Missing pieces of workflow automation

[on-hold music] I have before me the ominous task of making this presentation really interesting with a topic which is gonna sound like

a crib. What are those missing pieces for workflow automation to happen with AI? And I'm going to tell you really an enterprise story. Is it dry? Is it just gonna be about now I'm gonna find out who took my lunch sandwich?

We'll see. Um, and, you know, as I was preparing for this talk, and I realized in the schedule that this is gonna be just after lunch, I thought I should start off with a joke.

And since for all our daily needs, we go to AI tools, I tried to go to a AI tool for a joke, and they really suck. I couldn't find one decent joke.

If you can tell me a good joke about, you know, using AI for your real-world enterprise needs, I'd be happy to squeeze it in right now.

Uh, clicker doesn't work. Can someone help me? Oh, okay. Yep. So the graph looks a little different from all the graphs that I've been seeing this morning. We took this journey in our worlds, in our enterprise worlds, um, as we explored.

But before I dive into this, I think I should do an introduction. I'm Shirsha. I've come all the way from Bangalore to tell you the story that I see unfolding around me, not just in Thomson Reuters, where I work, trying to bring AI to my, um, to my, um, you know, teams and, um, different business processes, but

also the same story that I hear at meetups and, you know, um, different community events where, where I meet AI practitioners. Um, everyone started off trying to democratize the use of generative AI back in twenty twenty-three.

We have something called as Open Arena in, uh, Thomson Reuters, very similar to your, you know, um, playground where you can try different large language models. This is where it truly came home to, um, almost everybody in an enterprise to start using generative AI for their workflows.

Further along, we got onto the RAG, um, and, you know, prompt engineering world pretty quickly. We looked at, uh, automating various knowledge-driven tasks with the use of RAG, and very soon we were answering questions, um, at the enterprise level on what is the ROI.

Further along in twenty twenty-four, we started to play with tools and frameworks, and we heralded the rise of the agents. We are now here, where we are looking at automating entire workflows with the use of AI/agents, and not just one task at a time, right?

We are looking at a future where we wanna reimagine business processes because just automating a task seems redundant.

Okay, so what, what do we mean by workflow automation? And what I've got here is a very typical workflow for almost any company that's putting software out there. Customer calls you, calls your service desk, report a billing issue, an invoice issue, or a, a product feature not working as expected.

Your customer support, um, is gonna take the call. They are probably using RAG to sort of answer that question already, or they may be looking at, you know, internal tickets or connecting with, um, you know, um, their, uh, hierarchy to see if the answer can, if the answer can be given.

And if no answers are found, they then, you know, report a ticket to the IT ops teams, right? The internal IT ops teams do the level two support, and they're looking at, you know, launching investigations to support this.

Further along, if this doesn't work, you've got your engineering teams doing the L3, L4 support, and a fix is likely going to be identified with the use of various observability tools.

Scripts are launched to, you know, lo-- create that new build. Tests, regression tests, integration tests are launched, and finally, you've got either the bug fixed or the billing question answered.

Your tickets are updated, and SLA is met. Needless to say, all of you can spot so many of these tasks that can be automated with the use of agents.

But, you know, is automating each task the way we wanna do this? Is there something that can be done differently in reimagining this workflow? We are there. We're trying to reimagine this workflow.

Um, here's a slightly different take. Let's look at, um,

a workflow where content is getting created, right? It starts off with authors or content specialists perhaps identifying an alert or a trigger that's going to launch that content workflow.

You then have maybe approvals to say, "Yes, go ahead. Do your research. Find out what we want to write about this." And then you've got, you know, content getting created with research being done, and subsequently, your editors and your, um, associate editors and reviewers reviewing that content.

If it's, you know, very critical content, you, you will probably have several rounds of these reviews, and eventually, the finalized content goes to the publisher, which then, you know, the, the publishing teams then launch their own formatting, styling related workflows, and eventually the content is published.

Here too, you will realize that so many of these tasks can be done by AI, can be done by agents, and of course, humans being in the approval flow.

But here again, something seems amiss. Should we stick to the same design of the workflow, or should we be doing this a little differently?

Okay, so that's, that's where we are. We wanna be able to reimagine these workflows because it's a new world, because we have new capabilities with these technologies, um, and not just plug in capabilities into an existing business process, right?

But we're stuck. We're stuck. We are, um, missing certain parts of that reimagination.

