Sales augmentation agent

The challenge

A regional distributor’s reps were sitting on years of order history they could not use. Every account’s record held clear signals: the customer who orders fittings every month but never the matching sealant, the account whose volume on one line suggests an obvious adjacent line, the buyer whose reorder cadence just slipped. None of that was visible at the moment it mattered, when the rep was on the phone or standing in front of the customer. The data lived in reports nobody opened mid-call, so reps sold what they always sold and upsell opportunities expired quietly.

The brief was direct: get those signals in front of the rep at the moment of contact, without adding a new tool to their day.

What we built

We built a co-pilot engine that sits behind the distributor’s CRM and pushes ranked suggestions into the screens reps already work in.

Feature builder

The engine syncs three sources from the CRM and surrounding systems: order history down to line items and purchase frequency, the product catalog with SKU relationships and affinities, and account signals such as recent activity, open tickets, and commercial terms. The feature builder turns these into a per-account signal set, a compact picture of what each customer buys, what similar customers buy, and what is changing.

Opportunity scorer

The scorer evaluates each account’s signal set against the catalog and scores candidate offers for upsell and cross-sell fit. It asks, for every plausible product, how strong the evidence is that this account should be buying it and is not.

Suggestion ranker

Scoring produces more candidates than any rep can act on, so a ranker cuts the list to the top few offers per account: a short, defensible list rather than a wall of maybes, which is what keeps the panel trusted instead of ignored.

Suggestion panel, inside the CRM

The ranked offers surface in a suggestion panel embedded in the account view, on screen during the call or visit. This placement was a deliberate design decision. A separate dashboard would have meant asking reps to leave their workflow, check a second system, and carry findings back by hand; in practice such dashboards go unread and behavior never changes. Inside the CRM, the suggestion is simply there when the rep opens the account, at the exact moment a recommendation can become a line item.

Feedback that retrains the ranking

Every suggestion ends in a rep decision: accept, reject, or defer. Those outcomes flow back into the engine and retrain the ranking. This loop is what keeps the system honest. Reps know things the data does not, like which customer is price-sensitive this quarter, and their accept and reject decisions encode that knowledge. Over time the rankings drift toward what actually sells in this territory rather than what looked plausible in the initial model. Without the loop, suggestion quality would be frozen at launch; with it, the co-pilot improves as a side effect of reps doing their jobs.

How it was delivered

We started with the data, building the CRM sync and feature pipeline first and validating signal sets against accounts the sales team knew well. The scorer and ranker came next, tested offline against historical orders before any rep saw a suggestion. We then embedded the panel with a small group of reps, watched how they used it on real calls, and tightened suggestion volume and presentation based on what they accepted, rejected, and ignored. Once the feedback loop was live and demonstrably improving rankings, the rollout extended across the sales team.

What shipped

• A co-pilot engine syncing order history, product catalog, and account signals into per-account feature sets
• An opportunity scorer and suggestion ranker producing top-N upsell and cross-sell offers per account
• A suggestion panel embedded in the CRM account view, visible at the moment of contact
• A feedback pipeline capturing accept, reject, and defer decisions and retraining the ranking
• Measured results: a 12% increase in average order size and a 15% shorter sales cycle

The distributor got no new system to learn. Reps got better answers inside the one they had, and the numbers followed.