Internal AI Agent Platform for Enterprise Data Access and Product Development

amplitude · Tech · 2024 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Internal enterprise AI agents for product development and data access

Amplitude needed a way for employees across product, engineering, sales, support, marketing, and leadership to access enterprise-public knowledge, ask business questions, synthesize customer feedback, and generate product artifacts without manually searching many siloed systems. The organization also needed a shared AI fluency path because employees started from very different levels of AI experience.

Constraints:

Architecture

Amplitude built Moda, an internal AI agent platform on its Langley framework, with a Slack bot and proprietary web UI. The system uses Glean APIs as the enterprise-search/RAG backbone over enterprise-public sources, can also search externally, and exposes specialized workflows such as ask-anything, deep research, and PRD/prototype generation. The PRD workflow is configured in GitHub/YAML and chains problem exploration, solution exploration, detailed requirements, and prototype-prompt generation. Outputs can be reviewed, commented on, regenerated from higher levels in the stack, and used in tools such as Bolt, Figma, and v0.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  Enterprise_users["Enterprise users"] --> Slack_bot_interface["Slack bot interface"]
  Enterprise_users --> Web_UI["Web UI"]
  Slack_bot_interface --> Langley_framework["Langley framework"]
  Web_UI --> Langley_framework
  Langley_framework --> Glean_API_retrieval_layer["Glean API retrieval layer"]
  Langley_framework --> External_search["External search"]
  Glean_API_retrieval_layer --> Enterprise_public_data_sources["Enterprise-public data sources"]
  Enterprise_public_data_sources --> Data_access_policy_boundary["Data-access policy boundary"]
  Web_UI --> PRD_orchestrator["PRD orchestrator"]
  GitHub_YAML_configuration["GitHub YAML configuration"] --> PRD_orchestrator
  PRD_orchestrator --> Glean_API_retrieval_layer
  PRD_orchestrator --> Prototype_tools["Prototype tools"]
  PRD_orchestrator --> Human_review_workflow["Human review workflow"]
  Human_review_workflow --> Enterprise_users
  Contributor_workflow["Contributor workflow"] --> GitHub_YAML_configuration
  Enterprise_users --> Contributor_workflow
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Slack bot interface Social entry point where employees ask Moda questions and observe each other's successful prompts and results. Langley framework, Enterprise users
Web UI Richer proprietary interface for ask-anything, deep research, PRD creation, and specialized workflows. Langley framework, PRD orchestrator
Langley framework Amplitude's internal framework for processing AI requests from Slack and web surfaces. Slack bot interface, Web UI, Glean API retrieval layer, External search
Glean API retrieval layer Enterprise-search/RAG layer that retrieves content across approved internal systems. Langley framework, Enterprise-public data sources
Enterprise-public data sources Approved internal sources such as Confluence, Jira, Salesforce, Zendesk, Slack, Google Drive, Product Board, Zoom, Outreach, Gmail, Asana, Dropbox, GitHub, and HubSpot. Glean API retrieval layer, Data-access policy boundary
Data-access policy boundary Excludes private, personal, and restricted datasets from Moda's accessible corpus. Enterprise-public data sources, Glean API retrieval layer
External search Allows Moda to search public information in addition to internal knowledge sources. Langley framework
GitHub YAML configuration Version-controlled configuration for agent behaviors, Glean search, and PRD orchestration prompts. PRD orchestrator, Contributor workflow
PRD orchestrator Multi-step workflow that turns a product idea into problem exploration, solution exploration, detailed requirements, and prototype prompts. Web UI, GitHub YAML configuration, Glean API retrieval layer, Prototype tools, Human review workflow
Prototype tools Downstream tools such as Bolt, Figma, and v0 that receive generated prototype prompts for comparison. PRD orchestrator, Human review workflow
Human review workflow Product, design, and engineering review of generated problem statements, solution approaches, requirements, alternatives, and prototypes. PRD orchestrator, Prototype tools, Enterprise users
Contributor workflow GitHub pull-request path where engineers, product managers, and designers extend or modify Moda workflows. GitHub YAML configuration, Enterprise users
Enterprise users Amplitude employees across functions who ask questions, generate artifacts, observe Slack usage, and provide feedback. Slack bot interface, Web UI, Human review workflow, Contributor workflow

Data flow

Employees interact with Moda through Slack or the web UI. Requests flow into Amplitude's Langley framework, which calls Glean APIs to search enterprise-public sources such as Confluence, Jira, Salesforce, Zendesk, Slack, Google Drive, Product Board, Zoom, Outreach, Gmail, Asana, Dropbox, GitHub, and HubSpot while excluding private or restricted datasets. For PRD creation, a YAML-defined orchestrator starts from a user idea, retrieves context through Glean, performs problem and solution exploration, generates detailed requirements, and produces prototype prompts. Users review generated sections, leave comments, and can regenerate downstream sections when an upstream assumption changes.