So what are we missing? The first thing that we're missing is connectors. Um, I s-- I spoke to a few of the, um, you know, the stalls, uh, yesterday, and a common theme was how around providing a good AI agentic

solution, you always needed that layer which connected to your current IT systems, right? And, and connectors are a very, very much a missing part of reimagining these business processes.

Um, I also want to say that, you know, I come from a world where, um, you know, technology is not altogether new. We've been, you know, we've been doing AI, we've been doing NLP for several decades as Thomson Reuters.

And even, even in the different companies that, um, uh, that developers come from, in the different meetups and communities that I attend, they are also supporting IT systems of some of the, you know, different technology companies of our world.

Believe it or not, seventy-one percent of Fortune five hundred companies still use mainframe.

Sixty-eight percent of the world's IT production workloads still run on mainframe, right? And some of your major credit card, um, transactions still happen on the mainframe, which means we are that distant, right?

Like that technology spectrum, if you were to measure from the mainframe to, to an agentic workflow, how do we connect these worlds, right? And so that's one of, one of our major stumbling blocks, I believe, where, you know, how do you connect the worlds of the technology, um, stable technology stacks that exist with, um, with, with, with

the power of AI agentic workflows? The second thing is something that, um, I struggle with as I take new ideas to different stakeholders, and I see, you know, several startups whom I meet on a regular basis, them struggling as well.

It comes back to the question of ROI. It comes to the question of, you know, um, reliability, right? How will I be sure that my agent will be able to perform?

And often with stakeholders and, you know, from, from a business impact standpoint, it's a zero or one call, right? Am I gonna continue to need to have to pay manual hours, or can I consider that not needed anymore if I'm gonna pay for the agent, AI agent, right?

And so reliability becomes a big factor and a stumbling block for us. The third thing that we're finding missing as practitioners is to have visionaries who are able to reimagine this world with us.

It's, um, you know, as a practitioner, as somebody who's deeply entrenched in AI, you can only go that far in reimagining this world. You need the subject matter experts.

You need the specialists from that specific domain to sort of do this together with them to be able to, you know, reimagine your business processes. The, the fourth thing that, you know, we need, and I'm sure many of you will agree based on conversations that I've had, is we need a certain level of standardization, right?

We need to be able to say, "This is how agents will be built. This is how they are packaged. This is how they're deployed." It's too nascent yet, um, you know, in a, in a established tech ecosystem to say, um, "This-- Th-- We are gonna replace these, these bits with agents."

Data and systems. For an agent to truly, you know, um, get its full power, we need to give it access to context. The context is today distributed across different IT systems, business systems.

It is probably, um, partially located in logs. It is probably there in, um, chat messages or, you know, IT tickets and, you know, different, um, you know, d- different siloed systems, right?

Sometimes which are spread across different parts of the organization. And so bringing them together and even identifying which of these systems will have what and how do you correlate a single transaction across these systems becomes often a stumbling part of, you know, getting the AI in.

The sixth thing is one that, you know, I f-- I personally feel very strongly about, is creating a collaborative UX. Agents are going to be assistants, so what is the role of the human defining that and creating systems in which humans can support the work of the agents and vice versa is, I think, a very important part

of, um, creating, creating those workflows. And so what makes sense from a col- collaborative UX is something that, you know, I'm waiting to hear from, from any of your fresh ideas on, right?

AI governance. We s- we saw, um, in one of the talks about how different parts of... d- different aspects of security testing go down into parts of your agentic workflows.

And, and so, you know, aspects of your AI governance, which we'd established all this while around ethics and responsibility, how do you translate that into different levels of your agent architecture, right?

The next thing is control. We still want to give the human control. We wanna have certain steps which are deterministic and certain steps which the agent can control on its own, or, you know, how, how do you balance that, um, need for control, uh, between, between the agent and the human, and give the human the right, um,

you know, right time to act? And finally, what is the life cycle for the agent? All of you've, you know, spoken about that exponential growth of, um, evolution in our space.

How do we bring the capability, the latest capability into what we've already got deployed and that one that's ever-changing?

So that, that's what I had to share. Um, we are just at the start. A lot of good work from all of you. I'm waiting to bridge from the world that I'm seeing around me to the world that I come from, and so happy to have your questions and ideas, suggestions, feedback.

Thank you. [outro music]