The gallery's pain points are more specific than you think
There's a gallery in 798, Beijing. 15 people, focused on contemporary art and limited edition prints. The owner, Mr. Chen, has been at it for 12 years.
Their pain points are concrete: every piece that comes in needs an authentication report — era, material, style, market valuation, comparable sales records. One report takes two people two days. When collectors visit, sales reps match artworks from memory and experience. Hit rate isn't great.
Chen tried AI before. Last year, he had ChatGPT write authentication reports. The output read like encyclopedia entries — unusable. Valuations were unreliable, style descriptions too generic, collectors shook their heads.
This June, GPT-5.6 launched, a capability jump. Chen's tech consultant suggested retrying — not AI solo this time, but human + AI collaboration.
Authentication reports: AI drafts, human reviews
The workflow changed. Sales reps feed artwork photos, dimensions, and signature info into ChatGPT, along with 3-5 comparable sales record links. GPT-5.6 produces a structurally complete authentication draft in 20 minutes — era assessment, material analysis, style classification, market valuation range, horizontal comparison of comparable sales.
Then a senior appraiser spends 30 minutes reviewing and editing, focusing on era and valuation. One report: from 2 days to 2 hours.
The key isn't that AI writes fast — it's that AI can process dozens of sales records for cross-comparison simultaneously. A human doing this would flip through databases all day. And GPT-5.6's valuations are far more reliable than before, because it understands how variables like size, condition, and provenance affect price.
Collector matching: from memory to data-driven
Before, when collectors visited, sales reps relied on mental notes: "this one likes abstract expressionism," "that one collects prints." Now each collector's preference profile lives in ChatGPT — past collection records, budget range, style preferences, even artworks they've browsed.
When new works arrive, AI auto-recommends the top 5 best-matched collectors. Sales reps make targeted invitations. Match accuracy improved 40% over pure experience.
Sounds simple, but there's a barrier. Collector data must be cleaned first — Chen spent two weeks digitizing 5 years of collector archives, including purchase records, communication preferences, and even collection directions collectors mentioned. This data gets anonymized before feeding to AI, protecting collector privacy.
Two pitfalls hit
Pitfall 1: AI valuation almost caused embarrassment. For a young artist's piece, AI referenced auction records and valued it at 150-200K RMB. The gallery's actual price was 80K. The problem: AI mixed different-sized works in comparison — a 1m×1m painting and a 60cm×80cm painting obviously differ in price. They added a size-normalization rule, having AI standardize to per-square-meter pricing before comparison. Problem solved.
Pitfall 2: The authority problem. Collectors seeing AI-generated reports had doubts — "A machine wrote this, right?" Especially older collectors who felt machine authentication lacked a human touch. The gallery later added appraiser signatures and review records, labeling reports "AI-assisted analysis + human review," and kept a personal commentary paragraph from the appraiser. Trust issues eased.
An unexpected bonus
Midjourney V8.1 was used to generate artworks displayed in different spaces — living rooms, offices, hotel lobbies. Collectors' biggest hesitation when buying art is "will it look good in my home?" Previously, they could only imagine. Now the gallery provides 3-5 spatial display images for each key piece, and closing rates rose noticeably.
Chen said this feature was an unexpected gain. Originally just a sales communication tool, it became a deal catalyst.
Numbers and costs
- Closing rate: 12% → 25% (better matching, higher conversion)
- Authentication reports: 2 days → 2 hours per report
- Client match accuracy: ↑40%
- Labor cost: ↓30% (no layoffs — freed-up staff shifted to collector relationship management)
Tool costs: ChatGPT Team $25/user/mo × 5 = $125/mo. Midjourney for generating artwork display scenes, $30/mo. Total $155/mo, about 1,100 RMB.
No layoffs. Chen said the freed-up time goes to higher-value work — visiting artist studios, maintaining collector relationships, curating exhibitions. AI can't do those.
One more change: previously, sales reps hesitated to push lesser-known artists' work, because matching collectors relied purely on experience — get it wrong and it's awkward. Now AI provides data backing — "this collector has acquired 3 similar-style pieces before." Sales reps have confidence. Result: closing rates for lesser-known artists rose too, and the gallery is willing to sign more emerging artists.
Galleries are a trust-and-taste business. AI accelerates information processing, but the final call is still human. Chen's words: "AI saves me 80% of desk work. But that 20% of judgment — whether this piece is worth signing, whether this collector is worth deep investment — that's still on me."