Building AI into the tools professional sellers rely on

1,700 sellers pricing blind on a platform that had the data to help them

Back Market is Europe's largest marketplace for refurbished electronics, powered by 1,700+ professional resellers across Europe and North America. These aren't casual listers — they're businesses managing thousands of listings, repricing daily, and processing high volumes of orders. Their primary tool is the Seller Back Office: a web platform for pricing, inventory, orders, logistics, and performance analytics. It's open every working day.

Why the problem required AI

Sellers price using a mental formula: sourcing cost + operating costs + return risk + target margin. The problem is that several variables determining whether that price is competitive were either unavailable or required guesswork:

Back Market already had the data to answer these questions. The problem wasn't a data gap. It was a product gap: the intelligence existed but none of it was reaching sellers in a form they could act on.

The single design question that shaped every feature

I mapped the decisions sellers make daily, the information each required, and what Back Market could credibly provide. For each opportunity, the same question emerged: should we show the intelligence and let sellers act, or should the system act on their behalf? It turned out to be the most consequential design decision in the programme.

The stakes were unusually high. Back Market charges the highest commission rates in the market, and sellers assume the platform defaults to the customer. Any AI acting without explicit consent would be interpreted as the platform acting in its own interest.

The Automation Spectrum

To navigate this, I proposed a four-level spectrum — from passive data surfacing to full automation — and mapped every AI feature onto it. The spectrum became the governing framework for the entire programme: a feature can only occupy a level of automation proportionate to the trust the programme has earned with the audience it's serving.

Each feature answered the same design question at a different level of automation

• Level 1AI Support Assistant

  A floating chat panel where sellers ask product, policy, and operational questions in natural language. I designed the seller-facing UI for production and am now leading the 2026 evolution to task execution.

  Live · All markets · Q1 2024
• Level 2Sales Maximizer Price (SMP)

  An ML-powered price recommendation shown inline on the product page with an estimated sales forecast and a single button to apply it. I designed the recommendation card, defined the outcome-framing pattern, and iterated through three major versions.

  Live · 2,585 products covered · V3 · Q3 → Q4 2024
• Level 3BackPricer

  Sellers set a price floor and ceiling; the system reprices continuously within those bounds. I proposed the floor/ceiling model after observing that sellers think in limits, not optimums, and designed the setup flow to match.

  Live · All markets · Q4 2024
• Levels 2–3Pricing Opportunities

  Three categories of pricing actions (Quick Wins, Listings in Trouble, BackBoxes to Boost) with recommended prices and bulk-action controls. I defined the category taxonomy and worked with data science to set the thresholds.

  Live · 3 opportunity types · Q2 2025
• Level 4Dynamic Pricing Up

  An ML model that adjusts prices upward where demand can absorb it. Optimises platform revenue, not seller margin. I led the design response to the trust damage, including the transparency UI and the consent model now in build.

  In experiment · Q4 2025 · Enabler 2 in build

How a recommendation evolves

SMP went through three major versions. Each iteration moved the framing further from the algorithm and closer to the seller's mental model.

Design iteration

How the SMP recommendation evolved across three versions

V1 — Launch

Model-centric framing

Showed recommended price + model confidence score. Sellers didn't understand confidence percentages — "73% confident" felt uncertain rather than reassuring.

Problem: Low engagement, high confusion

→

V2 — Outcome framing

Sales forecast anchoring

Replaced confidence scores with outcome forecasts: "at €X, we estimate 12–18 sales in 7 days." Engagement improved, but sellers wanted to compare against their current price.

Progress: Better comprehension, still passive

→

V3 — Current

Comparative + one-click apply

Added current-price comparison, delta indicator, and a single "Apply price" button. Framing shifted from "here's a suggestion" to "here's what changes if you act."

Result: Highest engagement version

Key learning: Sellers respond to consequence framing ("what changes") over model framing ("how confident"). Each iteration moved further from the algorithm and closer to the seller's mental model.

Where AI lives in the product

AI currently surfaces across five distinct locations in the Back Office. Each was designed independently, but the programme is moving toward a coherent architecture where intelligence is present throughout the seller workflow.

Working with the ML team

Every feature required close collaboration with Back Market's data science team. The hardest negotiation was the cold-start problem: I pushed for a minimum 90-day accuracy window before surfacing track record data in the UI. The ML team resisted because it delayed a key engagement metric. We compromised on surfacing provisional accuracy earlier with explicit "early data" labelling. It was the right call. Showing unqualified numbers early would have undermined the trust the programme was built on.

