At AWS, our gross sales groups create customer-focused paperwork referred to as account plans to deeply perceive every AWS buyer’s distinctive targets and challenges, serving to account groups present tailor-made steerage and assist that accelerates buyer success on AWS. As our enterprise has expanded, the account planning course of has change into extra intricate, requiring detailed evaluation, evaluations, and cross-team alignment to ship significant worth to prospects. This complexity, mixed with the handbook overview 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 lowering the general time to finish, overview, and approve plans by 52%.
On this submit, we share how we constructed Pulse utilizing Amazon Bedrock to cut back overview time and supply actionable account plan summaries for ease of collaboration and consumption, serving to AWS gross sales groups higher serve our prospects. Amazon Bedrock is a complete, safe, and versatile service for constructing generative AI purposes 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 data bases to attach purposes to your newest knowledge so that you’ve got every thing you might want to rapidly 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 via 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 essential challenges had emerged:
- Disparate plan high quality and format – With groups working throughout quite a few AWS Areas and serving prospects in various industries, account plans naturally developed variations in construction, element, and format. This inconsistency made it troublesome to ensure essential buyer wants have been described successfully and persistently. 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 overview course of – The standard evaluation course of relied on handbook evaluations by gross sales management. Although thorough, these evaluations consumed invaluable time that would in any other case be dedicated to strategic buyer engagements. As our enterprise scaled, this strategy created bottlenecks in plan approval and implementation.
- Information silos – We recognized untapped potential for cross-team collaboration. Growing strategies to extract and share data would rework particular person account plans into collective finest practices to higher serve our prospects.
Answer overview
To deal with these challenges, we designed Pulse, a generative AI resolution 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 via an asynchronous batch processing pipeline.
- The info flows via a sequence of processing levels:
- Preprocessing to construction and normalize the information and generate metadata.
- LLM inference to research content material and generate insights.
- Validation to substantiate high quality and compliance.
- Outcomes are saved securely for reporting and dashboard visualization.
We’ve built-in Pulse immediately with current gross sales workflows to maximise person 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 elements 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 a daily cadence, facilitating steady evaluation of essentially the most present info.
Preprocessing
Contemplating the dynamic nature of account plans, they’re processed in day by day snapshots, with solely up to date plans included in every run. Preprocessing is performed at two layers: an extract, rework, and cargo (ETL) movement layer to arrange required information to be processed, and simply earlier than mannequin calls as a part of enter validation. This strategy, 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 day by day scheduled job that reads account plan knowledge 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 lately modified plans, considerably lowering computational overhead and prices. The processed textual content content material and metadata are then reworked 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 resolution makes use of Amazon Bedrock, which gives quite a lot of mannequin decisions and management, knowledge customization, security and guardrails, price optimization, and orchestration. We use the Amazon Bedrock FMs to carry out two key capabilities:
- 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 which may in any other case stay remoted in particular person paperwork.
We implement these capabilities via asynchronous batch processing, the place analysis and summarization workloads function independently. The analysis course of runs every account via 27 particular questions with tailor-made management prompts, and the summarization course of generates topical overviews for easy consumption and data sharing.
For this implementation, we use structured output prompting with schema constraints to supply constant formatting that integrates with our reporting instruments.
Validation
Our validation framework consists of the next elements:
- Enter and output validations are essential as a part of the OWASP Top 10 for Large Language Model Applications. 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 overview for outputs that don’t meet high quality thresholds.
- A suggestions mechanism that improves system accuracy over time.
Storage and visualization
The answer consists of the next storage and visualization elements:
- Amazon S3 gives safe storage for all processed account plans and insights.
