Partner Election 2026 · Oral

Planning,
made AI-native.

Phil Zhang · VP Data Science, Artefact
An end-to-end planning offer: machine learning forecasts, orchestrated by agents, deployable on open-weight models.

  1. 01Who I amFrom training models to building agents, in six years
  2. 02My offeringEnd-to-end planning, proven, plus one more thing
  3. 03Business caseThree bets on a €5.3M pipeline
  4. 04Close and Q&AWhat I need from the partner group
01 · Who I am

From training models to building agents, with agents.

2020joined Artefact
150+projects delivered
38clients
€3.6Mrevenue delivered, 2025–26 YTD
2020 – 22

A model you train

Propensity scoresForecastsRecommendations
score 0.82
19projects
11new clients
0%agentic100% ML & data
2023 – 24

A model you prompt

Knowledge botsContent generationInsight bots
What sells in Tier 2?Down 8%, driven by…sources
64projects
14new clients
5%agentic95% ML & data
2025 – 26

An agent that acts

Agents and toolsQuick demosFast iteration
data toolmodelreport
73projects
13new clients
50%agentic50% ML & data
My play
Lead the data science team
Frame the solution for every partner
Cover AI agent solutions and new use cases
Connect the China AI ecosystem and a leading research centre in Shanghai
2020

From 4 clients to 38, from Data Science to Agentic.

0clients
0offerings
0agentic projects
Agentic share of projects sold0%
02 · My offering · End-to-end planning

Plan the whole business as one loop, re-planned every week.

What it is
DemandAssortmentInventorySupplyPrice & promoFinanceOne planre-planned weekly
✕Stock you cannot sell→Buy what will sell
✕Sales lost to empty shelves→Right stock, right store
✕Plans that are stale on day one→Re-plan weekly
What to say, by client role
CEO

“Stop betting the season on a plan made a year ago. Re-plan every week and grow sales on the same stock.”

CFO

“Release the cash locked in unsold stock, without losing a single sale.”

Product & merchandising

“Know which new styles will sell, and how many, before you commit to a single unit.”

Supply chain

“Put the right stock in the right store before the customer asks, not after you run out.”

Retail & marketing

“Promote what sells at full price, and stop clearing the rest with blanket discounts.”

02 · My offering · Why speed matters

Shein re-plans in days. Most brands plan in months.

A typical brandone big bet per season
Plan and design6–12 months ahead
One big buy
Make and ship
Sell the season
Markdown
Shein-style loopa small bet, every week
Test a small batchRead live demandReorder winners, drop losers

One style, two ways to bet

Typical brandShein-style
Units in the first order
~5,000
100–200
Days to produce and restock
30–45
~5
Plan reviews per year
1–2
weekly, ~50

Shein revenue, USD billion (estimates)

3.22019
9.82020
15.72021
22.72022
32.22023

10× in four years, from small bets repeated every week.

To run a loop like this, plan end to end, continuously, automatically
Finance→Demand→Assortment→Inventory→Supply→Price & promo↺
Company-reported or third-party claims: Shein Group (100–200 unit batches, 5-day restock); sourcing-industry comparisons for the typical-brand figures; revenue from press estimates (Coresight, WSJ). Shein is private and publishes no audited figures.
02 · My offering · The planning cycle

Eight steps, three phases. Click a step.

Pre-seasonBuy the right thing · decided 6 to 10 months ahead
In-seasonPut it in the right place · weekly, from launch
After seasonClear what is left
Season actuals feed back into next season's strategy.Independent evidence   Vendor or brand-claimed
02 · My offering · What it delivers

Models and agents turn into four business levers.

