Detect Intent, Then Tailor Every Screen to the Person

Episode Summary

Al Shanmugam runs product for Dish Network and Sling TV at EchoStar, which means he owns what a viewer sees, how they search, the notifications they receive and the advertising they are served. He describes AI in that domain as three things: automation, personalization and intent detection. The old approach was segment based, grouping customers together and sending the whole group one offer. The new one learns about a person rather than a cohort, and works out what someone wants without being told, so the same page can look different to two people. Agents close the loop by acting on that signal and reaching back out on whichever channel the customer actually uses, inside guardrails and brand validation that stop a model speaking for the company unchecked. He measures all of it on human involvement removed, churn and conversion. His forecast is that integration layers and interim databases disappear while systems of record stay.

Key takeaways

  • He owns four surfaces at EchoStar and names them himself: what the customer sees in recommendations, how the customer searches and interacts with the platform, the notifications going out by email, text and push, and the advertising, including whether it is contextual and whether it is useful

  • His framing of AI in a product organization is three things, in this order: automation, personalization and intent detection. He notes at the end of the same answer that the three sit at different maturity levels inside his organization

  • Automation is different now because the old version ran on rules and this one is contextual, he says. Software and systems used to do that job, and what he describes in their place improves continuously from the mistakes being made

  • His own definition of AI, offered as a working description rather than a term of art, is a self understanding or goal-driven system

  • Segment-based personalization is the thing being replaced. The old pattern was to decide that a group of customers likes something, put them all in one segment and send that segment the offer

  • What replaces it is one-to-one, and he calls it hyper-personalization: AI learning about a person rather than a cohort, because the interest that matters is often nuanced

  • Intent detection is the third leg and the one that makes personalization act rather than describe. If he knows what you are looking for without your saying it, he can customize the product for you and tailor every screen

  • The endpoint he names is that a single dot com page renders differently for two different people, and he says that is where this is going rather than where it already is

  • He treats agents as the next step rather than a current state, and defines the step precisely: text and images were generation, and what he wants now is actions

  • The example he works through is abandonment. A customer searches, looks at a product, compares variations and then leaves, and the agent reads that behavior, predicts the next step, judges whether a better offer is what is missing, and reaches back out

  • His argument for agents is org structure rather than capability. Companies run those same steps today across several different teams, and one agent, or a set of connected agents, can run the whole workflow

  • Human in the loop is a design requirement in his account, not a safety add-on: you have to think about it while designing, because the guardrails are what stop the AI acting on its own

  • Brand validation is the specific guardrail he names for anything customer-facing. The model has to be trained on the brand through RAG, and it must not communicate with a customer without that validation

  • He makes brand concrete rather than abstract: the EchoStar brand is family based and sports centric, so the communication style, the colors and the creative all have to match, and that differs from other brands

  • The drudgery he says AI takes back is testing. Releases used to absorb days and sometimes months before anything shipped, and AI-based testing is what he names as the change

  • His Northstar KPIs are two: how much human involvement has been removed, and how much customer conversion is happening

  • The clearest saving he quantifies is deflection. A query solved in the chat bot rather than routed to a call center with a human saves the cost of that call, and at high call volume that is money on every one

  • The second saving is structural: skipping systems means fewer interfaces, less infrastructure cost, fewer people maintaining it, and fewer of the small partners whose business is piping data between systems

  • On the retention side he frames churn as a support problem. Customers used to contact support and then leave, so a falling number of departures after that contact is a cost saving he can point at

  • What he wants instead of several models is one that handles text, image, video, audio and user intent detection together, which he calls multimodal convergence

  • Personalization is multimodal in his answer, not just copy: personalized creatives, personalized messages, personalized in-app notifications, and the preferred channel personalized too, because an email person and a text person are different customers

  • His test for whether the churn of new releases matters is ownership of data rather than choice of model. If the model is working on your own proprietary information and you have a good RAG in house, you are good

  • He uses LLMs rather than small models, and puts the small-model case on the device. Companies like Apple and others might find them useful, he says, because the deployment target has limited memory capacity, while the bigger enterprise companies run most of their use cases server side with a lot more compute available

  • His rule for when to stop building is arithmetic. If maintaining and training the models costs a lot and the result is a three or four percent improvement, go with a partner instead

  • The model decision is made by an enterprise-wide committee, and his reason is sprawl: engineers cannot each pick a tool when there are too many options, and policy on what data may be used for training has to come from one place

  • Legal sits on that committee, which he flags as new. Old technology committees did not include legal, and now every data set exposed to a model and every API called is a legal question

  • The committee has no single reporting line, in his experience at EchoStar and previously. It is assembled from engineering, product, marketing, customer support and customer experience, and he says it depends on how the organization is structured, because one function alone ends up solving one set of problems

  • Agents get their own KPIs underneath the business ones, split the way the work is split: agent performance and reduction on the agent side, guardrails and error rate on the human side

