed alone is no longer enough. AI systems need to remember decisions, respect brand rules, know when to use specific tools, and understand when human review is required.
The discussion also reflects a broader change in how businesses evaluate AI platforms. Early adoption was often driven by output quality and productivity gains. As AI becomes more deeply embedded in operations, enterprises are starting to ask whether these systems can operate consistently over time, preserve context, document decisions, and remain accountable within defined boundaries.
In this TNGlobal Q&A, Ali Shaheen, Chief Technology, AI and Innovation Officer at Protaigé, discusses what it takes to make agentic AI useful in enterprise workflows. Drawing on more than 20 years of executive experience across AI-driven platforms and software ventures, including work with zenloop, WorkHub, and eKomi, Shaheen examines why memory, guardrails, workflow design, and domain-specific agents are becoming central to enterprise AI adoption.
Many companies have already adopted generative AI tools for content, customer engagement, and internal productivity. From your perspective, what changes when AI moves from assisting individual tasks to orchestrating live workflows across an enterprise?
We have built a lot of single-task agents, and the jump to orchestration is the point where it stops being a parlor trick and gets genuinely hard. A task agent answers a prompt and ends. The moment you ask one agent to run a whole workflow, it has to hold state and make decisions: what to do next, which tool to reach for, when to stop, and when to hand back to a person. That is a different class of system.
What surprised our users the most while using our system is where the gain actually comes from. It is not that the model does any single step better than a person. It is that you delete the handoffs. In a manual marketing stack, the bottleneck is almost never the work itself. It is the wait between steps. A draft gets finished, sits in a queue, gets reviewed, and goes back. Every handover is dead time and a place where context leaks out. When I let Maia run the sequence end to end, most of that overhead just disappeared, even on steps where a person could have done slightly better.
The catch, and this is the part you only see once you have started using this, is that coupling steps also couples their failure modes. A small error early in a chain does not stay put. It feeds the next step and the next, and by the end, you have something confidently wrong. The complexity of one of these systems tracks how tightly the steps are wired together, not how many steps there are. That is the whole reason we treat guardrails as core design constraints rather than as something you bolt on at the end.
A recurring concern with enterprise AI is that speed has improved faster than reliability. Where do you see the biggest reliability gaps today when AI systems are deployed into real business operations, rather than kept in pilots or isolated use cases?
Memory. It is the first thing that breaks when you move an agent out of a demo and into something that runs for days.
Every early agent we built looked great in a single session and came apart over time, because each turn started from nothing. The model would re-answer a question it had already settled, or contradict a decision it made yesterday, because it had no idea it had made one. We started calling it agent amnesia, and once you go looking, you realize it is not one bug, it is a few. The agent loses what happened earlier in the same task. It loses what it learned in previous sessions. And it loses the reasoning behind a decision, even when it still has the decision itself.
So with Protaigé, we stopped trying to cram history back into the prompt and built an actual memory layer that decides what to keep, what matters, and how a decision evolved. That is the line between an agent that responds and one that operates, and the raw capability of the underlying model is almost beside the point. A model can be brilliant for one turn and useless across a quarter of work, because it never remembers the quarter. It is also why we are convinced the value is in narrow, domain-specific agents rather than a general assistant. Memory is only worth anything if it is memory of something specific.
Agentic AI is often discussed as a way to automate end-to-end work. In practical terms, what does an autonomous workflow need before it can be trusted to act across tools, teams, and channels without creating operational risk?
Four things, and we built each one because its absence bit us.
One, hard boundaries on what the agent does on its own versus what needs a human. You decide this up front and wire it in. Oversight you add afterward does not really catch anything. It just slows people down and makes everyone feel better.
Two, control over tool use. This is the least intuitive one and the one I got wrong first. Maia can reach over hundreds of tools, and our instinct early on was to give the agent access to all of them. That makes it worse, not better. Every extra tool in front of the model is another way for it to pick wrong and another source of hallucination. The real engineering problem is selection: put the two or three tools that fit this step in front of it and hide the rest. We route per step rather than exposing the whole set.
Three, governance that rides along with every output instead of sitting in a separate review stage. For us, that is Brand DNA: a brand’s strategy, tone, visual rules, and audience encoded once and applied automatically to everything Protaigé produces, with a trail of why each choice was made. Without that, you are back to a human eyeballing every asset, which does not scale.
