Powered by Lemma Series
/
Bryant Lewis - Operations Strategist at Thread AI
Ben Allison - SVP, Head of Media Operations at VaynerMedia
December 9, 2024
In this second installment of our customer series, we showcase our partnership with VaynerMedia, a leading global digital marketing agency.
From our initial engagement, VaynerMedia demonstrated a clear commitment to innovation and a strong interest in exploring AI securely, responsibly, and at scale. VaynerMedia strongly believes its employees are its most valuable resource, and that the most compelling AI-powered workflows would augment how they work, to focus on optimizing efficiency and freeing up more time to deliver high-value client services.
AI-Powered Proposal Management Leads to 70% Improvement in Response Time.
Proposal management, regardless of the specific type of proposal (e.g., solicited or unsolicited), is a resource-intensive process for companies. It requires extensive research, drawing on both internal knowledge bases (past responses, internal data sources) and external sources (research reports, market data). The collaborative nature of proposal development, often involving multiple stakeholders working asynchronously, can further add to the complexity and time required. This is true whether the proposal is a less formal request for information (RFI) used in earlier procurement stages, or a more detailed and project-specific request for proposal (RFP) used later in the process. Both frequently necessitate leveraging existing resources and prior submissions to create a compelling and competitive response.
VaynerMedia gives each client's unique business challenges careful consideration. Developing a tailored proposal therefore involves a rigorous, multi-stage process. Upon receiving a proposal request, a designated lead coordinates information gathering, often seeking input from various teams, including senior executives. This individual also manages the collaborative drafting, review, fact-checking, and timely submission of the proposal. Oftentimes, this involves answering similar questions across different opportunities, but adapting the responses to the specific circumstances of each situation, such as the current market, the client's unique needs and priorities, and the particular project scope. This entire process takes several days, and occasionally weeks, to complete.
Both VaynerMedia and Thread AI have a similar philosophical approach when applying AI to improve and augment existing workflows.
1
Data Quality is Key
AI effectiveness depends on vetted, clean, well-prepared data, which requires investment.
2
Pragmatic Pilots
Begin AI implementation with internal, low-risk, non-customer-facing workflows.
3
Human Oversight is Crucial
Implement human guardrails throughout the AI workflow.
4
Vendor Selection with Guarantees
If and when evaluating various third-party model providers, choose providers who offer guarantees aligned with your business needs and ensure data flow transparency and auditability.
5
Develop a Robust Risk Assessment Framework
Establish a framework to assess the risks and implications of using probabilistic AI systems within your business processes.
6
Data & Information Security is Critical
Prioritize data security by employing robust risk assessment, vendor selection, and workflow management to ensure strict client data segregation, prevent unauthorized third-party model training or data access, and maintain client confidentiality, meeting all internal and external compliance mandates.
Given the above framework, VaynerMedia has adopted a more comprehensive proposal management strategy. This new approach leverages a Retrieval Augmented Generation (RAG) framework, efficiently built using Lemma Workers. The process is divided into two key stages, mirroring the core components of RAG. First, a "Data Hydration Worker" ingests and processes all relevant information to build a corpus of data. Second, a "Retrieval Worker" retrieves from the processed data the most pertinent information needed for proposal development and generates a “first pass” response. This AI-augmented proposal management workflow streamlines data aggregation and information retrieval, enabling VaynerMedia to dedicate more resources to developing and articulating strategic, client-centric solutions that address complex business challenges.
Data Hydration Worker
Below is an example of one version of the data hydration workers. Data exists in both structured and unstructured formats, ingested in the form of spreadsheets, documents, pdfs, and other representations. The data preparation ranges in complexity, but at a high level, the data ingested for processing must align with the access patterns for which it tends to be used (e.g. match expected query patterns). The structured data, or corpus, is then stored in a vector database (subject to defined retention policies) and immediately ready to be accessed.
Data Hydration WorkerRetrieval Worker
The retrieval worker offers access to multiple model providers at runtime, a feature that was important in VaynerMedia's workflow design. In addition, the worker allows for both precise (deterministic) and generative (stochastic) search. Deterministic search, where there is zero tolerance for any level of probabilistic output, ensures accuracy for factual queries (e.g. company founding date), while stochastic search supports more nuanced questions where some generation is required. For generative search, the worker has ensures all responses contain citations, which allow a human reviewer to reference and verify source data used in the response. Every question submitted to the retrieval worker is logged and audited using the same Runs product primitive employed by all workers. The modular worker design facilitates experimentation with different model providers, allowing VaynerMedia to tailor and configure the worker to meet specific goals or objectives.
Retrieval WorkerAt Thread AI we strongly believe that the best technology is invisible. It aids, augments, and ideally reduces the cognitive overhead of learning new tools and frameworks. This is why we designed the Lemma Workers to be unobtrusive, easy to build and easy to integrate for users. They can easily be plugged directly into familiar tools, like Google Sheets, Docs and other applications, empowering consumers of workers to leverage advanced capabilities without disrupting their established processes.
By redesigning its proposal management process to include AI-powered components, VaynerMedia has achieved a significant 70% improvement in proposal response time. The success of the work stems from a shared philosophy of prioritizing data quality, iterative implementation, human oversight, and thoughtful workflow design. The implementation and allowing the worker to be accessible where VaynerMedia employees currently do their work, both aid in the efficiency of incorporating new technology into an existing workflow. The results highlight how thoughtful and strategic AI implementation can significantly improve business processes without requiring drastic changes to existing workflows and technologies.
We are excited about the direction of our work with VaynerMedia and look forward to sharing more about other, innovative use cases VaynerMedia is building on the Lemma platform.
