At AWS, our gross sales groups create customer-focused paperwork known as account plans to deeply perceive every AWS buyer’s distinctive targets and challenges, serving to account groups present tailor-made steering and help that accelerates buyer success on AWS. As our enterprise has expanded, the account planning course of has change into extra intricate, requiring detailed evaluation, opinions, and cross-team alignment to ship significant worth to clients. This complexity, mixed with the guide evaluate effort concerned, has led to important operational overhead. To deal with this problem, we launched Account Plan Pulse in January 2025, a generative AI device designed to streamline and improve the account planning course of. Implementing Pulse delivered a 37% enchancment in plan high quality year-over-year, whereas reducing the general time to finish, evaluate, and approve plans by 52%.
On this publish, we share how we constructed Pulse utilizing Amazon Bedrock to scale back evaluate time and supply actionable account plan summaries for ease of collaboration and consumption, serving to AWS gross sales groups higher serve our clients. Amazon Bedrock is a complete, safe, and versatile service for constructing generative AI functions and brokers. It connects you to main basis fashions (FMs), companies to deploy and function brokers, and instruments for fine-tuning, safeguarding, and optimizing fashions, together with information bases to attach functions to your newest information so that you’ve every part that you must shortly transfer from experimentation to real-world deployment.
Challenges with growing scale and complexity
As AWS continued to develop and evolve, our account planning processes wanted to adapt to satisfy growing scale and complexity. Earlier than enterprise-ready giant language fashions (LLMs) turned accessible by Amazon Bedrock, we explored rule-based doc processing to judge account plans, which proved insufficient for dealing with nuanced content material and rising doc volumes. By 2024, three important challenges had emerged:
- Disparate plan high quality and format – With groups working throughout quite a few AWS Areas and serving clients in numerous industries, account plans naturally developed variations in construction, element, and format. This inconsistency made it tough to verify important buyer wants had been described successfully and constantly. Moreover, the analysis of account plan high quality was inherently subjective, relying closely on human judgment to evaluate every plan’s depth, strategic alignment, and buyer focus.
- Useful resource-intensive evaluate course of – The standard evaluation course of relied on guide opinions by gross sales management. Although thorough, these opinions consumed precious time that might in any other case be dedicated to strategic buyer engagements. As our enterprise scaled, this method created bottlenecks in plan approval and implementation.
- Data silos – We recognized untapped potential for cross-team collaboration. Creating strategies to extract and share information would remodel particular person account plans into collective greatest practices to raised serve our clients.
Answer overview
To deal with these challenges, we designed Pulse, a generative AI answer that makes use of Amazon Bedrock to research and enhance account plans. The next diagram illustrates the answer workflow.
The workflow consists of the next steps:
- Account plan narrative content material is pulled from our CRM system on a scheduled foundation by an asynchronous batch processing pipeline.
- The info flows by a sequence of processing phases:
- Preprocessing to construction and normalize the information and generate metadata.
- LLM inference to research content material and generate insights.
- Validation to verify high quality and compliance.
- Outcomes are saved securely for reporting and dashboard visualization.
We’ve built-in Pulse immediately with present gross sales workflows to maximise consumer adoption and have established suggestions loops that constantly refine efficiency. The next diagram exhibits the answer structure.
Within the following sections, we discover the important thing parts of the answer in additional element.
Ingestion
We implement a batch processing pipeline that extracts account plans from our CRM system into Amazon Easy Storage Service (Amazon S3) buckets. A scheduler triggers this pipeline on an everyday cadence, facilitating steady evaluation of probably the most present data.
Preprocessing
Contemplating the dynamic nature of account plans, they’re processed in each day snapshots, with solely up to date plans included in every run. Preprocessing is performed at two layers: an extract, remodel, and cargo (ETL) circulate layer to arrange required information to be processed, and simply earlier than mannequin calls as a part of enter validation. This method, utilizing the plan’s final modified date, is essential for avoiding a number of runs on the identical content material. The preprocessing pipeline handles the each day scheduled job that reads account plan information saved as Parquet information in Amazon S3, extracts textual content content material from HTML fields, and generates structured metadata for every doc. To optimize processing effectivity, the system compares doc timestamps to course of solely just lately modified plans, considerably lowering computational overhead and prices. The processed textual content content material and metadata are then remodeled right into a standardized format and saved again to Amazon S3 as Parquet information, making a clear dataset prepared for LLM evaluation.
Evaluation with Amazon Bedrock
The core of our answer makes use of Amazon Bedrock, which gives a wide range of mannequin selections and management, information customization, security and guardrails, value optimization, and orchestration. We use the Amazon Bedrock FMs to carry out two key features:
- Account plan analysis – Pulse evaluates plans towards 10 business-critical classes, making a standardized Account Plan Readiness Index. This automated analysis identifies enchancment areas with particular enchancment suggestions.
- Actionable insights – Amazon Bedrock extracts and synthesizes patterns throughout plans, figuring out buyer strategic focus and market developments that may in any other case stay remoted in particular person paperwork.
We implement these capabilities by asynchronous batch processing, the place analysis and summarization workloads function independently. The analysis course of runs every account by 27 particular questions with tailor-made management prompts, and the summarization course of generates topical overviews for easy consumption and information sharing.
For this implementation, we use structured output prompting with schema constraints to offer constant formatting that integrates with our reporting instruments.
Validation
Our validation framework contains the next parts:
- Enter and output validations are important as a part of the OWASP High 10 for Giant Language Mannequin Purposes. The enter validation is crucial by the introduction of vital guardrails and immediate validation, and the output validation makes positive the outcomes are structured and constrained to anticipated responses.
- Automated high quality and compliance checks towards established enterprise guidelines.
