A predictive application is an operational SaaS where the data layer doesn't just store transactions — it predicts the next one. Predictive e-commerce means search results that re-rank for the customer in front of you, recommendations that learn from every click and purchase, demand forecasts that drive replenishment, inventory reorders that come with revenue-at-risk-in-€, markdowns that pick the discount that clears in three months at the highest recoverable margin, and churn risk surfaced before a customer goes silent.
Aito is the predictive database underneath. Every prediction comes back with a probability, top-3 alternatives, and a $why factor decomposition. No retraining, no MLOps, no separate ML pipeline alongside your storefront or merchandising tables.
ecommerce.aito.ai is our open-source reference implementation of a predictive e-commerce platform — the shop, the owner's views and back-office tools, built on _predict, _relate, _search, _recommend, _estimate, and _evaluate. PetNord is a synthetic Nordic-flavoured pet-store dataset generated deterministically from data/generate_fixtures.py (seed 42) to drive the demo: 658 SKUs, 3,000 customers, 10,594 orders, 36,357 order lines, 5,202 reviews, 5,626 recorded shop actions, a 27,003-row customer-month panel for churn modelling, and 15,157 SKU-month rows for demand forecasting. The data has planted effects, so the numbers cited from PetNord below show how Aito works and that it recovers those effects, not how much it would lift a real shop. They are reproducible against the live demo.
Read more: Every screen a decision, the tour of the whole shop and its routines.
The demo is arranged by who is asking, and every view reads from a single Aito database and shows the query behind it.
Shop (what a shopper sees)
? on each card opens the $why behind its percentage. A search ranks products by their words and by what this shopper is likely to buy.Run the Shop (the owner's decisions)
Operate (back-office tools)
$why.Evidence (how well; accuracy numbers live only here)
_evaluate runs on PetNord, including the ones that fail.Each use case below runs in the demo and has its own page with the query, the numbers and the honest limits.
For the shopper
For the owner: supply and pricing
For the owner: customers
Running on their own
Each number below is computed on the generated PetNord data (release 3b). Its effects are planted, so these show that the queries recover them, not an uplift in a real shop. The one real-data result is the UCI reorder figure above.
| View | What it shows | Number |
|---|---|---|
| Patterns | Dog dry food and dental treats in the same basket | 3.17× more often than chance, interval 2.75–3.66 (2,044 of 10,594 orders); dog dry food with cat foods about half as often as chance |
| Retention | Next best action, against the generator's own best action | Aito's top action is the best one for 81 % of 300 customers; mean regret €0.45 per customer, against €9.36 for always emailing and €10.72 for never acting |
| Retention | Signs before churn, for customers who had just ordered | An NPS answer of 0–3 comes with 1.93× the churn within three months, a 1–2 star review 1.23×, a late or damaged delivery 1.17×; an NPS of 9–10 0.71×. The generator caps these contrasts |
| Supply | The reorder queue for May 2026 | Reorder 83 products (833 units, €12,620 at cost): if May sells like the last three months, about €14,400 of sales can't be served without them; €27,691 tied up in 79 overstocked products |
| Model Checks | Held-out _evaluate | Pet type from the name 90.0 % against 40.0 % for the most common value; dietary 66.7 % against 26.2 %; churn +14.7 pp (interval +11.4 to +18.0) |
| Model Checks | The honest failure | Return risk 99.0 % against 99.0 %: +0.0 pp, flagged FAIL. About 3 % of lines come back, and nothing in the data predicts which. Demand ties the naive forecast |
Two things are true at once.
The verifiable facts. Aito has shipped predictive infrastructure in production since 2018; on the operational side, Nordic enterprise AP deployments have run at 95%+ accuracy on thousands of invoices per month — same _predict, same $why decomposition. On the commerce side, the structural regularity of consumer purchase data (catalog taxonomies, segment behaviour, seasonality, replenishment cycles) makes _recommend and _relate produce statistically meaningful results on a few hundred impressions per segment, and _predict produce calibrated forecasts on a panel of SKU-month rows.
The conditional answer. All commerce transactions are super-regular at their core. Catalog taxonomies are stable across years, suppliers map to categories, customer segments behave predictably, and seasonality repeats on calendar cycles. What differs between commerce tenants is how much of your incoming volume matches prior patterns with sufficient history vs. arrives as a first-time pattern. Predictive e-commerce is a conditional-probability engine that works the same way on each transaction; how many of your transactions benefit from it depends on the volume mix.
