From manual data chaos to Salesforce-ready in minutes
How Navikenz built a six-stage agentic AI pipeline that eliminated ~80% of manual CRM data preparation for the world’s leading events services partner.
|
~80% reduction in manual data prep |
|
Zero duplicate records at ingestion |
|
10× faster processing (723-record batch) |
|
First first agentic AI app in EMEA |
THE CLIENT
A Global Events Services Leader
The client is the world’s leading global events services partner, supporting hundreds of tradeshows, exhibitions, and live events across the UK and international markets. Its Salesforce CRM estate is the operational backbone for account management, sales reporting, and territory planning — a system that must remain accurate as exhibitor volumes grow, show schedules accelerate, and data arrives from an increasingly distributed network of field teams, partners, event organisers, and third-party providers.
The sheer scale and heterogeneity of incoming exhibitor data had outpaced the firm’s ability to process it reliably. The client engaged Navikenz to design and deliver a production-grade, AI-driven solution capable of replacing a largely manual and error-prone data preparation workflow with an intelligent, automated pipeline.
THE CHALLENGE
A data intake process that could not keep pace
Each exhibition cycle generated a new wave of exhibitor data files — Excel workbooks of varying structure, quality, and provenance. Before a single record could be loaded into Salesforce, data stewards and account managers faced a time-consuming and error-prone remediation process: hunting for missing mandatory fields, correcting zone misassignments, resolving non-standard country codes, and reformatting values to comply with Salesforce’s field specifications. The work was repetitive, skilled, and entirely manual.
The consequences were predictable. Transcription errors propagated directly into the CRM. Service zone and operations zone assignments were frequently mismatched, degrading the accuracy of downstream territory planning and sales routing. Uploaded records arrived without the firmographic attributes — industry classification, company size, verified address — that account teams depended on. And without deduplication, bulk uploads periodically seeded Salesforce with duplicate account and contact records, generating reporting conflicts that were costly to unwind.
A further complication: several source files arrived password-protected to satisfy data governance requirements, making them incompatible with any existing automated processing approach. The combination of volume, variability, and security constraints had created a bottleneck that no incremental improvement to the existing manual process could resolve.
OUR APPROACH
A two-phase programme: clean and enrich, then automate the upload
Navikenz approached the engagement in two sequential phases, each building on the last. Phase 1 established the core data cleansing and enrichment capability; Phase 2 completed the automation chain by eliminating the remaining manual handoff between prepared data and Salesforce.
Phase 1 — The Agentic Data Preparation Pipeline
We designed and delivered the SFDC Agentic AI Upload Backend: a FastAPI microservice wrapping a six-stage AI agent orchestration pipeline, exposed via REST API and underpinned by Azure Blob Storage for file and job-state management. The service accepts single or multi-file exhibitor uploads in .xlsx and .xls format, processes them asynchronously in background jobs, and returns Salesforce-ready output files.
The pipeline’s six stages were sequenced to address each failure mode in the existing process. An LLM-powered preprocessing agent (Azure OpenAI GPT-4o-mini) ingested raw data in configurable row batches, standardising field values, correcting zone assignments, resolving stand-type ambiguities, and flagging residual quality issues — all while preserving every original column. A second enrichment agent filled gaps in firmographic data, drawing on the model’s knowledge base and live Serper web lookup to populate industry classifications, contact details, and company attributes. Deterministic validation then enforced Salesforce’s field-length constraints across every column. A deduplication step removed redundant records before handoff. Dun & Bradstreet and Serper lookups provided authoritative firmographic verification. Finally, a country-code normalisation stage resolved free-text country entries to ISO-standard values compatible with Salesforce picklists.
The architecture was built for scale from the outset. Password-protected Excel files — previously an outright processing blocker — were decrypted transparently using msoffcrypto-tool ahead of pipeline entry, and optionally re-encrypted on output to preserve governance controls. The asynchronous job model allowed concurrent multi-file uploads to return immediately with job IDs, with teams polling for results against a blob-backed job store. An auto-cleanup scheduler managed stale job data on a configurable cadence.
