Mimica Chief Says Missing Workflow Data Threatens Enterprise AI
Ai4
Gartner predicts nearly half of agentic AI initiatives will be canceled by 2027 because organizations lack operational visibility into actual day-to-day employee workflows.
## Background
This presentation was delivered by Tuhin Chakraborty, Co-Founder and CEO of Mimica, and published on YouTube by the channel Ai4 on August 19, 2026. The page description outlines how organizations face pressure to deploy AI agents quickly, but risk misplaced investment without operational visibility into business processes—a problem solved by process intelligence.
## TL;DR
Tuhin Chakraborty argues that the primary hurdle in deploying agentic AI is not writing code, but gaining full visibility into how enterprise work actually happens. While traditional employee interviews miss crucial edge cases and process variations, desktop process capture technology automatically maps workflows to identify high-ROI tasks. He outlines a five-phase deployment lifecycle and traces the evolution of agents from deterministic V1 systems to full end-to-end autonomous processes. Ultimately, human workers will transition from performing manual tasks to acting as supervisors who manage multiple AI agents.
## Atmosphere & style
The presentation carries an authoritative, pragmatic, and forward-looking executive tone. Chakraborty balances market warnings—such as high project cancellation rates—with a structured, solution-oriented roadmap for enterprise AI adoption. The style is clear, strategic, and polished, typical of an industry keynote aimed at demystifying complex operational transitions.
## Key takeaways
- [00:08](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=8s) The primary barrier to scaling enterprise AI agents is not the capability of underlying models, but a lack of visibility into how work is actually executed day-to-day.
- [00:08](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=8s) Gartner predicts that approximately 50% of agentic AI initiatives will be abandoned or canceled by 2027 due to operational challenges.
- [02:45](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=165s) Modern AI coding tools have reduced the cost of writing code to near zero, shifting the main bottleneck to 'knowing what to write' regarding process rules and variations.
- [02:45](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=165s) Traditional employee interviews fail to capture true workflows because workers naturally describe ideal paths while omitting edge cases, exceptions, and regional variations.
- [05:27](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=327s) Desktop sensor technology captures OS-level click and keystroke data with built-in employee privacy controls to map processes accurately without relying on human memory.
- [05:27](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=327s) Successful agent deployment relies on a 5-phase lifecycle: Selection, Mapping, Genesis, Harden/Scope Expansion, and Maturity.
- [08:22](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=502s) Process intelligence tools like Mimica Miner evaluate tasks across the enterprise based on volume, variability, and suitability for deterministic or reasoning-based automation.
- [08:22](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=502s) When addressing tribal knowledge or process ambiguities, automated systems prompt Subject Matter Experts with specific screenshot evidence rather than broad questions.
- [08:22](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=502s) Captured process recordings serve as exact blueprints to auto-generate baseline agent workflows across APIs, GUI automations, and unstructured data tools.
- [12:10](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=730s) AI agents will evolve through three distinct generations: Deterministic/V1 Agents, Planning Agents for non-linear workflows, and fully autonomous End-to-End Agents.
- [12:10](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=730s) Human employees will transition from manual task execution into supervisory roles, monitoring agent output quality and handling rare exceptions to amplify individual productivity.
## Detailed walkthrough
### [00:08](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=8s) The Great Transition and the Disconnect in Agentic AI
Tuhin Chakraborty, Co-Founder and CEO of Mimica, opens his presentation by introducing the core bottleneck facing organizations today: while businesses are rushing to adopt AI agents at scale, the single biggest hurdle is actually understanding how work is truly executed within an enterprise. Framing this shift as "the great transition," Chakraborty outlines a bold projection that within ten years, the vast majority of manual, digital work across the world will be handled by autonomous agents. With LLMs unlocking computer-based tasks, automation is inevitable.
However, there is a stark disconnect between this high-minded promise and the reality on the ground. Despite the inspirational feeling leaders experience when testing modern AI models, Gartner predicts that nearly half of agentic AI initiatives will be canceled by 2027.
- **Key Takeaway:** The ultimate challenge in enterprise AI is not the capability of the underlying models, but the lack of operational visibility into actual day-to-day workflows.
- **Market Forecast:** Gartner projects that ~50% of agentic AI projects will be abandoned or canceled by 2027.
