DealTrace

01 / The Problem

The patterns were already there.

Every sales deal has its own circumstances. Different customers, different concerns, different reasons for moving forward or walking away.

Still, certain behaviors repeat.

A prospect raises a technical objection. A competitor enters the conversation. A proposal goes unanswered. A previously hesitant customer suddenly becomes interested.

Individually, these events tell us something about a particular deal. Across hundreds of conversations, they may reveal something more useful: patterns associated with winning and losing.

The trouble is recognizing them.

Sales Engineers accumulate experience through their conversations, but the patterns behind those experiences aren’t always easy to surface or examine systematically.

I wanted to make those patterns visible and find a way to apply them while a deal was still in progress.

02 / Finding the Idea

Start with a problem you understand.

When I joined the Sales AI Challenge, I didn’t have a particular project in mind.

Rather than begin with an AI feature and search for somewhere to put it, I looked at the people I already built software for.

Much of my work at Sophos involves developing tools for Sales Engineers. I was familiar with their workflows and the applications they used.

Outlook seemed a natural place to start.

Sales conversations were already taking place there. If those conversations contained useful patterns, why should someone need to leave their inbox to find them?

I began exploring an Outlook add-in that could examine conversations, compare them with historical deal outcomes, and provide practical recommendations.

That idea became DealTrace.

03 / Recognizing the Patterns

From conversations to predictions.

DealTrace examines sales conversations to identify behavioral signals, including engagement, sentiment, urgency, objections, and changes in intent.

But detecting signals is only the beginning.

The engine compares those observations against patterns found in historical winning, losing, and stalled deals. Those patterns provide context for evaluating an ongoing opportunity.

The analysis follows a progression:

1. Conversation analysis
2. Signal recognition
3. Historical win/loss patterns
4. Predictive assessment
5. Recommended action

DealTrace uses an AI-based scoring process informed by those historical patterns to assess deal health on a scale of 0 to 100.

It also identifies factors that support or threaten the deal.

A score alone would tell a Sales Engineer very little. I wanted the application to explain the assessment well enough for someone to decide what to do with it.

04 / Bringing It Into Outlook

Keep the analysis close to the conversation.

I built DealTrace as an Outlook add-in rather than a separate dashboard.

The interface allows Sales Engineers to examine deal health alongside the email conversation they’re working on.

It brings together the health score, detected signals, potential risks, and recommended actions. The application can also suggest a reply based on the conversation and its assessment.

The design needed to make several kinds of information understandable within the limited space of an Outlook task pane.

I organized the experience around a simple progression: understand the deal’s condition, examine the reasons, and consider the next action.

The recommendations are there to support judgment, not replace it.

05 / Learning From Completed Deals

The outcome is part of the evidence.

A completed deal has something an ongoing conversation doesn’t: an outcome.

DealTrace uses historical outcomes to examine which behaviors appear across successful and unsuccessful opportunities.

I extended the project beyond live deal assessment with two reporting capabilities.

Deal Case Studies

Deal Case Studies examine individual opportunities, their important turning points, and the patterns that may have contributed to their outcomes.

The Rep Playbook

The Rep Playbook examines patterns across multiple deals, bringing recurring findings and recommendations into a broader view.

Both support PowerPoint report generation, allowing Sales Engineers to review and share the findings.

As completed deals are incorporated into the historical dataset, their outcomes can inform subsequent analysis and playbook generation.

The objective is to make past experience more useful to the next conversation.

06 / Engineering the Experience

Making AI output useful, not merely impressive.

DealTrace combines Next.js and TypeScript with Microsoft Graph, Office.js, Microsoft authentication, and AI inference through Groq.

I structured the analysis into separate stages for recognizing conversation signals and generating a contextual deal assessment.

Zod helps validate structured data, while Zustand and TanStack Query support application state and asynchronous operations.

One of the more important design and engineering challenges was handling the uncertainty of AI-generated assessments.

The application needed to present a recommendation without making it appear more authoritative than the evidence supporting it.

That meant treating explanations, supporting signals, and risk factors as important parts of the experience, rather than concentrating solely on the numerical score.

The project also gave me my first experience developing an Outlook add-in, which introduced its own integration and interface constraints.

07 / The Outcome

An idea that won the challenge.

DealTrace won the Sophos Sales AI Challenge.

The recognition was gratifying, particularly because the project began without a predetermined idea and involved several technologies I hadn’t previously worked with in this combination.

I had to learn the Outlook integration model, design an AI-assisted analysis workflow, develop the interface, and bring the different components together.

There are still questions worth investigating, particularly the accuracy of predictive assessments and how well the historical patterns generalize across larger datasets.

Winning the challenge doesn’t settle those questions.

What it demonstrated was that an observation about an existing workflow could become a working application with a purpose beyond showcasing AI.

That’s the part of the experience I value most.