The Phase 1 data architecture followed a Microsoft Fabric / OneLake medallion pattern. Files entered a Landing Zone before a Preprocessor Agent ingested them into the Bronze Layer for null-value and pre-enrichment checks. An Enrichment Agent then elevated records to the Silver Layer, applying DNB and Loqate remote services to verify and enrich addresses and firmographic values. Validated data flowed to an Event Data Lake, where legacy and incoming records were reconciled before the final Salesforce-ready Excel output was produced.
Phase 2 — End-to-End Salesforce Upload Automation
Where Phase 1 produced clean, Salesforce-ready files that still required a manual upload step, Phase 2 eliminated that final handoff entirely. A dedicated Salesforce Upload Automation layer — comprising a React.js frontend, a FastAPI backend deployed on Azure App Service, and a six-step data processing pipeline integrated with Salesforce’s Bulk API v2 — closed the loop from exhibitor file submission to confirmed Salesforce record creation.
The processing pipeline transformed uploaded files through conversion, encoding normalisation, schema transformation, field cleaning, header and country mapping, and line-ending standardisation before submitting records to Salesforce via an OAuth 2.0-authenticated Bulk API v2 job lifecycle. A custom Salesforce staging object, Booth_Staging__c, held the full exhibitor field set pending ingestion. Supporting Apex automation — batch classes, a deduplication handler, and triggers — managed post-ingestion processing. On job completion, the system produced four structured output reports: a performance summary, successful records, failed records, and a retry file for any rows requiring reprocessing.
THE OUTCOMES
Measurable gains across speed, quality, and operational scale
The combined Phase 1 and Phase 2 delivery transformed the client's exhibitor data onboarding from a manual, error-prone operation into a fully automated, scalable capability.
The automated six-stage pipeline removed approximately 80 percent of manual data preparation effort, freeing data stewards and account managers from field-by-field Excel review and redirecting their time toward higher-value activities. Deduplication at both the pipeline and Salesforce ingestion layers ensured that no duplicate accounts or contacts entered the CRM through the bulk upload channel — eliminating a source of reporting and assignment conflicts that had previously required manual remediation.
Phase 2 delivered substantial throughput improvements that scaled with file size. A batch of 100 records processed 1.6 times faster than the prior approach; at 723 records, the same pipeline ran more than ten times faster, reducing processing time from nearly five minutes to under thirty seconds. A batch of 1,107 records completed in just over one minute, against a previous baseline of nearly seven. These gains reflect both the efficiency of the automated pipeline and the elimination of the manual upload step that had previously formed a hard constraint on throughput.
Password-protected files — previously excluded from any automated workflow — are now handled transparently, enabling the client to maintain its data governance controls without sacrificing automation. CRM data quality improved materially: zone assignments, country codes, and firmographic attributes are now consistently populated before ingestion, reducing post-load correction cycles.
The solution also carries strategic significance as the first deployed agentic AI application for the client’s EMEA operations, establishing a proven architecture and delivery pattern that the business can extend to adjacent data intake challenges.
WHY IT WORKED
Deterministic where it had to be. Agentic where it paid
The programme succeeded because Navikenz treated the problem as a pipeline engineering challenge, not a transcription exercise. Rather than attempting to build a single model capable of resolving every data quality issue, the team decomposed the workflow into discrete stages, applying LLM-based reasoning precisely where variability and ambiguity required it — field standardisation, zone resolution, firmographic enrichment — and deterministic logic where consistency and auditability were paramount: length validation, deduplication, country-code normalisation.
This hybrid architecture delivered the best of both approaches: the flexibility of AI reasoning for the messy, context-dependent work of data interpretation, and the reliability of rules-based processing for the structured, compliance-sensitive steps that Salesforce’s ingestion layer demanded. Every artifact — landing zone input, pre-processed file, enriched output — was persisted to Azure Blob Storage, providing a complete audit trail for each job.
Equally important was the decision to build for operational reality from the outset. Password-protected files, concurrent multi-file uploads, large batch sizes, and the need to re-encrypt output for governance compliance were all treated as first-class requirements, not edge cases. The result is a system that the client’s teams can rely on at production scale, with the confidence that the architecture will hold as exhibitor volumes continue to grow.
Have a CRM intake bottleneck like this? We would like to help.
info@navikenz.com · navikenz.com