# Ward (getward.ai)

> Ward is an AI analytics and observability platform for multi-store retail. It connects read-only to the systems retailers already run (POS, inventory, ERP, e-commerce, data warehouse, BI) and delivers automated insight cards, not dashboards. Every card states what changed, the root cause, and the recommended action. Ward also publishes practical guidance on the AI infrastructure underneath: LLM-agnostic architecture, AI orchestration, observability, and AI-driven workflow automation.

Ward is operated by askotter. Pricing scales with operational complexity (store count, data volume), not seats. First insight cards arrive within ~48 hours of connection; robust baselines form within ~2 weeks. The underlying numbers come from auditable statistical and time-series models; the LLM frames the result, it does not invent the math.

## Architecture: the four layers

Ward is four layers over a retailer's existing stack. Each has a stated boundary and can be evaluated separately.

1. **Retail observability.** Continuous reads across POS, ERP, WMS, labor, finance, and supplier feeds, joined to weather and event data. Anomalies score against each store's own baseline rather than a chain aggregate. Every finding carries a reasoning graph: sources touched, SQL run, model fit, causes ruled out. Boundary: reads only; no data is copied out of the customer's warehouse.
2. **AI orchestration.** An abstraction layer in front of Anthropic, OpenAI, Gemini, and Ollama, running on the customer's own API keys and commercial terms. Each task routes to the cheapest model that clears the customer's accuracy bar, with cross-provider failover. An eval harness scores candidate models against the customer's real cases before promotion. Forecast math runs on ARIMA, Holt-Winters, Bayesian hierarchical, and gradient-boosted models, backtested over 24 months with MAPE attached. Boundary: the LLM frames answers and never produces the numbers.
3. **Workflow automation.** 43 prebuilt retail playbooks, each specced with a trigger, a procedure, a write-back target, and the KPI that must move before the case closes. Write-back targets include SAP, Oracle Retail, Blue Yonder, Relex, Manhattan, and pricing and promo engines. Boundary: every write gates on a named approver in a customer-defined role, and is logged before and after.
4. **Security and data governance.** Federated query against Snowflake, BigQuery, Redshift, Postgres, and SAP HANA, with no second copy and no shadow lake. Read-only service accounts locked to SELECT. Cedar policies scope each agent by role, tenant, and resource. Agent charters and Cedar policy live in the customer's Git repository, versioned, reviewed, and signed. Every query, model call, and approval streams to the customer's SIEM as JSONL. SSO/SAML, SCIM, RBAC, customer-managed keys, TLS 1.3, AES-256 at rest. SOC 2 Type II underway.

## For technical evaluators (CIO, Head of IT, platform teams)

Ward is designed to sit inside an AI strategy the customer already owns rather than replace it. It inherits the customer's model vendor contracts, warehouse, identity stack, and policy framework.

- **Customisable.** Cedar policies and agent charters live in the customer's repo, so scoping an agent is a reviewed pull request. Each charter declares scope, sources, allowed actions, owner, and version. Playbooks are specced objects tuned per vertical and per store cluster.
- **What the platform team runs.** A read-only service account per source plus an identity integration. No ETL to schedule, no warehouse to stand up, no model to host.
- **No model lock-in.** Providers are configuration, not code. Switching is a config change validated by the eval harness, not a migration.
- **Graduated write scope.** Read-only for the entire first phase. Write access opens one playbook at a time against a named approver role.
- **Review artifacts before the pilot.** Data flow diagram, Cedar policy bundle, network topology, sub-processor list, SOC 2 Type II status, and a pre-answered security questionnaire.
- **Onboarding path.** [getward.ai/cio](https://getward.ai/cio) sets out the six-week deployment week by week: what connects, what the customer signs off at each gate, and the internal case for the CFO, security, the board, and the platform team.
- **Evaluation packet.** [getward.ai/cio-packet](https://getward.ai/cio-packet) is a self-contained, printable document covering architecture, data handling, access control, audit, the model layer, deployment gates, available documents, commercial terms, and exit. Written for security and procurement review before signature.
- Full detail: [architecture for Head of IT](https://getward.ai/icp-head-it), [trust and governance](https://getward.ai/trust).

