Back
Blog
How Human Work Becomes the Source of Context: Accelerating Sales Rep Onboarding at a Fortune 500 Enterprise Software Company
What Granola and Cursor teach us about the intelligence hidden in how we perform everyday work
Reading time: ~6 min
Word count: ~1,100
Workfabric AI™
Focused
Other
Customer 1
Vision sync
12:55PM
Yesterday
Customer 2
Team pictures
12:55PM
@Katri, I uploaded all the pictures from our workshop
Yesterday
Customer 3
Tomorrow's Sync
12:55PM
Can you share a link to the marketing assets?
Customer 4
Coaching workshop
10:12AM
Hey Katri, I know this is last minute, but do you have
Customer 5
Vision sync
12:55PM
We look forward to meeting our fall interns for team pictures!
Tomorrow, 11:00 AM (30m)
RSVP
Customer 6
Fw: Volunteers needed
12:55PM
Hey Alumni! We're looking for volunteers
Customer 7
Fw: Volunteers needed
12:55PM
Hey Alumni! We're looking for volunteers
Honeybee Propo
+1
Upsell Proposal Training Task
Medium Difficulty
Customer
Subject:
RE: Follow-up on Proposal
Draft Response
Sales Rep
Dear Customer,
Thank you for your enquiry, .........................................
Suggested response based on best practices: Acknowledge concern about timeline, to commit the customers shall...............
Use Suggestion
Edit
Favorites
Inbox (elviaatkins@outlook.com
11
Expenses (elviaatkins@outlook.com
2
Folders
Inbox
Drafts
Sent Items
Deleted Items
Junk Email
Archive
Expenses
January Expenses
Add account
Practice with training bots based on real scenarios and proven internal behavior.
2x
Faster time to
first closed deal
Fortune 500 Case Study
New reps closed their first deal in just 3 weeks versus the 7-week company average
ContextFabric built a personalised sales coaching tool from the decision traces of top performers.
By capturing how the best reps actually work - across email, meetings, CRM, and internal tools - ContextFabric transformed sales onboarding with coaching on proven internal behavior.
The hardest decisions in GTM, finance, legal, and operations are rarely driven by what is written down. They are driven by judgment: preferences, risk tolerance, precedent, and experience that almost never get articulated explicitly.
That context shows up through digital interactions at the moment of action:
- What someone includes or removes before committing
- Which signals they trust versus ignore
- Where they pause, hesitate, escalate, or decide not to proceed
- What they open repeatedly versus skim once
- What they compare side by side, copy from, edit out, or rewrite
These actions happen too quickly and at too much granularity to be captured as fields, notes, or summaries. Yet they are where real decision making happens.
These interactions are the raw decision traces of work. Not just final artefacts or database updates, but interface-level micro-actions that reveal intent, confidence, uncertainty, and reasoning. When persisted over time, they form true decision traces that explain not just what happened, but how and why it happened.
This is why tools like Cursor and Granola improve so quickly.
Cursor continuously learns how to generate better code by observing how developers interact with its output in real time. It tracks which suggestions are accepted as-is, which are partially accepted and then edited, which are rejected entirely, what gets deleted or rewritten moments later, how often users undo changes, and how these patterns evolve as developers gain familiarity and confidence. Every accept, modify, reject, and rewrite becomes a live training signal, allowing Cursor to refine its understanding of intent, style, and correctness so each subsequent interaction is measurably better than the last.
Granola does the same for writing. It tracks how drafts evolve into final versions: which sentences are repeatedly rewritten, which sections remain unchanged, where users consistently shorten, soften, or clarify language, what structure gets preserved, and what gets cut before sending. Over time, it learns stylistic judgment, not from instructions, but from behavior.
Single-Surface Learning (Cursor & Granola)
Cursor Workflow (Code Autocomplete)
Granola Workflow (Note Generation)
Code Generation
(Tab Autocomplete)
User Accepts,
Edits, or Rejects
AI Generated Notes
User Makes Stylistics
Choices & In-line Edits
Continuous Live Training Signal
(Refines Future Code Generation)
Continuous Live Training Signal
(Learns Writing Style)
The advantage is not just the model itself. It is continuous exposure to human judgment, captured through granular digital interactions at the moment work is done. Each action leaves a behavioural trace that compounds over time, turning real work into training signal rather than static configuration.
The problem is that most real enterprise work does not happen inside a single surface.