  • His forecast for enterprise AI starts with goals rather than technology: companies will carry AI-driven OKRs, with a set percentage of productivity expected to come from AI, and will treat AI as an investment

  • The systems that survive are the ones that hold the record. CRMs stay because the data has to live somewhere and you need a single source of truth

  • The systems that go are the integration layers, and he puts them first. Anything whose job is moving data between two other systems gets replaced by an agent doing that job

  • Monolithic systems go the same way, and so do the replicas. Instead of keeping your own copy you read live from the source of truth, which for inventory means reading from the supplier

  • The casualty he names that is a team rather than a technology is the database function: interim databases, database teams and database systems, unless the data is proprietary to you

  • SaaS does not die wholesale in his account, it splits. SaaS that holds a source of truth stays; SaaS that only sends email or runs social posting is in trouble unless it adds something

  • His first instruction to anyone starting is data, not models: build a single source of truth for the customer, a CDP or something like it, because personalization needs to know the person first

  • His second is to experiment with pre-trained LLMs and not train your own, and to build the RAG in house with your own vector database and your own embeddings, because AI reads through embeddings and will not query your Oracle database directly

  • His third is to have KPIs at all, and he has seen people run experiments without them. The point is not the cool factor, it is whether the business moved

  • His closing takeaway is the goal-driven framing again, with a number attached: humans set a goal, devise strategies and iterate over six months, and he says AI can run that loop and reach the goal in ten or fifteen days

About Al Shanmugam

Al Shanmugam is Head of Product at EchoStar, where he leads product for the satellite communications and streaming division, working across Dish Network and Sling TV. His remit covers recommendations, search, customer notifications and ad tech, and his focus is personalization, automation and AI-driven streaming that lifts engagement and monetizes content delivery. He was doing AI at scale at Amazon and Meta before generative AI arrived, which is the background he brings to intent detection and one-to-one personalization at consumer scale. He studied engineering at VIT University.

 

In this episode

00:41 Welcome
01:05 Introducing Al Shanmugam of EchoStar
01:22 The five companies inside EchoStar, and which two he works on
02:36 What Sling carries: sports, then cricket and soccer, then news
04:00 Christina: are you the one helping me find things, or selling me ads
04:20 His remit: recommendations, search, notifications and ad tech
04:52 Christina: the four obvious AI use cases, search, ads, recommendations, notifications
05:17 AI at scale at Amazon and Meta, before the generative wave
05:33 The three areas: automation, personalization, intent detection
05:33 His definition of AI: a self understanding or goal-driven system
06:09 Segment-based personalization, and what replaces it
06:28 A person rather than a cohort
06:46 Intent detection: knowing without being told
06:46 Tailoring every screen, so one page renders two ways
07:01 Maturity levels differ by area
07:26 Agents as the next step, and the shift from answers to actions
08:31 The agent reads the behavior and reaches back out
09:00 One agent, or connected agents, instead of several teams
09:28 Own agents, for privacy reasons
09:40 Human in the loop as a design requirement
09:58 RAG, and brand validation before anything reaches a customer
10:14 What brand means concretely: family based, sports centric
11:04 Testing as the drudgery AI takes back
11:28 Christina: how are you judging the ROI of that workflow
11:51 Northstar KPIs against system-level technical ones
12:06 Deflection: the call that never reaches the call center
12:23 Why the old chat bots could not do it
12:42 Skipping systems, and the costs that go with them
13:32 Churn as a support problem
14:22 Christina: would the images be personalized too
14:42 Multimodal convergence, in one model rather than several
14:56 Personalized creatives, messages and notifications
15:15 Personalizing the channel itself
16:34 A separate model for recommendations, another that understands ads
16:51 New models every week, and the catching-up problem
17:13 Own your data and your RAG, and the release churn stops mattering
18:02 Small models are for on-device, enterprises have server-side compute
18:35 A high maintenance cost against a three or four percent lift
18:53 Spending more to get five percent, and the case for a partner
19:15 The clean room, where partner and first-party data meet
19:44 Llama for self-managed, Claude when Anthropic carries the patching
20:06 Christina: how does a new model get decided internally
20:39 The enterprise-wide committee, and the sprawl it prevents
21:04 Policy on what data may be used for training, from one place
21:31 Legal on the committee, which older technology committees lacked
21:53 No single reporting line, at EchoStar or before it
22:15 Assembled from engineering, product, marketing and customer experience
22:35 Christina: is there someone who leads it
22:37 Up into the CEO, and in some companies the CIO
23:08 Christina: do the agents have KPIs of their own
23:26 Agent KPIs underneath the business ones
23:55 AI-driven OKRs, and AI as an investment
25:04 Christina: do you see any of those systems going away
25:18 What stays: CRMs, and the single source of truth
25:31 What goes first: the integrators
26:21 Interim databases, database teams, database systems
26:58 Which SaaS is in trouble, and which is not
27:27 Christina: what is your guidance for someone early
27:47 Start with a source of truth for the customer, a CDP
28:45 Build the RAG in house, with your own vector database
29:21 And have KPIs, because he has seen experiments without them
29:54 Christina: what resources do you recommend
30:33 Codex, and reading a code base without being a coder
31:22 Hugging Face Transformers, LangChain and LlamaIndex
32:11 Christina: what should listeners take away
32:37 Set the goal, let AI iterate the strategies
33:02 His closing line: set the goal, use AI to reach it
33:07 Sign-off, and the invitation to the November summit

