The AI-Native Sales Stack
Sales and marketplace software was built to record what happened. The next generation will decide what happens next, and that changes what product managers build.
Most sales software is a very expensive diary. A CRM records the call, the stage change, the discount, the lost reason. Dashboards summarize the diary. Reps spend hours writing in it and very little time being helped by it.
I learned this the hard way building a sales enablement app at Hanging Valley. The data a rep needed was spread across on-prem databases, an ERP, a CRM and a pile of spreadsheets. The biggest win wasn’t a smarter chart. It was getting that data into one place and pushing the one thing that changed, a stalled deal or a price shift, to the rep’s phone before it was too late. Sales cycles got 40% shorter. The win didn’t come from a clever model. It came from no longer making reps do the assembly.
That project convinced me that the next generation of sales and marketplace software won’t be about recording. It will be about deciding.
From system of record to system of action
AI changes three things at once.
- Unstructured data becomes usable. Call transcripts, emails, contracts and support tickets used to be where insight went to die. Language models can read all of it and turn it into structured signals: buying intent, objections, risk, commitments made.
- The interface becomes a conversation. “Which of my accounts are most likely to churn this quarter, and why?” is now a reasonable thing to type, and the answer can cite its sources.
- The software can act, not just advise. Drafting the follow-up, updating the forecast, flagging the risky discount for approval: the machine does the busywork, and the human makes the call.
What this looks like in sales
- A copilot for every rep. Before each meeting: a one-paragraph account brief, the last three conversations summarized, open issues, and a suggested next step. After it, the notes and follow-up are drafted automatically.
- Forecasts built on evidence. Instead of reps guessing a stage, the forecast is based on what customers actually said and did, with the reasoning shown.
- Coaching at scale. Every call becomes coaching material. Managers see patterns across the team instead of sampling one call a week.
What this looks like in marketplaces
Marketplaces are matching engines with payments attached, and AI improves every part of that.
- Matching: understanding what a buyer means, not just the keywords they typed, and what a seller can actually deliver.
- Pricing: dynamic, explainable price recommendations for sellers, based on demand, seasonality and competition.
- Trust: fraud and risk detection that learns, along with automated dispute summaries so support teams resolve cases in minutes. In payments work I’ve seen that layering a model over existing rules, rather than replacing them, cuts false positives while keeping every decision explainable. Trust is a product feature, not a back-office cost.
The winners won’t have the best model. They’ll have the best loop: data in, decision out, and outcomes fed back.
What changes for product managers
This is where I think most teams underestimate the shift.
- Evaluation is the new QA. A feature that’s right 80% of the time is a different product from one that’s right 99% of the time. PMs need test sets, quality thresholds and a clear plan for when the model is wrong.
- Data is the moat, and the plumbing is the product. Every AI feature in sales depends on clean, connected, current data. The unglamorous integration work I did at Hanging Valley is exactly what makes the AI layer possible.
- Inference has unit economics. Every AI call costs money. Choosing the model, caching and routing work between large and small models are now product decisions with a direct effect on margin, not just engineering details.
- Keep humans in the loop where trust matters. At Circadence we kept a human review step in our AI pipeline even though it cost speed, because our buyers needed to trust what went live. In finance and sales, discounts, credit and fraud decisions deserve the same care.
What I’d build
A deal copilot for mid-market B2B sales teams that plugs into the systems they already use rather than replacing them. It would read calls, email and the CRM, and produce three things: a daily prioritized action list for every rep, an evidence-based forecast for every manager, and an early warning when a deal, a customer or a price is drifting.
I’d measure it on one number: hours given back to selling. If it doesn’t create more selling time, it’s just a nicer diary.