4 supporting quotes
“We we have a framework that we've used internally and we have both a little bot that sits in Slack that can call out to that framework.”transcript, stated
“we use that as um part of our rag to make sure that we go get the right content and then pass it off and it's able to evaluate all of that and come back with a much better answer.”transcript, stated
“If it needs more details, it's going to ask you for that. It's then going to take all of this and break it down of where it will break things into let's make sure that we do proper problem exploration, solution exploration, detailed requirements”transcript, stated
“you can create a comment um associated with it and then you can tell it to go re-evaluate things and so you're not stuck with the answer that you got”transcript, stated

Why this architecture

Amplitude built Moda internally because the leverage came from deep integration with company data and workflows, but they kept the implementation lightweight by using off-the-shelf pieces such as Glean APIs and external prototype tools. Slack made AI use visible and viral, the web UI enabled richer workflows, GitHub/YAML made the system extensible, and the review/regeneration loop let teams use AI for acceleration without treating generated artifacts as final.

4 supporting quotes
“the more that you can do it internally, as long as you don't invest a lot in, you know, overbuilding it and overengineering it, uh, you'd probably be better off doing it yourself.”transcript, stated
“we were able to pull a lot of things off the shelf like glean APIs and things along those lines, which just allowed us to move really quick.”transcript, stated
“if I can see what other people are doing and it's adjacent to the work that I'm doing, I'll b I'll borrow the prompt.”transcript, stated
“You actually have to apply critical reasoning to see where it may have failed you.”transcript, stated

Models used: Moda internal agent · Langley framework · Glean API · large language models behind Moda · Bolt · Figma · v0 · Cursor

Scale

Alternatives considered (roads not taken)

Buy off-the-shelf AI tools

Rejected because: Amplitude wanted internal leverage from its own data and workflows, and peer advice suggested building internally if they avoided overengineering.

“About half of them were saying, "Hey, I'm I'm pulling things off of the shelf and using it." And the other half were kind of like, no, the way that you're going to want to use this and how you're going to get more leverage”transcript, stated
“as long as you don't invest a lot in, you know, overbuilding it and overengineering it, uh, you'd probably be better off doing it yourself.”transcript, stated

Build all enterprise search and data-access plumbing from scratch

Rejected because: Glean APIs simplified retrieval, data access, and controls so the team could move quickly.

“we were able to pull a lot of things off the shelf like glean APIs and things along those lines, which just allowed us to move really quick.”transcript, stated
“you're using the Glean API for a bunch of your enterprise search. So that simplifies a little bit the data access and controls and all that piece the rag piece of it.”transcript, stated

Expose the tool only through a private web interface

Rejected because: Slack visibility helped employees learn from each other and drove viral adoption.

“Because if I can see what other people are doing and it's adjacent to the work that I'm doing, I'll b I'll borrow the prompt.”transcript, stated
“The more that I can make this publicly visible and do real solid work on the other side of it, provide great answers on the other side of it, then it should catch like fire.”transcript, stated

Use a purely iterative chat flow for PRD generation

Rejected because: Amplitude designed a multi-step PRD workflow that can generate problem exploration, solution exploration, requirements, and prototype prompts with little user intervention, while still allowing comments and regeneration.

“What I like about the design of this internal flow is it's clearly multi-step without user interaction”transcript, stated
“you can create a comment um associated with it and then you can tell it to go re-evaluate things and so you're not stuck with the answer that you got”transcript, stated

Patterns

rag-retrieval (retrieval)

Moda uses Glean APIs to retrieve enterprise-public content from many internal systems, then passes that context into AI workflows for answers, thematic analysis, and PRD generation.

“we use that as um part of our rag to make sure that we go get the right content and then pass it off and it's able to evaluate all of that and come back with a much better answer.”transcript, stated
“what's the right context to provide to Moda to make sure that that we got it right or can it provide us that context by searching through all of the enterprise data that we've got”transcript, stated

tool-use-orchestration (architecture)

Moda orchestrates Slack, web UI, Glean search, external search, GitHub-configured workflows, and prototype tools into one internal agent platform.