---
[02:45](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=165s) The True Cost of Building Agents: "Knowing What to Write"
To explain why so many agentic projects stall, Chakraborty breaks down the cost of building an AI agent into two distinct components:
1. **Writing the code** for the agent.
2. **Knowing what to write** (understanding the process rules, variations, and context).
Thanks to cutting-edge AI coding tools like Claude Code and complex multi-agent architectures that plan, debug, and test code automatically, the cost of writing code has essentially collapsed to zero. Consequently, the remaining bottleneck is "knowing what to write." Enterprise practitioners frequently struggle because organizations lack end-to-end visibility into their own processes.
When employees are interviewed about how they perform their duties, they naturally describe the ideal "happy path" and inadvertently gloss over critical exceptions, edge cases, and localized variations. Furthermore, global operations add massive complexity—teams in Germany, the US, and Brazil often execute the exact same core process using vastly different steps and hundreds of subtle variants.
- **The Core Paradox:** "You can't automate what you can't see."
- **Why Interviews Fail:** Human workers omit edge cases and rare exceptions during traditional workflow discovery interviews.
---
[05:27](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=327s) Process Capture Technology and the 5 Phases of Agentic Deployment
To overcome the blind spots of human reporting, Mimica utilizes desktop sensor technology to record precise click-and-keystroke data at the OS level (e.g., within Windows). This sensor captures low-level interactions—such as field entries and button clicks—and anonymizes the data to predict automation potential. Crucially, the tool puts full control in the employee's hands with clear start, stop, and pause controls on-screen to protect privacy.
Chakraborty outlines the five distinct phases required to successfully roll out AI agents across an enterprise:
1. **Selection:** Identifying the most ripe candidates for automation out of thousands of employees and tasks.
2. **Mapping:** Documenting every micro-step, variance, and exception across teams.
3. **Genesis:** Generating the baseline agent code or automation logic directly from captured data.
4. **Harden / Scope Expansion:** Operating the agent side-by-side with human workers; the agent manages happy-path items while passing exceptions to humans, gradually expanding its scope over time.
5. **Maturity:** The agent takes over full execution while a human transitions into a supervisory role.
- **Privacy & Control:** Employees maintain complete authority over recording using an on-screen sensor widget with explicit start, stop, and pause functions.
---
[08:22](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=502s) How Process Intelligence Powers Selection, Mapping, and Genesis
Chakraborty details how process intelligence software transforms the initial stages of agent deployment. Tools like Mimica's *Miner* are deployed across hundreds or thousands of desktop environments to automatically surface high-volume, standardized tasks that present the strongest automation ROI. It evaluates whether a workflow requires complex reasoning or is a deterministic sequence suitable for an LLM wrapper.
During mapping, process intelligence captures exact decision thresholds, business rules, and micro-actions. If ambiguous steps occur or tribal knowledge is required, the software automatically prompts Subject Matter Experts (SMEs) with targeted questions paired with actual screenshots of their historical actions, rather than subjecting them to blank-slate whiteboard sessions.
In the genesis phase, these detailed process maps directly auto-generate agent workflows. Whether converting 15 desktop clicks into a direct backend API call, setting up GUI automation, or parsing data across unstructured PDFs, SAP, and Excel spreadsheets, the recording serves as the exact template for the agent.
- **Mimica Miner:** Scans enterprise operations at scale to categorize tasks based on volume, variability, and suitability for deterministic vs. reasoning-based automation.
- **Human-Driven Prompts:** Uses concrete screenshot evidence from historical user sessions to ask SMEs specific questions about edge-case logic.
- **Core Philosophy:** "The best agents are powered by people" — top-performing employees supply the exact behavioral blueprints that agents mimic.
---
[12:10](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=730s) The Evolution of Agents and the Human Supervisor Role
Chakraborty outlines the roadmap for how agentic capabilities will evolve across three major generations:
1. **Deterministic / V1 Agents:** These tackle repetitive, workflow-based tasks that legacy RPA failed to solve due to slow development cycles or minor reasoning gaps. They combine GUI and API automation with targeted LLM calls for unstructured tasks like reading invoices or drafting emails.
2. **Planning Agents:** Designed to handle non-linear workflows with variable user interactions (e.g., insurance claims where a customer postpones a phone call). These agents reason dynamically about the "next best action."
3. **End-to-End Agents:** Autonomous systems capable of handling multi-step enterprise workflows—such as complete Know Your Customer (KYC) onboarding or end-to-end claims processing—from start to finish.