- Further evaluate for outputs that don’t meet high quality thresholds.
- A suggestions mechanism that improves system accuracy over time.
Storage and visualization
The answer contains the next storage and visualization parts:
- Amazon S3 gives safe storage for all processed account plans and insights.
- A each day run cadence refreshes perception and allows progress monitoring.
- Interactive dashboards provide each government summaries and detailed plan views.
Engineering for manufacturing: Constructing dependable AI evaluations
When transitioning Pulse from prototype to manufacturing, we applied a sturdy engineering framework to handle three important AI-specific challenges. First, the non-deterministic nature of LLMs meant similar inputs might produce various outputs, probably compromising analysis consistency. Second, account plans naturally evolve all year long with buyer relationships, making static analysis strategies inadequate. Third, completely different AWS groups prioritize completely different elements of account plans based mostly on particular buyer {industry} and enterprise wants, requiring versatile analysis standards. To take care of analysis reliability, we developed a statistical framework utilizing Coefficient of Variation (CoV) evaluation throughout a number of mannequin runs on account plan inputs. The aim is to make use of the CoV as a correction issue to handle the information dispersion, which we achieved by calculating the general CoV on the evaluated query degree. With this method, we are able to scientifically measure and stabilize output variability, set up clear thresholds for selective guide opinions, and detect efficiency shifts requiring recalibration. Account plans falling inside confidence thresholds proceed robotically within the system, and people outdoors established thresholds are flagged for guide evaluate. We complemented this with a dynamic threshold weighting system that aligns evaluations with organizational priorities by assigning completely different weights to standards based mostly on enterprise influence. This customizes thresholds throughout completely different account sorts—for instance, making use of completely different analysis parameters to enterprise accounts versus mid-market accounts. These enterprise thresholds bear periodic evaluate with gross sales management and adjustment based mostly on suggestions, so our AI evaluations stay related whereas sustaining high quality and saving precious time.
Conclusion
On this publish, we shared how Pulse, powered by Amazon Bedrock, has remodeled the account planning course of for AWS gross sales groups. By way of automated opinions and structured validation, Pulse streamlines high quality assessments and breaks down information silos by surfacing actionable buyer intelligence throughout our world group. This helps our gross sales groups spend much less time on opinions and extra time making data-driven selections for strategic buyer engagements.
Wanting forward, we’re excited to boost Pulse’s capabilities to measure account plan execution by connecting strategic planning with gross sales actions and buyer outcomes. By analyzing account plan narratives, we goal to determine and act on new alternatives, creating deeper insights into how strategic planning drives buyer success on AWS.
We goal to proceed to make use of the brand new capabilities of Amazon Bedrock for enhanced and sturdy enhancements to our processes. By constructing flows for orchestrating our workflows, use of Amazon Bedrock Guardrails, introduction of agentic frameworks, and use of Strands Brokers and Amazon Bedrock AgentCore, we are able to make a extra dynamic circulate sooner or later.
To be taught extra about Amazon Bedrock, discuss with the Amazon Bedrock Person Information, Amazon Bedrock Workshop: AWS Code Samples, AWS Workshops, and Utilizing generative AI on AWS for numerous content material sorts. For the most recent information on AWS, see What’s New with AWS?
In regards to the authors
Karnika Sharma is a Senior Product Supervisor within the AWS Gross sales, Advertising, and International Providers (SMGS) org, the place she works on empowering the worldwide gross sales group to speed up buyer progress with AWS. She’s enthusiastic about bridging machine studying and AI innovation with real-world influence, constructing options that serve each enterprise targets and broader societal wants. Outdoors of labor, she finds pleasure in plein air sketching, biking, board video games, and touring.
Dayo Oguntoyinbo is a Sr. Knowledge Scientist with the AWS Gross sales, Advertising, and International Providers (SMGS) Group. He helps each AWS inside groups and exterior clients make the most of the facility of AI/ML applied sciences and options. Dayo brings over 12 years of cross-industry expertise. He focuses on reproducible and full-lifecycle AI/ML, together with generative AI options, with a concentrate on delivering measurable enterprise impacts. He has MSc. (Tech) in Communication Engineering. Dayo is enthusiastic about advancing generative AI/ML applied sciences to drive real-world influence.
Mihir Gadgil is a Senior Knowledge Engineer within the AWS Gross sales, Advertising, and International Providers (SMGS) org, specializing in enterprise-scale information options and generative AI functions. With 9+ years of expertise and a Grasp’s in Info Know-how & Administration, he focuses on constructing sturdy information pipelines, advanced information modeling, and ETL/ELT processes. His experience drives enterprise transformation by modern information engineering options, superior analytics capabilities.
Carlos Chinchilla is a Options Architect at Amazon Net Providers (AWS), the place he works with clients throughout EMEA to implement AI and machine studying options. With a background in telecommunications engineering from the Technical College of Madrid, he focuses on constructing AI-powered functions utilizing each open supply frameworks and AWS companies. His work contains growing AI assistants, machine studying pipelines, and serving to organizations use cloud applied sciences for innovation.
Sofian Hamiti is a expertise chief with over 10 years of expertise constructing AI options, and main high-performing groups to maximise buyer outcomes. He’s passionate in empowering numerous expertise to drive world influence and obtain their profession aspirations.
Sujit Narapareddy, Head of Knowledge & Analytics at AWS International Gross sales, is a expertise chief driving world enterprise transformation. He leads information product and platform groups that energy AWS’s Go-to-Market by AI-augmented analytics and clever automation. With a confirmed monitor document in enterprise options, he has remodeled gross sales productiveness, information governance, and operational excellence. Beforehand at JPMorgan Chase Enterprise Banking, he formed next-generation FinTech capabilities by information innovation.