“We we have a framework that we've used internally and we have both a little bot that sits in Slack that can call out to that framework.”transcript, stated
“we have a proprietary interface. Actually, let's jump over there.”transcript, stated

prompt-chaining (workflow)

The PRD workflow decomposes one product idea into staged problem exploration, solution exploration, detailed requirements, and prototype prompts.

“it will break things into let's make sure that we do proper problem exploration, solution exploration, detailed requirements that come out the other side.”transcript, stated
“We like to move as quickly as possible to a prototype. And so if it can do prototype generation or at least the prompts that then we can plug in”transcript, stated

human-in-the-loop-gate (governance)

Teams review generated PRDs, evaluate problem and solution statements, compare alternative prototypes, and apply critical reasoning before treating Moda's outputs as usable.

“almost every document that we produce on um this side will have a review segment that we go through and so we'll have people go through actually look at the problem statement look at the um solution statement”transcript, stated
“You actually have to apply critical reasoning to see where it may have failed you.”transcript, stated

observability-tracing (operations)

Amplitude uses Moda itself to inspect usage patterns, adoption by team, failed interactions, friction points, and data-access gaps.

“if I can understand where we have any friction inside of the the company um and what types of questions are being asked and and where are people not having success”transcript, stated
“you can see product management engineering sales customer support marketing our CEO um our our head of sales um I've seen our chief product officer in here”transcript, stated

structured-output-contract (interface)

Moda produces structured artifacts such as PRDs, detailed requirements, prototype prompts, and Confluence documents rather than only chat answers.

“it will talk about the problem exploration. It will talk about the actual uh solution exploration. and it will talk about the detailed requirements”transcript, stated
“for everything else it will generate and put it into the confluence document.”transcript, stated

Tradeoffs

Build a lightweight internal tool instead of buying a full off-the-shelf platform.

Chose internal build using off-the-shelf components such as Glean APIs over a purchased generic AI tool disconnected from Amplitude's workflows.

Rationale: The team wanted deep leverage over internal data and workflows while keeping the investment small enough to throw away or rebuild if needed.

“it's almost so cheap. I think, you know what, if in 3 months I throw the whole thing away, it will have been worth it anyway.”transcript, stated
“It's really unlocking people and the data that sits in the enterprise that I wanted to do the most.”transcript, stated

Make AI interactions visible in Slack rather than only private.

Chose a social Slack bot plus web UI over only a private individual AI interface.

Rationale: Visible usage helped employees learn from trusted peers and discover useful prompts and outputs.

“if I can see what other people are doing and it's adjacent to the work that I'm doing, I'll b I'll borrow the prompt.”transcript, stated
“If you can see people more credible or equally credible as yourself having great effect with this, it's it's an obvious thing that I want to use.”transcript, stated

Use Glean as the enterprise-search layer instead of building retrieval and permissions from scratch.

Chose Glean API retrieval over approved internal data over custom ingestion, indexing, and access-control infrastructure.

Rationale: Glean simplified access and controls so the team could focus on workflows and agent behavior.

“you're using the Glean API for a bunch of your enterprise search. So that simplifies a little bit the data access and controls and all that piece the rag piece of it.”transcript, stated
“we were able to pull a lot of things off the shelf like glean APIs and things along those lines, which just allowed us to move really quick.”transcript, stated

Generate a complete PRD/prototype package with review hooks rather than ask the user for every intermediate step.

Chose multi-step autonomous generation with comments and regeneration over constant interactive prompting for each section.

Rationale: The workflow gets close quickly, but preserves the ability to correct the highest wrong assumption and regenerate dependent sections.

“it's clearly multi-step without user interaction, which is quite quite interesting. And so instead of this sort of iterative, do you like this? Do you like this?”transcript, stated
“you can create a comment um associated with it and then you can tell it to go re-evaluate things and so you're not stuck with the answer that you got”transcript, stated

Gotchas (what bit them)

Enterprise AI adoption starts unevenly because employees have different levels of AI fluency.

Root cause: Some employees used ChatGPT or Claude at home while others had little experience, so a private tool alone would not create shared habits.

Resolution: Amplitude used Slack visibility and social learning so employees could see prompts, borrow successful interactions, and build a common language.