Consumer purchase data has more structure than typical ML training data. Catalog taxonomies are durable: product categories, brands, attributes, and tax classes are stable across years. Repeat customers behave consistently: someone who bought cat litter and lactose-free cat treats this month will likely behave the same way next month. Seasonality repeats: Q4 gifting, summer outdoor goods, back-to-school. Segment effects compound: similar customers do similar things, so a few hundred orders per segment carries real predictive signal.
This is why on regular commerce data a stable pattern becomes predictable after a handful of observations; how many depends on how consistent the pattern is, so measure it on your own history with _evaluate rather than assume a fixed count. The top confidence bucket carries near-zero realized error in production — Aito's confidence is calibrated, so when it says >95%, the realized rate matches.
Underneath all three, the engine is the same: same _recommend / _predict / _relate / _estimate shapes, same calibrated confidence. What changes between deployments is the volume mix.
| Profile | Examples | Volume mix | Range we typically see |
|---|---|---|---|
| Mostly-recurring | Grocery, pet supplies, replenishment commerce, subscription boxes, household staples, pharmacy, baby goods | Stable repeat customers, calendar-driven baskets, narrow tail | Cross-sell at 1.5-3× lift on common pairs; demand forecast accuracy strong on top SKUs; churn modelable with a 3-month panel; near-zero-touch reorder on staples |
| Mixed | Multi-channel general retail, apparel, home goods, electronics, beauty, sporting goods | Combination of repeat-pattern carts and one-offs; meaningful share of trend-driven novelty each season | Persona-conditioned re-ranking visibly outperforms popularity; demand forecast accurate on the core, weak on newness; markdown decisions reliable for stable categories |
| Mostly-snowflake | Gifting, made-to-order, art and collectibles, project-driven B2B commerce, gallery/curated commerce, antiques | Most incoming orders are first-time patterns. Cross-sell signal sparse at the SKU level; demand forecasting drops to category-level | Cross-sell shifts to the category band rather than SKU; demand forecast useful at the brand/category level; predictive search still works via Text tokenisation on name + description |
The conditional-probability machinery is the same across all three profiles. What differs is how much of your incoming volume matches patterns already seen many times vs. arrives as a snowflake.
Find your profile in 15 minutes: run _evaluate on a held-out slice of your real data. Per-target accuracy and per-confidence-bucket calibration tell you which range your tenants sit in.
A predictive e-commerce stack splits naturally into three layers, and the volume mix above determines how much each layer carries:
$why rendered.A predictive e-commerce product covering tenants with a heavier snowflake share will see lower headline accuracy at the SKU level — that's the data, not the model. Both numbers are correct, and both are useful, as long as your UI is field-level confidence-aware.
Merchandiser bumps the recommended reorder qty from 12 to 18 at 14:00:23. The next request for a similar SKU — different SKU, same category, similar velocity — reflects the new pattern in its conditional probability. No retrain queue. No model deploy. No A/B holdout.
| In a typical ML pipeline | In a predictive database |
|---|---|
| User correction → label store | User correction → INSERT |
| Nightly retraining job | (none) |
| Model artifact + deploy | (none) |
| Canary, holdout, regression monitoring | (none) |
| 24-72 hours from correction → live | <1 second |
Same loop covers a merchant tagging a recommended cross-sell as "no thanks", a buyer overriding the markdown level for an SKU, a CRM lead overriding the recommended retention action, an analyst correcting a category prediction during catalog enrichment. Every override trains the next prediction.
This is why predictive e-commerce scales differently than rule-based personalisation engines. Rules rot; predictions improve with every override.
The signature interaction in predictive e-commerce is one loop that crosses the consumer-facing / merchandiser-facing boundary without ever crossing a model boundary:
_recommend.$why.$why behind the risk.Every override trains the next prediction. Same operators across the whole loop. Same Aito DB. No model boundary between the consumer-facing and merchandiser-facing surfaces.
Run the same predictions used by ecommerce.aito.ai directly from this page. These queries hit our hosted PetNord database — no signup, no setup.
Run live queries against ecommerce.aito.ai's PetNord (Nordic pet store) database
Which categories go in the same basket as dog dry food. Dental treats lead at about 3.2× chance; cat foods come in at about half.
{
"from": "orders",
"where": {
"line_categories": {
"$match": "dog_dryfood"
}
},
"relate": "line_categories",
"limit": 8
}If you're building a predictive e-commerce product for multiple merchants, the same Aito table serves every tenant. merchant_id (or shop_id / customer_id) in the where clause scopes the conditional probability to that merchant's history. accounting.aito.ai runs 255 tenants this way on one shared Aito DB, sub-100 ms predictions, no per-tenant model or retraining schedule.