In Al’s words

“a self understanding or a goal-driven system. That’s how I would describe AI”

— Al Shanmugam   (05:33)

“About you as a person, not as a cohort or not as a segment”

— Al Shanmugam   (06:28)

“So if I know, what you’re looking for. Without you explicitly saying, I can actually, customize my product for you”

— Al Shanmugam   (06:46)

“a lot of companies do all of these steps using like different teams, right? So now AI can do all of this as a single agent”

— Al Shanmugam   (09:00)

“guardrails are very important. So you have to implement all these guardrails so you don’t really let AI do on its own”

— Al Shanmugam   (09:40)

“it has to be aligned with your brand. So it should not, communicate to customers without, brand validation”

— Al Shanmugam   (09:58)

“we are talking about personalized creatives, personalized messages, personalized notifications”

— Al Shanmugam   (14:56)

“As long as your model is actually working with your own proprietary information, you’re good”

— Al Shanmugam   (17:13)

“the systems that are gonna go away are integrators”

— Al Shanmugam   (25:18)

“think about, a goal setting and then using AI to reach the goal”

— Al Shanmugam   (33:02)

 

Resources

Websites and projects

Al Shanmugam

What he recommends

  • OpenAI Codex: The tool he says at 30:33 he loves, and the reason he gives is that you can understand a code base without being a coder or a computer science expert. He mentions meeting OpenAI three or four weeks before this conversation to talk about it

  • Hugging Face Transformers: He describes it at 31:04 as GitHub for AI projects and AI apps, and at 31:22 says it is where to look at what is currently going on

  • LangChain: One of two frameworks he names at 31:22, and at 31:40 he gives the reason: you do not need a lot of insight into these technologies to run them

  • LlamaIndex: The second of the two frameworks he names at 31:22, for the same reason: you do not need a lot of insight into these technologies to run them

  • Chroma: The vector database he names as an example at 28:45 when telling anyone starting out to build their own RAG, and his own words on the choice are that anything is fine

Named on air

  • Llama: Meta’s open-weight family, which he names at 19:29, and at 19:44 he gives the reason for reaching for it: you do not want to rely on a company, so you do your own training and your own maintenance

  • Claude: The managed option he names at 19:44, where Anthropic carries the patching and the model updates so his team can stay on business innovation rather than model maintenance

  • Amazon and Meta: Where he says at 05:17 he used AI at scale before the generative wave. His current employer is EchoStar

  • Apple: His example at 17:42 of where small models earn their place, and at 18:02 he gives the reason: the deployment target is a device with limited memory capacity

  • AWS: Named at 18:02 as the kind of cloud compute that makes server-side inference practical for a large enterprise

  • Oracle: Used at 28:45 as the example of the database AI will not query directly, which is his argument for embeddings and a vector store

  • Boost Mobile, Hughes: Two of the five EchoStar companies he lists at 01:28, alongside Dish Network, Sling TV and EchoStar itself. He does not work on these

  • CNN, CNBC, Bloomberg, Fox: The news channels he names at 02:54 when describing who comes to Sling for what

Ideas and terms discussed

  • Intent detection: The third of the three areas he names, alongside automation and personalization. He describes it as knowing what someone is looking for without their saying it, so the product can be customized for them

  • Hyper-personalization: His word at 06:09 for one-to-one rather than segment-based, and he defines it by contrast: AI learning about you as a person, not as a cohort

  • Segment-based personalization: The practice being replaced. Decide a group likes something, put them in one segment, send the segment the offer

  • Brand validation: The guardrail he names for outbound customer communication. The model is trained on the brand through RAG, and nothing reaches a customer without passing that check

  • Northstar KPI: His term for the top-level measure, against the system-level technical KPIs underneath it. His two are human involvement removed and customer conversion

  • Clean room: The privacy-controlled environment where partner data and first-party data are transacted under compliant infrastructure, with the AI trained on top of it

  • CDP: Customer data platform, his name for the single source of truth about a customer that he says to build before anything else

  • Multimodal convergence: One model that understands text, image, video, audio and user intent together, which he wants because keeping up with several separate models is the problem

  • Source of truth: The property that decides which systems survive his forecast. Hold the record and you stay, move data between systems that do and you are the integrator being replaced

Related AI Realized episodes and events

 

Frequently Asked Questions

 
 
 
 
 
 
 
 
 
 
 
Previous
Previous

More Podcast Episodes

Next
Next

AI Agent Sprawl: Govern the Lifecycle Before You Backtrack