Four requirements for trust

Research surfaced four trust requirements. Each one shaped specific design decisions across the product.

• Seller control — 51.8%

  BackPricer's adoption is higher than SMP's, not because it's simpler, but because sellers configure it themselves. Automation outside seller-defined parameters is experienced as the platform acting in its own interest.
• Transparent explanation — 47.3%

  Explain in seller terms, anchored to consequences. Not "coefficient of 0.73" but "at this price, you're in the top 10% of sales velocity." Full formula explanations increased confusion and reduced confidence.
• Proven accuracy — 41.1%

  After 90 days, we surface accuracy in the UI: "73% of sellers who applied our recommendation sold more units within 7 days." Early adoption is infrastructure, not a success metric.

What it actually takes to get AI right

Every AI initiative in 2024 had pressure behind it — pressure to ship features with "AI" in the name, show progress to the board, keep pace with competitors. We pushed back on that. The question was never how do we add AI to the Back Office — it was where does AI actually make a seller's job easier, and how do we earn the right to put it there? That framing shaped every decision in this programme. The data below shows where it worked and where we still fell short.

The consistent finding: sellers express a coherent form of partial trust — they'll take the data, but not the recommendation. They're not disengaged; they're careful. The programme's job isn't to replace seller judgment — it's to give sellers the information to exercise it more confidently.

These findings pointed to a single conclusion: sellers didn't reject AI — they rejected AI that hadn't earned their trust yet. The order we introduced features mattered as much as the features themselves.

What we'd do differently

The accountability we wanted required getting the order right. Each feature either builds credibility or creates suspicion that every subsequent feature inherits. Shipping AI that works isn't a technology problem — it's a planning problem.

If we were starting again — with the luxury of sequencing the programme around trust instead of around shipping deadlines — we'd move through four stages:

• 1
  Stage 1: Surface data, make no recommendation

  Before any AI recommendation, give sellers the raw data Back Market has that they don't: competitor pricing on their exact products, sales velocity rank, demand trends. No suggestion about what to do with it. Just visibility they've never had before. This proves the data is accurate before you attach a recommendation to it. Minimum duration: 2–3 months.

  Automation Spectrum: Level 1
• 2
  Stage 2: Single-product recommendation (SMP)

  A price recommendation on one product at a time. Recommendation only — seller always decides. Framed around an outcome forecast: suggested price + estimated units sold in 7 days. This is the first moment sellers are asked to consider an AI recommendation. It needs to be right more often than it's wrong, for long enough to build the accuracy track record that 41% of sellers require. Minimum duration: 3–6 months before BackPricer.

  Automation Spectrum: Level 2
• 3
  Stage 3: Automation with consent (BackPricer)

  Sellers configure price bounds; the system reprices continuously within those bounds. Sellers have already experienced the pricing model as a recommendation — they've seen it work, they understand how it thinks. BackPricer asks them to extend that trust into automation, with the safeguard that the system can never act outside their defined parameters.

  Automation Spectrum: Level 3
• 4
  Stage 4: Catalogue-level intelligence (Pricing Opportunities)

  AI-surfaced actions across the seller's entire catalogue — Quick Wins, Listings in Trouble, BackBoxes to Boost — with bulk action controls. This requires sellers to trust the intelligence across their entire catalogue simultaneously. That trust is only possible after 6–12 months of experience with single-product recommendations and automated repricing.

  Automation Spectrum: Levels 2–3

The actual programme deviated under shipping pressure. BackPricer launched with a limited SMP track record. Dynamic Pricing Up went to Level 4 before the programme had earned it — exactly the kind of "ship AI because we can" decision we'd set out to avoid. The trust damage cast doubt on every subsequent AI feature.

Where this goes next

This is Phase 1 of a multi-year programme. The roadmap operates across three horizons:

• Near-term — closing the €15 gap

  The €15 gap between AI-recommended and actual seller prices is the 2026 North Star metric. Focus: improving recommendation legibility, surfacing accuracy track records, and evolving the chat assistant from Q&A to task execution.
• Medium-term — system-level intelligence

  Moving from isolated features to a coherent AI layer: proactive anomaly detection on the dashboard, AI-assisted pricing at listing creation, and demand forecasting in inventory management.
• Long-term — the transferable framework

  The Automation Spectrum, trust requirements, and sequencing model aren't Back Market-specific. They apply to any professional platform where users have structural reasons to be cautious about automation.