{
"from": "order_lines",
"where": { "merchant_id": "SHOP-0123", "customer_pet_size": "large" },
"recommend": "product_sku",
"goal": { "customer_segment": "dog_owner" },
"limit": 12
}
Same query shape, different merchant_id returns a different ranking — because the conditional probability is computed only over that merchant's rows. No per-tenant index, no per-tenant model, no per-tenant retraining schedule.
Three small, reusable components carry the predictive-application UX. Anyone porting this pattern keeps them.
PredictionExplanation — the $why rendererPure component that turns Aito's $why factor tree into:
baseP)baseP × lift₁ × lift₂ × … = finalPLiftHint / ConfidenceBar — the calibrated annotationA small badge anchored to a recommended SKU, a cross-sell, or a forecast value. Renders the lift or confidence as a coloured pill: green for strong (lift ≥ 1.5, confidence ≥ 85%), grey for neutral, faint red for the protective band (lift < 0.7). Merchant tooling — and savvy consumers, via a "why am I seeing this?" tooltip — gets a calibrated signal instead of a black box.
SmartField — three-state inputSingle DOM input that's the same field whether the value came from a prediction or the user. Three visual states: empty / predicted (gold-italic with a 🤖 badge) / user (normal). Tab promotes, Esc rejects, typing replaces. Used in Product Filling (catalog enrichment) and any merchandiser form where Aito has a credible guess.
These match the patterns in the smart form fill use case and the $why explanations in the API docs.
Traditional commerce personalisation means: write a segmentation rule, write a recommendation rule, write a cross-sell rule, write a markdown rule. Maintain them. Watch them rot. Hire vendors to fill the gaps.
A predictive e-commerce stack flips it: load your orders and impressions, query for predictions, promote patterns to merchant-visible rules with governance. No retraining, no MLOps, no feature engineering. The data is the model.
| Approach | What it covers |
|---|---|
| Rules-only personalisation engine | Static segmentation, hand-tuned cross-sell, popularity ranking — ceilings at top-of-funnel, no operational coverage |
| Bolted-on commerce search & rec vendor | Search and recs only; demand forecast, markdown, churn, win-back stay separate vendors |
| Predictive e-commerce on Aito | One substrate covers consumer-facing AND operational — search, recs, demand, inventory, markdown, churn, win-back, catalog enrichment, anomaly detection |
The fastest way to disqualify Aito for your e-commerce use case is to run _evaluate on a slice of your real data.
_evaluate on a held-out slice for top-k recommendation accuracy, demand forecast accuracy, churn AUC, or whatever target matters for your roadmap.If the lift over baseline is convincing, schedule a technical review. If not, you've spent an afternoon.
_evaluate is the same operator that powers the Model Checks view on ecommerce.aito.ai. Same code path you'd use in production to monitor accuracy per merchant.
For a 5-15 person product team building a predictive e-commerce SaaS:
Build it yourself. A working prediction loop with multi-tenancy, retraining, evaluation, and override capture takes ~6 months and ~1.5 FTEs (one ML engineer, ~0.5 backend for plumbing). At €100-130K loaded ML salary plus infrastructure: year-1 TCO €200-400K, year-2 €120-200K maintenance. Plus you're operating retraining infrastructure in production while shipping new features.
Drop in Aito. API key and SQL-like queries, 1-2 weeks to a working prediction loop. Multi-tenancy is merchant_id (or shop_id) in the where clause. Aito plans start at €0 (sandbox), €75/mo (small production), €350/mo (growth) — see pricing. You give up some control over model internals; you gain ~5 months of engineering and the right to delete user data without a model-retraining ticket.
The breakeven question. Time to first paying merchant matters more than build-vs-buy purity. Six months later means six months less revenue and a smaller dataset to learn from when you do ship.
The honest alternatives we hear:
vs. Algolia AI Recommendations / Klevu / NoSto / Searchspring / Boost (commerce search & rec vendors). These bolt on as services with their own data ingestion, taxonomy, and UI patterns. They work, but: (a) they own your prediction layer and the taxonomy decisions inside it; (b) they cover search and recs but stop at the storefront, leaving demand forecast, markdown, churn, and win-back for separate vendors; (c) the same pixel reports to several SaaS products rather than one substrate. Aito is the database underneath your own features, so the predictions are yours, the taxonomy is yours, and a single substrate covers consumer-facing AND operational.
vs. Shopify Magic / Magento Live Search / BigCommerce native AI. Great if you're a merchant on that platform. Useless if you're a platform vendor building a competing or complementary product. Aito is the substrate for the team building the platform feature.
vs. Klaviyo / Recharge / Bloomreach (retention and personalization vendors). These ship CRM + email + retention workflows. They own segmentation and campaign delivery. Aito does the predictive scoring underneath: which customer to win back, which product to offer, how to price the offer. Most predictive e-commerce roadmaps end up running one of these and Aito — the CRM for the workflow, Aito for the score.
vs. GPT-4 + a vector database for product reasoning. LLMs reason well over unstructured input but cost ~1-3 sec per call, are opaque (no $why), and don't learn from corrections. For volumetric commerce work — thousands of impressions/day per merchant — the unit economics break. Use LLMs for free-form customer-message classification or product-description generation, Aito for the structured prediction.
vs. building it on scikit-learn / Vertex AI / SageMaker. You can. You'll spend the 6 months above. The thing you build is a worse predictive database with retraining infrastructure you operate.