## Product
- [Product overview](https://getward.ai/product): What Ward is and how insight cards work.
- [LLM-agnostic platform](https://getward.ai/product-llm-agnostic): Why Ward is not locked to a single model provider, and what model-agnostic architecture means in practice.
- [Insight cards](https://getward.ai/product-insight-cards): Cards instead of dashboards, with root cause and recommended action.
- [KPI monitoring](https://getward.ai/product-kpi-monitoring): Continuous monitoring of retail KPIs between report cycles.
- [Closed loop](https://getward.ai/product-closed-loop): Detect, attribute, recommend, execute with write-back, audit, measure.
- [Live pilot](https://getward.ai/product-live-pilot): 90-day pilot with pre-agreed metrics and close criteria.
- [Pricing](https://getward.ai/pricing): Tiers by store count and complexity.
- [Demo, security review, or packet](https://getward.ai/demo): One intake for three separate asks. A 30-minute product walkthrough, a 90-minute architecture and security review, or the evaluation packet with no call attached.
- [Engagement status](https://getward.ai/pilot-status): Where Ward's enterprise grocery work actually is. AI strategy, change management, AI orchestration and the reporting layer, running against live POS, ERP and inventory at nine-figure revenue. Operators are not named and their numbers are not published without permission; reference calls are arranged during evaluation instead. Carries no measured KPI result yet and says so; the delta is published here at day 90 whichever way it goes. The nine logos on the homepage are askotter platform customers, not Ward customers.
- [Document room](https://getward.ai/documents): Self-serve. One form and a click-through NDA opens the MSA, DPA, pre-answered CAIQ Lite questionnaire, sub-processor list, reference architecture, SOC 2 status, certificate of insurance and DPIA template. No sales call in between.
- [For data & analytics leaders](https://getward.ai/analytics-leader): How Ward coexists with Power BI, Looker and Snowflake without building a second semantic layer. SQL lineage on every number, model cards and MAPE on every forecast, and what does and does not come off the ad-hoc query queue. Parent of the three comparison pages.
- [The board case](https://getward.ai/board): For the CEO and board. Where the modelled 200 basis points come from and the same arithmetic in dollars, the four-option shortlist including doing nothing, the day-90 close criteria, and the 30-day exit. States plainly that the margin figure is modelled and unbooked.

## AI infrastructure (pillar guides)
- [AI infrastructure hub](https://getward.ai/ai): Orchestration, observability, and LLM-agnostic architecture, the three problems every production AI system hits.
- [AI orchestration](https://getward.ai/ai/orchestration): What orchestration is, how it differs from automation, model routing, agent workflows, context passing, guardrails, and build vs buy.
- [AI observability](https://getward.ai/ai/observability): What AI/LLM observability is, observability vs monitoring, and the signals to instrument (traces, cost, latency, quality, hallucination, drift).
- [LLM-agnostic architecture](https://getward.ai/ai/llm-agnostic-architecture): What model-agnostic means, the LLM gateway/abstraction-layer pattern, switching providers, and avoiding vendor lock-in.

## AI strategy & orchestration (advisory)
- [AI orchestration](https://getward.ai/advisory-ai-orchestration): Designing the system around models, model selection, routing, agent workflows, retrieval/RAG, build vs buy.
- [AI readiness](https://getward.ai/advisory-ai-readiness): Whether your data and stack are ready for AI, and what to fix first.
- [Data lake / AI-ready data layer](https://getward.ai/advisory-data-lake): The data foundation that makes AI useful, and its real total cost of ownership.
- [Systems modernization](https://getward.ai/advisory-systems-modernization): Modernizing legacy retail systems for AI workloads.
- [Operational deployment](https://getward.ai/advisory-operational-deployment): Getting AI from proof-of-concept to production.

## Key explainers (blog)
- [LLM-agnostic AI, explained](https://getward.ai/blog/llm-agnostic-retail-ai-explained): What model-agnostic means and why it matters.
- [AI agents in operations: 7 use cases](https://getward.ai/blog/ai-agents-retail-operations-7-use-cases): Where agentic workflows replace manual analysis.
- [AI vs traditional demand forecasting](https://getward.ai/blog/demand-forecasting-ai-vs-traditional): How the approaches differ and when each wins.
- [GenAI operator's guide (2026)](https://getward.ai/blog/genai-for-retail-2026-operators-guide): A practical guide to deploying generative AI in operations.
- [Observability vs analytics](https://getward.ai/blog/retail-observability-vs-retail-analytics): How observability differs from classic analytics.
- [The hidden TCO of a data lake](https://getward.ai/blog/hidden-tco-retail-data-lake): The real cost of building AI-ready data infrastructure.

## Reference
- [KPI glossary](https://getward.ai/glossary): Definitions, formulas, and benchmarks for retail KPIs.
- [Blog](https://getward.ai/blog): Operator-level writing on retail analytics, AI, and infrastructure.
- [About](https://getward.ai/about): Who builds Ward.
- [Trust & security](https://getward.ai/trust): Security posture, governance, and data handling.

## Contact
- Demo / sales: https://calendly.com/askotter
- Email: brian@askotter.ai