In closing, Chakraborty addresses the future role of human workers in an agentic world. Rather than executing manual tasks, humans will step into **supervisory roles**, reviewing output quality, handling rare exceptions, and monitoring performance. This shift allows an individual's productivity to scale exponentially—from doing the work of one person to managing the output of five to ten agents. Success will ultimately belong to organizations that cultivate a corporate culture incentivizing employees to embrace this transition.
## Notable quotes
> [00:08](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=8s) the great transition — Tuhin Chakraborty
> [02:45](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=165s) You can't automate what you can't see. — Tuhin Chakraborty
> [02:45](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=165s) Knowing what to write — Tuhin Chakraborty
> [08:22](https://www.youtube.com/watch?v=z0ARu4kvWz4&t=502s) The best agents are powered by people — Tuhin Chakraborty
## Video description
Tuhin Chakraborty - Co-Founder & CEO, Mimica
Organizations are under immense pressure to move quickly on AI. Leaders are told to invest aggressively, experiment broadly, and deploy agents across the enterprise. But without a clear understanding of their business processes, that urgency can lead to overbuilding, misplaced investment, and AI initiatives that fail to deliver meaningful business value.
Successful AI agent deployments begin with understanding how work is actually done, as agents can only be as effective as the operational context they are given. AI-powered process intelligence provides that visibility by mapping how work is executed across teams. It enables organizations to identify where agents can have the greatest impact based on evidence rather than assumptions.
In this session, Mimica Co-Founder and CEO, Tuhin Chakraborty will explore how enterprises can build a stronger foundation for agentic AI – identifying the most valuable opportunities, reducing wasted investment, and scaling initiatives that deliver measurable operational impact.
Attendees will learn how to:
-Identify the processes and workflows best suited for AI agents and automation
-Use process intelligence to understand how work actually happens across teams, applications, and systems
-Avoid overbuilding and misplaced investment by prioritizing opportunities based on operational evidence
Tuhin Chakraborty, Co-Founder and CEO of Mimica, argues that the primary hurdle to enterprise agentic AI is not model capability, but organizations' fundamental lack of operational visibility into real-world day-to-day workflows.
The Workflow Visibility Deficit and Coding Shift (2:45)
Chakraborty projects that within ten years, autonomous agents will handle the vast majority of manual digital work. However, enterprise implementations face significant operational headwinds; market research firm Gartner forecasts that nearly half of agentic AI initiatives will be abandoned or canceled by 2027. Chakraborty attributes this high failure rate to an operational blind spot rather than software development limits. With the advent of multi-agent architectures and AI coding tools like Claude Code, the technical cost of writing code has effectively collapsed to zero. Consequently, the central bottleneck has shifted entirely to understanding process context, rules, and variations.
Traditional methods of uncovering process requirements consistently fall short. When interviewed about their daily responsibilities, human workers reliably recount an idealized 'happy path,' unintentionally omitting crucial edge cases, exceptions, and institutional workarounds. This discovery problem is compounded across multinational organizations, where teams in regions such as the United States, Germany, and Brazil execute the same core business process using hundreds of localized step variations.
Five-Phase Deployment via Desktop Capture Technology (5:27)
To capture unvarnished workflow data, Chakraborty advocates deploying desktop sensor technology directly within operating systems to record granular interactions, such as keystrokes and field clicks. To address privacy concerns, the capture utility incorporates on-screen controls that allow employees to start, pause, and stop recording at will, while scrubbing sensitive data prior to analysis.
From this captured behavioral data, Chakraborty structures agent implementation into five formal phases: selection, mapping, genesis, hardening, and maturity. The organization first identifies high-opportunity candidate tasks from thousands of potential workflows. After micro-steps and variances are mapped, baseline agent logic is generated. During hardening, the agent works alongside human staff, executing standard transactions while routing edge cases back to employees. Over time, the agent expands its operating scope until reaching operational maturity, where human employees step out of execution entirely and transition into supervisory roles.
Translating Process Intelligence into Automation Blueprints (8:22)
The front end of deployment relies on process intelligence tools such as Mimica Miner to monitor desktop activity across hundreds of machines, categorizing tasks by volume, variance, and suitability for deterministic automation versus complex reasoning models. Instead of relying on abstract whiteboard sessions to resolve workflow ambiguities, the software automatically surfaces screenshots of historical employee actions to ask subject-matter experts targeted questions regarding specific decision thresholds and edge cases.