For commerce SaaS platform vendors. Shopify-class, headless commerce, niche vertical platforms, marketplace platforms. Bake predictions into your product. Personalised search and recommendations, demand forecast, markdown, churn and retention actions become default features instead of premium add-ons or third-party integrations. Multi-tenancy is merchant_id in where — same pattern accounting.aito.ai runs across 255 tenants on a single Aito DB.
For e-commerce agencies and integrators. Shopify Plus shops, headless commerce agencies, BigCommerce / Magento / WooCommerce dev shops, freelancers running multi-client retainers. Aito is forkable (Apache 2.0), the PetNord demo doubles as a client pitch deck, and one Aito plan can drive predictions across a portfolio of small clients. See the Partner page for the agency programme.
For ERP vendors with a commerce module. Oscar Software, ERPly, Visma, Lemonsoft, NetSuite — if your product spans accounting + commerce + inventory, the same Aito instance running AP coding and PO routing also runs demand forecast, markdown, churn, and win-back for the commerce side. One substrate, three product surfaces.
Latency is per-prediction in the 20-50 ms range, sub-200 ms P95 at scale. Throughput is what you provision for:
| Plan | Cost | API calls / month | Storage | Right for |
|---|---|---|---|---|
| Sandbox | €0 | 5,000 | 100 MB | Evaluation, PoC, _evaluate on sample data |
| Dev | €75 | 500,000 | 500 MB | One small production merchant or pilot |
| Prod | €350 | Unlimited | 1 GB (extendable) | Single-instance production at SME scale |
| Enterprise | Contact | Custom | Custom | Multi-tenant SaaS, multi-region, dedicated infra |
Worked example. A predictive commerce SaaS with 200 merchants averaging 500 product impressions + 50 cart events per merchant per day = ~110K events/day. Each impression triggers ~2 predictions (re-rank + cross-sell), each cart event triggers ~1 (completion suggestion) = ~210K predictions/day. Comfortable on Prod (€350/mo, unlimited). Storage at ~10K rows per merchant fits 1 GB; extends as you grow.
For multi-tenant SaaS beyond ~500 merchants or multi-region deployments, contact us — Enterprise pricing is per-instance with dedicated infrastructure.
DELETE the rows and predictions reflect it from the next query forward. Important for GDPR Article 17 in consumer-facing e-commerce.customer_id (CUST-NNNNN). The recommended pattern: keep PII in your own DB, send Aito the segment / behaviour / event signals scoped per customer_id.We do not currently hold a SOC 2 Type II attestation; we share our security posture and pen-test results under NDA. For procurement that requires SOC 2 today, contact us.
E-commerce platforms ship for 10+ years. Before you bet a product roadmap on Aito:
_predict, _relate, _search, _recommend, _estimate, _evaluate, _match) dates from the public v1 release in 2019. We version at the API level, and behavior changes are listed in the changelog._search. There is no proprietary format you can't export. Worst case migration: read your rows, write them somewhere else.For company posture, funding, and team — see About or contact us for current details under NDA. Multi-year procurement reviews welcome.
Ready to ship a predictive commerce product — or add predictive features to one you already operate?
Try ecommerce.aito.ai → — the shop, the owner's views and the evidence, live.
Read the source on GitHub → — Apache 2.0, with a use-case guide per view and the architecture decision records.
Schedule a Technical Review → — walk through your specific commerce automation use cases with our team.
Start Free Trial — point Aito at your own order and impression data in our sandbox environment.
Predictive Accounting, Predictive ERP, and Predictive E-commerce are not three separate products. They are three instances of the same pattern: a SaaS application designed around the premise that the user should never make a decision the data has already made for them. The substrate is a predictive database. Prediction becomes a query rather than a project.
For the full architectural argument, see The Predictive Application. The canonical explanation of why per-prediction cost determines what gets built, and what changes when that cost collapses.
Episto Oy
Putouskuja 6 a 2
01600 Vantaa
Finland
VAT ID FI34337429