AI doesn't dashboard. AI decides.
The brain in the middle of the operating system. It runs every night. Reads sales, stock, supplier files, calendars. Forecasts. Decides. Drafts. By 09:00, the work is done — explained, ready for human approval.
- 1Reads26 weeks of sales, current stock at every store, open POs, supplier knowledge files (SIFs).
- 2ForecastsPer SKU per location. Picks the best-fitting model and writes down why.
- 3DecidesOrder now / wait / watch per supplier — with the reasoning, the cover horizon and the freight flag.
- 4DraftsPre-fills POs and store transfers ready for human approval at 09:00.
Sits on top of your ERP. Reads stock, writes transactions, never duplicates the source of truth.
Most 'AI inventory' tools just put a chatbot on a dashboard.
If a system shows you charts and waits for you to figure out what to do, it's not AI — it's a search engine with extra steps. Stokk's brain decides. Then it writes down why.
- Reads everything overnight, decides before your buyer logs in
- Reasoning attached to every recommendation
- Supplier knowledge captured in human-written files
- Your data is yours — never used to train shared models
- Claude prompts scoped per request, not persisted
Capabilities, in plain language.
Overnight intelligence
A nightly run that evaluates every supplier, every store, every replenishment cycle. By 09:00 the work is done.
Supplier intelligence files (SIFs)
Capture buyer knowledge — shutdowns, recall histories, freight quirks, allocations — in plain text. Feeds into every brief.
Explainable decisions
Every recommendation comes with a written reason. No black-box rules, no opaque scores.
Auditable history
Every brief, every draft, every override is versioned. You can replay last May's decisions next May.
Tenant-isolated data
Your data stays in your tenant. Never used to train shared models. Encrypted at rest.
Human always approves
AI drafts; the buyer approves. The system never sends a PO, a transfer, an SMS or a wallet-pass update without a human signing off.
What 'explainable' actually looks like.
Every recommendation comes with reasoning and an audit trail. Replay last May's decisions next May.
- Lead time slipped from 18d → 21d in Q4
- Top-mover RC-AD-3KG up 14% YoY at A-class velocity
- MOQ €110k met by current draft (€118.4k)
- Supplier shutdown Jul 22 — order in this cycle
- 02:14 · AIEvaluated 14 suppliers · 4 ready
- HM08:51 · Helena M.Approved Acme Home Goods · €185,400
- 09:02 · AIDrafted 3 store transfers from suggestion
- ML09:14 · Marcus L.Edited Berlin · -8 lines, +reasoning note
- 09:18 · AIUpdated supplier intelligence file from edit
The numbers Stokk customers report after the first cycle.
After the first full ordering cycle, lost-sale events on A-classified items roughly halve.
Buyers stop building proposals from scratch. The Brief lands; they review and approve.
Stokk's per-store sizing and lateral transfers free cash that was sitting on a pallet.
Continuous, offline-capable counts replace the once-a-year shutdown count.
Powered by Claude — wired into every module.
Stokk runs Claude (Anthropic's frontier AI model) for the reasoning that turns numbers into recommendations. The brain is tightly scoped: it sees the data it needs for the decision in front of it, nothing more.
Same brain, different surfaces: it writes the Monday Brief, it flags count discrepancies, it sizes a fulfilment queue, it suggests a member segment. The interaction is always the same — Stokk decides, you approve.
How AI gets used here
Decisions stay explainable. Every recommendation has a written reason and a human approval step. Your data isn't used to train shared models. Claude's prompts are scoped per request.
Plays nicely with the systems you already pay for.
Common questions about this module.
Does Stokk train Claude on our data?
No. Claude prompts are scoped per request and not persisted by Anthropic. Your data is never used to train shared models.
What stops the AI from doing something stupid?
A human approval step on every action that leaves Stokk — POs, transfers, SMS, wallet-pass updates. The AI drafts; you approve.
What model exactly?
Claude Opus on the heavyweight reasoning surfaces (Monday Brief, supplier eval). Lighter models for high-volume helpers (NLQ assistant, draft labels).
Other parts of the operating system.
See The AI brain on your data.
A 20-minute demo on your ERP, your SKUs, your stores. We do the work; you decide if it earns its place in your operating system.