“the problem is is everybody's starting in a different place, right? And and so uh you use Chad GPT at home or you use Claude or you use something else.”transcript, stated
“if I can see what other people are doing and it's adjacent to the work that I'm doing, I'll b I'll borrow the prompt.”transcript, stated

Internal retrieval agents can expose too much if data boundaries are not explicit.

Root cause: Moda connects to many enterprise systems, so unrestricted retrieval would risk private, personal, or restricted information exposure.

Resolution: The system limits access to enterprise-public datasets and excludes private, personal, and restricted datasets.

“Confluence, Jira, Salesforce, Zindesk, Slack, Google Drive, Product Board, Zoom, Outreach for um transcribed uh meetings, those sorts of things, even some Gmail, Asauna, Dropbox, GitHub, HubSpot.”transcript, stated
“So, it does not have access to private, personal or restricted uh data sets. It's generally the public ones uh well, enterprise public ones um that we have internally.”transcript, stated

Generated PRDs and prototypes can look complete before the reasoning is actually right.

Root cause: The workflow compresses research, PRD generation, and prototyping into minutes, which makes it tempting to skip review.

Resolution: Amplitude reviews generated problem statements, solution statements, cons, alternatives, and multiple prototype variations before proceeding.

“You have created five beautiful detailed assets in I don't know three minutes. Does anybody actually read them or do they just click right to that prototype and say yes this is what we want?”transcript, stated
“if the person who is doing the generating hasn't also done some follow-up queries to say, you know, what are the cons on this? And is there evidence that suggests other answers would be better?”transcript, stated

Multi-step generation can fail upstream and contaminate every downstream artifact.

Root cause: Problem exploration, solution exploration, requirements, and prototypes depend on earlier generated sections.

Resolution: Users can comment on an incorrect section, regenerate from the highest wrong point, and let downstream sections refresh.

“if you don't think it got it right, like the downstream consequences are are fairly minimal because you can just go to as high up in the stack as possible”transcript, stated
“Just regenerate everything that's beneath it and you can just keep going down the stack as you need to.”transcript, stated

Internal agents do not eliminate the need for traditional product diligence.

Root cause: Some ideas are complex, market-facing, UI-heavy, or deeply integrated with the product and need more research and design coherence.

Resolution: Amplitude slows down for new product directions, market research, UI-heavy work, and deeply integrated features.

“If it's something that is a product direction or a entirely new product, um you're probably going to need to go do some market research.”transcript, stated
“if it's something that is UI heavy or, you know, deeply integrated with a part of our product, uh, we'll probably need to slow down a little bit”transcript, stated

Evaluation

Amplitude evaluates Moda through adoption and usage visibility, direct inspection of team usage, friction analysis, prompt/data-source adjustment, review segments in generated PRDs, follow-up queries for cons and alternatives, and human assessment of problem, solution, and prototype outputs. The talk reports strong adoption and workflow compression, but does not present a formal offline benchmark or quantitative quality score.

Metrics: company-wide usage within a week · usage across product management, engineering, sales, customer support, marketing, CEO, head of sales, and chief product officer · questions asked and unsuccessful interactions · data-access gaps · review of problem statement · review of solution statement · evidence for alternatives · prototype variation comparison · workflow compression from weeks to a single meeting

4 supporting quotes
“and then a week later it seemed like the entire company was using it.”transcript, stated
“if I can understand where we have any friction inside of the the company um and what types of questions are being asked and and where are people not having success”transcript, stated
“almost every document that we produce on um this side will have a review segment that we go through”transcript, stated
“what are the multiple solutions that you looked at? What did you generate from a prototype perspective so that we can see three different variations, maybe four different variations”transcript, stated

Outcomes

Moda was initially built in about three to four weeks of spare engineering time, went live after a Friday prototype and Monday push, spread across the company within about a week, accesses many enterprise-public systems, and can reduce some product-development workflows from weeks to a single meeting. The claims come from a transcript demo, not a formal performance report.

4 supporting quotes
“I started showing a couple of people um internally he's like oh this is really cool. You got to look at this thing. Uh and then a week later it seemed like the entire company was using it.”transcript, stated
“you can see product management engineering sales customer support marketing our CEO um our our head of sales um I've seen our chief product officer in here”transcript, stated
“we can actually put those three different roles together and actually produce that in a single meeting where we're going through and using MODA or um other tools”transcript, stated
“we've actually intentionally done that at times of where we've said, "Okay, you're going to take on a different role."”transcript, stated