These detailed empirical records function as exact templates during the genesis phase. The system translates observed user steps into executable software flows, converting multi-click desktop routines into backend API calls, configuring graphical user interface automations, or deploying LLMs to process unstructured content across spreadsheets, PDFs, and enterprise resource planning software like SAP.
Three Generations of Agents and Human Supervisors (12:10)
Chakraborty outlines a three-stage evolutionary roadmap for enterprise agents. The initial generation consists of deterministic V1 agents that resolve the maintenance and development delays that hindered legacy robotic process automation by combining GUI and API automations with targeted LLM calls for tasks like reading invoices. The second generation introduces planning agents that use dynamic reasoning to determine next best actions in non-linear workflows with variable customer interactions, such as delayed insurance claims. The third generation encompasses fully autonomous, end-to-end agents capable of completing complex, multi-step enterprise workflows like Know Your Customer compliance from beginning to end.
In this framework, the future enterprise workforce shifts from executing manual digital tasks to managing agent outputs. Chakraborty projects that human workers will oversee five to ten agents simultaneously, focusing on exception handling, quality reviews, and performance monitoring.
Chakraborty concludes that organizational success during this transition depends on fostering a workplace culture that actively incentivizes workers to provide behavioral blueprints for automation, though enterprise viability ultimately relies on whether human supervisors can reliably manage the exception loads of evolving multi-agent systems.
[0:08] Tuhin Chakraborty: the great transition
[2:45] Tuhin Chakraborty: You can't automate what you can't see.
[2:45] Tuhin Chakraborty: Knowing what to write
[8:22] Tuhin Chakraborty: The best agents are powered by people
Original transcript
[0:08] Hello everybody. My name's Toucan. I'm Mimica's
co-founder and CEO. Thank you all for being here.
[0:17] So today I'm going to talk about what we believe
is the biggest challenge deploying AI agents
[0:26] at scale, which is understanding how the work
in your organization is actually happening.
[0:37] So taking a step back,
[0:39] we are about to embark on a journey that
we call within Mimica the great transition.
[0:50] We're going to talk about what is the
great transition. Why will it be hard?
[0:55] How can we make it easier? And finally,
our take on a question that might be on
[1:01] a lot of people's minds, which is what is
the role of the human in a world of agents.
[1:10] Before we get started a
little bit about Mimica, um we
[1:16] help organizations understand
how work is actually happening.
[1:20] We work with a lot of Global 2000 and Fortune
500 companies. Some of the customer names you
[1:25] see on the screen here, and we are helping
these companies go through this transition.
[1:32] So, let's start with a bold probably not very
controversial statement. In 10 years time,
[1:41] the vast majority of the world's manual
digital work will be done by agents.
[1:49] The advent of LLMs has meant that a lot
of the work that people are doing today
[1:56] on their computers can and will be automated.
[2:04] And yet at the same time, we constantly hear
in the news that there's this gap between the
[2:09] promise of agentic AI and today's reality on
the ground. Gartner predicts that about almost
[2:17] half of agentic AI projects will be canceled by
2027. So, there's this massive disconnect. We see
[2:23] the power of the models. We've all experienced
it ourselves. We believe that this technology
[2:29] will eventually automate this work, and yet the
picture on the ground somehow doesn't match up to
[2:38] um the inspirational feeling that we get
when we play with these models ourselves.
[2:45] So, let's talk a little bit about the cost
[2:48] of building an agent and try to
understand what might be happening.
[2:54] So, building an agent essentially has two
cost components. One is writing the code,
[3:01] writing the code for the agent, and
then two is knowing what to write.
[3:17] Now, coding agents are essentially eliminating
the first cost, right? We've seen Claude code. The
[3:27] multi-agent structure that they've built behind
that technology is capable of planning the code
[3:33] to write, debugging the code, testing the code.
So, writing the code has essentially collapsed
[3:40] to zero, which means the bottleneck that remains
is the second, knowing what to write. This is the
[3:46] key challenge. But knowing is the hardest part. A
lot of the practitioners in the room probably know
[3:53] that companies don't have visibility into their
end-to-end processes. The tools and technologies
[4:00] that we have today to understand what people do on
their computers is pretty limited. Before Mimica,
[4:10] a big problem is that employees don't
accurately describe the way that they
[4:15] work. If you ask someone how they do
their process, they tend to talk about
[4:18] the happy path and they forget many of the
variations and exceptions in the process.