In core operational workflows, work is spread across inboxes, documents, spreadsheets, CRMs, calendars, Slack threads, approvals, side conversations, and even entire teams. No single vendor owns the interface end to end. Integrations expose outcomes and timestamps. They show that something happened, but not how the decision was made.
In a real deployment with a Fortune 500 enterprise software company's sales organization, we observed this gap firsthand. The CRM accurately recorded pipeline stages, call outcomes, approved discounts, and closed-won deals. What it did not capture was how the company actually sold: which materials top performers reviewed before calls, how they framed objections in real conversations, what they followed up on immediately after meetings, which competitive arguments they trusted, and where they slowed down or escalated.
Fragmented Multi-App + Internal Tool Sales Workflow (Decision Logic Between Systems)
Outlook
Gmail
Customer pushes back on pricing after receiving proposal.
Rep reviews similar past deals, discount bands, margin thresholds, and exception policies.
Teams
Slack
Rep messages manager asking whether a discount exception is acceptable.
Call
Slack Huddle
Manager and rep discuss precedent, deal risk, customer importance, and alternatives
Salesforce
Hubspot
Rep updates opportunity stage, discount percentage, and brief notes.
DocuSign
Google Docs
Rep edits pricing and terms and sends updated contract to customer
Internal Pricing/
Deal Desk Tool
Internal Chat
Verbal /
Ad Hoc Discussion
CRM
Contract Tool
Systems of Record
Missing Decision Context
Rep compared three prior deals, ignored one dur to churn risk, and anchored on a similar customer in the same region - none of this comparison is recorded
Manager approved the discount this customer was a lighthouse logo and nearing renewal, not because of deal size-that reasoning is never captured.
Manager considered 15% rejected it as setting a bad precedent for similar accounts, and settled on 10% - the rejected alternatives are lost.
CRM shows ‘10% discount approved’ but not why 10% was safe here and risky elsewhere,
Rep rewrote the termination clause twice, softened language after legal pushback, and removed a liability cap - only the final document remains.
Integration Reality: Integrations only show outcomes:
“Approval Granted,” “Discount Applied”,” and “Contract Sent.”
They do not capture how decisions were evaluated or made.
Result
The organization has no durable record of how pricing decisions were made, what precedent mattered, or how top performers exercised judgement.
{
By capturing granular digital interactions across email, meetings, documents, CRM usage, and internal tools, we were able to reconstruct true decision traces for the company’s best sales reps. Those traces revealed not just what the top performers did on calls, but how they prepared, how they navigated objections in the moment, and how they executed follow-up afterward. That behavioural pattern became the foundation of a personalised sales onboarding and coaching tool grounded in how this specific company actually sold, not in generic sales theory.
New reps were onboarded against real precedent. Instead of static playbooks and generic role-play, the system coached them using decision traces extracted from the company’s highest-performing sellers. Reps could practice conversations modelled on real calls, see how their preparation and follow-through compared to top performers, and receive guidance tied directly to proven internal behaviour.
The ROI was immediate and measurable. Ramp-up time for new reps was reduced significantly, and new hires hit their initial sales targets nearly twice as fast as prior cohorts. Managers spent less time correcting basic execution and more time on strategic deals. Deal quality improved because reps adopted proven patterns earlier, leading to fewer escalations and more consistent outcomes.
This outcome was not driven by better scripts or more integrations. It came from capturing how decisions were actually made and turning real work into durable training signal. Instead of documenting outcomes, the organisation learned from behaviour, and that learning compounded with every deal.
It’s precisely because this decision-making lives outside any single system that many upstart AI companies take a different approach. Some attempt to solve the problem by building broad, end-to-end systems of action so all workflows through their own interfaces, allowing them to directly observe and capture the digital interactions that reveal how decisions are made. This is an enormous lift for organisations with entrenched tools and heterogeneous workflows, and it requires shipping interfaces faster than users adopt external ones.
What this approach often misses is a simpler truth.
Humans are the real source of context,
not systems.
Rather than trying to radically change or consolidate tools just to capture interactions, the right approach is to observe digital behaviour across the entirety of a person’s digital footprint as it exists today, shaped by the processes, tools, and ways of working humans
already use.
That requires capturing context during execution by observing digital interactions one level below the application itself, across all apps, systems, and devices. Done this way, context is learned directly from real work without relying on vendor integrations, while causality and human judgment are preserved in full fidelity.
APIs show outcomes. Digital interactions show intelligence.
And that difference is everything.
© Workfabric AI