- A day by day run cadence refreshes perception and permits 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 carried out a sturdy engineering framework to deal with three essential AI-specific challenges. First, the non-deterministic nature of LLMs meant similar inputs may produce various outputs, doubtlessly compromising analysis consistency. Second, account plans naturally evolve all year long with buyer relationships, making static analysis strategies inadequate. Third, totally different AWS groups prioritize totally different features 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 purpose is to make use of the CoV as a correction issue to deal with the information dispersion, which we achieved by calculating the general CoV on the evaluated query degree. With this strategy, we are able to scientifically measure and stabilize output variability, set up clear thresholds for selective handbook evaluations, 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 handbook overview. We complemented this with a dynamic threshold weighting system that aligns evaluations with organizational priorities by assigning totally different weights to standards based mostly on enterprise impression. This customizes thresholds throughout totally different account varieties—for instance, making use of totally different analysis parameters to enterprise accounts versus mid-market accounts. These enterprise thresholds endure periodic overview with gross sales management and adjustment based mostly on suggestions, so our AI evaluations stay related whereas sustaining high quality and saving invaluable time.
Conclusion
On this submit, we shared how Pulse, powered by Amazon Bedrock, has reworked the account planning course of for AWS gross sales groups. Via automated evaluations and structured validation, Pulse streamlines high quality assessments and breaks down data silos by surfacing actionable buyer intelligence throughout our world group. This helps our gross sales groups spend much less time on evaluations and extra time making data-driven selections for strategic buyer engagements.
Wanting forward, we’re excited to reinforce 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 establish 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 strong enhancements to our processes. By constructing flows for orchestrating our workflows, use of Amazon Bedrock Guardrails, introduction of agentic frameworks, and use of Strands Agents and Amazon Bedrock AgentCore, we are able to make a extra dynamic movement sooner or later.
To be taught extra about Amazon Bedrock, seek advice from the Amazon Bedrock Person Information, Amazon Bedrock Workshop: AWS Code Samples, AWS Workshops, and Using generative AI on AWS for diverse content types. For the newest information on AWS, see What’s New with AWS?
Concerning 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 development with AWS. She’s obsessed with bridging machine studying and AI innovation with real-world impression, constructing options that serve each enterprise targets and broader societal wants. Exterior of labor, she finds pleasure in plein air sketching, biking, board video games, and touring.
Dayo Oguntoyinbo is a Sr. Information Scientist with the AWS Gross sales, Advertising, and International Providers (SMGS) Group. He helps each AWS inside groups and exterior prospects benefit from the facility of AI/ML applied sciences and options. Dayo brings over 12 years of cross-industry expertise. He makes a speciality of reproducible and full-lifecycle AI/ML, together with generative AI options, with a give attention to delivering measurable enterprise impacts. He has MSc. (Tech) in Communication Engineering. Dayo is obsessed with advancing generative AI/ML applied sciences to drive real-world impression.
Mihir Gadgil is a Senior Information Engineer within the AWS Gross sales, Advertising, and International Providers (SMGS) org, specializing in enterprise-scale knowledge options and generative AI purposes. With 9+ years of expertise and a Grasp’s in Data Know-how & Administration, he focuses on constructing strong knowledge pipelines, advanced knowledge modeling, and ETL/ELT processes. His experience drives enterprise transformation via revolutionary knowledge engineering options, superior analytics capabilities.
Carlos Chinchilla is a Options Architect at Amazon Net Providers (AWS), the place he works with prospects 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 purposes utilizing each open supply frameworks and AWS companies. His work consists of creating AI assistants, machine studying pipelines, and serving to organizations use cloud applied sciences for innovation.
Sofian Hamiti is a know-how chief with over 10 years of expertise constructing AI options, and main high-performing groups to maximise buyer outcomes. He’s passionate in empowering various expertise to drive world impression and obtain their profession aspirations.
Sujit Narapareddy, Head of Information & Analytics at AWS International Gross sales, is a know-how chief driving world enterprise transformation. He leads knowledge product and platform groups that energy AWS’s Go-to-Market via AI-augmented analytics and clever automation. With a confirmed observe file in enterprise options, he has reworked gross sales productiveness, knowledge governance, and operational excellence. Beforehand at JPMorgan Chase Enterprise Banking, he formed next-generation FinTech capabilities via knowledge innovation.