[4:24] And it's hard to know of all
the people in my organization,
[4:27] what is the work that is the low-hanging fruit,
the most ripe for automation with agents today.
[4:38] And to make matters certainly not easier,
the processes themselves are very complex.
[4:46] Folks rarely perform work the same way.
Those of you that work in multinational
[4:52] organizations know that the work is happening
differently in Germany, in the US, in Brazil.
[4:58] These processes include dozens
or hundreds of variants.
[5:05] So,
[5:09] you can't automate what you can't see. This
is the crux of the problem, you might say.
[5:17] Writing the code is the easy part. Knowing what to
write is the challenge. Now, it turns out that um
[5:27] there might be a way to get better visibility
into what people are doing on their computers.
[5:33] The Windows operating system, for example,
gives some information about the buttons
[5:39] that are being interacted with, the
fields that are being interacted with.
[5:44] And in the bottom right-hand corner, you'll
see a sensor. That's a Mimic a sensor. It
[5:50] captures all the click and keystroke data of what
people do on their computers. It anonymizes it.
[6:00] And then using the anonymized data, it makes
predictions about work that can be automated, what
[6:07] the processes are. We'll talk more about that a
little bit later. Now, you'll notice that it has a
[6:14] start and a stop and a pause button on it, very
important. So, the user is in complete control of
[6:21] the recording at all times. They turn it on, they
turn it off, they can pause it whenever they want.
[6:28] So, I'm going to take a step back from this
recording technology for a second and just talk
[6:35] about the five phases of Agentyc deployment.
And what I'd like you guys to do is imagine,
[6:43] as we go through these phases, how a recording
technology might assist in the execution of these
[6:51] five phases. So, first, we call selection,
choosing what to automate. I've got anywhere
[6:57] from 500 to 50,000 people in my company. What of
the work is the best candidate for automation?
[7:06] Two is mapping, traditionally done
through employee interviews. Uh
[7:12] people give anecdotes about the way they work.
It's not the best way to understand a process,
[7:16] but it's a critical first step to deploying an
agent. Then the third is what we call Genesis,
[7:23] using the data that we've captured
to build a first version of an agent.
[7:29] Now, the way these Agentyc automation platforms
work is you put the agent into production
[7:36] and anything the agent's not able to accomplish,
like an exception, gets punted to the human.
[7:43] And there's this period of time where the
human and the agent are working in tandem,
[7:48] where the agent's handling the happy path
cases, the human's handling the exceptions.
[7:53] And the goal is to iteratively harden or expand
the scope of the work of the agent over time
[8:00] until it finally reaches maturity, where the agent
now does the work. Now, agents make mistakes,
[8:08] so you will always have someone in a supervisory
role confirming that the agent's work is happening
[8:14] well. But these are the five phases of agentic.
So now I'm going to talk about when you have
[8:22] a technology that can automatically capture
the work that people do on their computers,
[8:27] how does it impact each of these five phases?
So I'm going to talk about three tools
[8:37] that Mimica has. You know, other companies do this
in in different ways. But they're going to take us
[8:42] through this journey of selection, mapping, and
genesis. So first, there's a tool called Miner,
[8:49] which you deploy at scale across hundreds or
thousands of individuals. And what it does is
[8:55] identify the individual tasks and processes those
teams are doing and determine how good they are
[9:03] from the perspective of how good a candidate they
are from the perspective of agentic automation.
[9:09] Uh does there need to be any reasoning? Is it
more of an LLM-powered deterministic workflow?
[9:18] How standard is the process? How many variations
are there? How high volume is it? All questions
[9:24] that you might want to ask to kind of choose
which candidates to go after first. Then
[9:29] a critical step, which is mapping. We are going
to map in extraordinary detail all the clicks,
[9:37] keystrokes, exceptions, variants, and we're going
to capture from the data the decision logic,
[9:44] the thresholds, the business rules.
Very critical step. And of course,
[9:50] if there's any ambiguity in the recorded data or
there's tribal knowledge that we need to capture,
[9:56] we can use the software to pose a question to
the SME. So, rather than a blank slate interview
[10:04] where they're expected to fill a whiteboard from
scratch, you're showing them a screenshot evidence
[10:10] of something that they did at a particular point
in time and posing a question to them. Hey,
[10:15] can you give me a little bit more information
about why you did this particular step?