Four, monitoring that runs continuously. If the agent acts continuously, checking it once at launch tells you nothing. The checks have to run while the workflow runs.
What did building for existing workflows teach you about how enterprises actually adopt AI?
We’ve learned that enterprises don’t adopt AI as a blank slate. They need to integrate it into existing systems, constraints, and processes. Adoption accelerates when AI aligns with how work already flows, rather than requiring the organization to adapt to AI.
This is why the market is moving beyond copilots and task assistants toward agents that can actually take responsibility for work. This is what we’ve created with our agent Maia. She operates from within existing channels such as email, WhatsApp, and Slack, and can maintain context, reason across steps, and carry work through end to end without constant human stitching.
In theory, once a brand is onboarded to Protaigé, teams don’t need to engage with the platform. Maia can do it on their behalf from their day-to-day channels. It’s why we’ve called her the first autonomous account director. She operates like an actual teammate.
Alongside ease of adoption, in practice, reliability and control are more important for enterprise organizations than novelty. At the end of the day, it’s about trust, and this comes from systems that are predictable, controlled, and able to operate reliably within the way work already happens.
Marketing is a useful test case for agentic AI because it combines brand judgment, speed, compliance, localization, and performance data. Which parts of marketing execution are ready for greater autonomy, and which parts still need close human review?
We built Protaigé to take the execution load off marketers, not to take judgment away. The parts ready for autonomy are the high-volume, repeatable parts of marketing execution: versioning, localization, asset production, channel adaptation, performance optimization, and keeping every execution consistent with the brand. These are the parts that break first when teams try to scale them manually.
Most marketing teams still run this work across a fragmented stack, with the human marketer acting as the glue between the brief, the copy, the design, the media plan, the approvals, and the performance data. That works at a small scale, but it falls apart once you add more markets, more audiences, and more channels, especially if you’re on a deadline.
That is the obvious work to hand to an agent. Three markets x four audience segments x five channels means 60 assets. In a manual workflow, that usually means 60 separate execution briefs, 60 rounds of direction, and 60 chances for the work to drift. Maia takes one brief and produces the full set: copy written, layouts composed, hero images generated, variants animated, and each execution tuned to its audience and channel. Global can provide the master creative, and our system will understand the vibe, then match it for each variant. The result isn’t one asset stamped out 60 times. It’s 60 distinct executions from one act of direction.
So what we wouldn’t hand over is the work where the trade-off is a judgment call, for example, the big creative bets, the emotional storytelling, and the gut intuition of what will actually land with people. That is where human taste and instinct still matter. We think of it as inspiration versus perspiration. Humans should own the inspiration. Machines can carry the perspiration.
When an AI agent takes action across multiple tools, accountability becomes more complex. How should companies define responsibility when decisions are influenced by a model, executed through software, and approved or supervised by people?
Accountability in AI systems has to shift from individual outputs to system-level responsibility. In practice, that means organizations need to be explicit about a few things: who sets the decision rules and owns them, what the system is actually allowed to do autonomously, and where human approval is required as a control point.
AI shouldn’t be treated as an independent actor operating without boundaries. At the end of the day, accountability must be to the user. When organizations define these boundaries up front, it creates the conditions for trust, because the system isn’t operating unpredictably. It’s operating within rules that people understand, control, and are responsible for.
Looking at the next 12 to 24 months, how do you expect enterprise expectations of AI platforms to change? Will buyers still evaluate tools by output quality, or will reliability, orchestration, and accountability become the more important differentiators?
Over the next 12 to 24 months, we’ll see a clear shift in how enterprises evaluate AI platforms. Output quality will remain important, but it won’t be the primary differentiator. The real focus is moving toward reliability, orchestration, and accountability at a systems level.
Enterprises are starting to treat AI less as a collection of tools and more as operational infrastructure. That means they’re asking: can this system work consistently across workflows, does it know what context to carry forward, and can it make decisions about when and how to act?
Ultimately, organizations will expect AI to behave more like a reliable partner, embedded in how they operate, aligned to their domain, and accountable for outcomes, not just outputs.