Grow sales
+2 to +5%
from the same buy
≈ +€10–25M a year
Protect margin
+5 to +20 pp
full-price sell-through
fewer markdowns, same buy
Release cash
−15 to −30%
inventory and coverage
≈ €15–30M freed
Free people
−35 to −50%
time on data work
≈ 10–15 of 30 planner FTE
powered by, with the evidence↑↑↑↑
Modelspredict and optimise
Accuracy
−20–50%
forecast error vs manual methodsMcKinsey 2022
−40–50%
error on new stylesMyntra, peer-reviewed
15% → 6%
retailer forecast errorbrand case
Flexibility
4 → 12
planning cycles a year (target)Anta
+3.6–5.8%
revenue from localised rangingWharton field tests
Optimisation
+3–4%
sales from constraint-aware orderingZara, Operations Research
+2–4%
sales, with −4% unsold unitsZara experiment
Agentsact, align and explain
Automation
Minutes
vs 2–3 days for a ~100-person team2,700-store client
up to 40%
of merchant time reclaimedMcKinsey 2026
Collaboration
−40%
order cycle with a digital order fairvendor case
−50%
order confirmation timemanufacturer case
Explainability
Audit trail
every quantity explained, every override keptdesign principle, no benchmark yet
Independent: McKinsey, BCG, Zara field work, Wharton, Myntra. Brand-claimed: Guess, Boggi, Anta, retailer and fair cases. Illustration: a brand with €500M sales, €100M stock and 30 planners (my assumption). Ranges give way to the client's own baseline in the diagnostic.
02 · My offering · See it work

A demo you can click through.

demo link: TBD
Click to open the live demo ↗
02 · My offering · Proof

Already proven: a lot delivered across the chain, and a strong pipeline to connect it.

Already delivered

11 clients
One logo per step of the planning chain we have worked on. Connecting the steps is the headroom.
Forecast & demand
Buying & allocation
Pricing & promotion
Financial & sales planning

In the pipeline

~€2.1M GM
Planning and forecasting work, 8 opportunities, 7 clients. PepsiCo is the lead for a fully agent-native, end-to-end planning offer.
AS WatsonPromo forecast, pilot build
€670K
PepsiCoAgent-native company: end-to-end planning, Ph1 + Ph2
€589K
adidasDemand planning, Phase 2
€303K
Ralph LaurenScenario-based decision support
€269K
Walmart ChinaForecasting optimisation
€170K
ANTA · ECCODemand and franchise order forecast
€126K
Draft · multi-count GM · to verify
03 · One more thing

Some clients want agents inside their own walls. We should be ready.

We have done itlululemon China and CIMB Singapore
Plugged into China's AI ecosystemopen-weight labs and chip makers
Close to China's researchuniversities and a leading centre in Shanghai
Let's build it togetheran open-source, on-premise stack, with our global colleagues
Agents
Frameworks
LangGraphGoogle ADKQwen-Agent
Studio
Difyn8nCoze Studio
Harness & tools
HermesOpenCodeOpenClawMCP
Knowledge
RAG
RAGFlowDoclingMilvusQdrant
Knowledge graph
Neo4jLightRAGGraphRAGNebulaGraph
Models
Language
QwenDeepSeekKimiGLMMiniMaxLlamagpt-oss
Embed · vision · speech
BGEQwen-VLFunASR
Compute & inference
GPU
NVIDIAHuawei AscendHygonCambricon
Serve & optimise
vLLMSGLangTensorRT-LLMllama.cpp
Platform
KubernetesLiteLLM
Across every layer: security · observability · evals
04 · Business case

Three bets on a pipeline we already have.

€5.3M
GM pipeline closing in 2026of which €2.1M is planning work
Bet 1 · What we sell

Agentic offerings

Today
50%
Aim
70–80%
agenticML & data
Agentic share of projects sold, by year
20
21
22
23
24
25
26
aim
0% → 50%2020 to 2026 YTD · 25 agentic projects for 11 clients in 2025–26
Bet 2 · Where we sell

China + MENA + global

Today
18%
Aim
30–40%
GM sold outside Chinaaim: to confirm
China€4.0M
Southeast Asia€0.87M
MENA€16K
9 projectsdelivered outside China, 4 countries, €1.5M revenue
Halo · CIMB · BMMB, co-sold and co-delivered with the Southeast Asia team
Bet 3 · Who we sell to

Roll out across industries

Today
Aim
LuxurySportBeautyRetailHealthFinance
Luxury51%→~30%
Retail1%→~18%
Health1%→~12%
Finance11%→~12%
38 clientsin 9 industry groups already; the aim mix is illustrative
What I needGlobal Tech to co-build the stackIntroductions in MENAPilot accounts in other partners' portfoliosTBD people and budgetTBD 2027 ambition
Thank you
Make planning an AI-native capability: machine learning forecasts, orchestrated by agents, deployable on open-weight models.

Phil Zhang · VP Data Science · Q&A

← → navigate · N scorecard tags