[10:21] And then finally, it turns out that this process
map can in fact be the genesis of the agent. Now,
[10:31] you may want to re-engineer the process a little
bit first, maybe make some modifications to it,
[10:36] but we can see a human doing 15 clicks and
keystrokes and understand the exact API call that
[10:46] that represents. Or you can turn it into a GUI
automation. Or going back to the invoice example,
[10:55] we have 15, 50, 500 production examples
of how these invoices were processed. So,
[11:03] what are the steps that the person does with the
invoice? Does it go into an Excel spreadsheet?
[11:08] Does it go into SAP? Does the spreadsheet
need to be uploaded somewhere? How do they
[11:12] determine the GL code associated I'm using,
you know, fi- finance examples, but of course,
[11:17] all back office operational work um falls in falls
into this category. So, from this you can actually
[11:26] generate an agent automatically. This is the
future, we believe, of deploying agents at scale.
[11:37] And there's a very interesting idea that
falls out of this, which is that the best
[11:44] agents actually are powered by people. We're
using the top performers in the company today,
[11:52] their examples of the way that they work. That
information is being used to mimic, in a sense,
[12:02] that individual's most repetitive work.
The best agents are powered by people.
[12:10] Now, I'll just talk very briefly about where
agents mostly are today and how they're going to
[12:17] evolve over time. So, a lot of the work that is
still manual and repetitive today is frankly work
[12:26] that probably could have been automated with
RPA. It just takes so long to actually get a
[12:31] RPA bot into production that a lot of that work
remains unautomated. Or there's some step in the
[12:41] process that required reasoning, human judgment,
and so the process just went unautomated.
[12:48] So, in this way, the first generation of agents,
or what you might call deterministic agents,
[12:54] they're powered by workflows. They do
click, keystroke, GUI automation, LLMs,
[13:00] uh sorry, API automation. And then, if you need
to read an unstructured email at one point in
[13:06] the process, or generate an email response, or you
need to read an order form, or an invoice, or a um
[13:13] you know, freight bill, or whatever whatever
the form is, the LLM can do that step,
[13:18] and then the agent executes the rest of the
process deterministically. This is our like
[13:23] V1 nice and easy agents. We're not giving them
too much responsibility. We just really need to
[13:31] have a clear understanding of what is the business
process. Then you get into planning agents. Let's
[13:39] say the customer you're talking about claims in
the insurance industry and you have a series of
[13:45] discussions with the customer. You call the
customer and you ask for some documentation.
[13:49] Customer says, "Hey, I'm in a meeting right
now. I don't have time to take this call."
[13:52] This is where an agent can start to reason
about given the input that I was given,
[13:58] what is the next best step? And that's where you
get into planning and reasoning. And then finally,
[14:06] the end-to-end agent. This is the agent that can
do KYC from end to end or the agent that can do
[14:12] the entire claims process from end to end.
And no matter what you throw in front of it,
[14:17] it's going to be able to accomplish that task.
Again, we believe that no matter where you are
[14:24] on this journey, the examples, the production
examples of how your process is done today, those
[14:32] clicks and keystrokes that humans are performing,
that becomes the harness, the guardrails,
[14:38] it becomes the foundation of how these agents
actually work in production and how you get
[14:44] confidence that this agent is doing work
that is ultimately powered by your people.
[14:53] Now, I touched on this earlier today.
[14:57] I just want to talk briefly about it. What is
the role of the human in the world of agents?
[15:02] We believe that humans become supervisors.
So, you're not necessarily doing the work.
[15:11] You're supervising the work of agents. You're
ensuring that they're doing the work well.
[15:17] You are ensuring that it's high quality.
You're doing checks along the way, and in this
[15:23] way the productivity of that one person improve
significantly. They go from being able to do one
[15:30] person's work to being able to do five people's
work or seven people's work or 10 people's work.
[15:37] And it's an exciting future for the people
that think that it's an exciting future.
[15:46] And a lot of companies that we work with that
are really deploying agentic AI at scale are
[15:52] thinking a lot about how to create a culture that
incentivizes people to embrace this transition
[16:00] and move into a role where they're not necessarily
doing the work, but but supervising it.
[16:12] All right, that concludes my
talk. Thank you all for joining.
[16:15] I'm happy to take any questions
that you guys might have.
