Resources

Most AI transformation decisions get made before anyone has explained what actually gets built. This one-page primer sets out the layered engineering stack behind Intelligent Enterprise AI: foundation models chosen per use case for cost, latency and data residency; RAG over your own ERPs, SOPs and contracts so answers carry citations; agentic orchestration that plans multi-step work and calls your systems over protocols like MCP; and the guardrails, evals and human-in-the-loop gates that keep agents acting inside defined authority.

It then makes the case function by function — lead scoring and proposal drafting in sales, AP/AR extraction and reconciliation in finance, exception handling in supply chain, multilingual support with escalation logic in service — and closes on the playbook: audit for the document-heavy, rule-heavy work, pilot one agentic workflow in six to eight weeks against measurable KPIs, then scale on a shared platform with governance. Written for the leader signing off on the programme, not the team implementing it.

Agents multiply faster than anyone reviews them. Past a few thousand, an inventory isn’t governance — someone still has to judge each one, and that is exactly where retirement stalls while idle and duplicate agents keep consuming tokens, licences and review time. Agent Lifecycle Management reviews every agent in a Glean workspace, keeps the ones people rely on and retires the rest, running through the Glean APIs and admin controls a customer already has.

The service runs as four stages you can run again for the next pass without rebuilding: Stabilize (discovery, catalog, classify, cleanup), Standardize (an agent development lifecycle), Operationalize (AgOps) and Maintain/Govern. Across a 10,000+ agent registry, inactivity and update checks feed a decision engine behind subscription and routing safeguards — retiring 71% and retaining 29%, with a full decision audit trail. The criteria stay the customer’s: which agents are protected, how long inactive is too long, who signs off. Human review sits in front of every deletion, with a restore window behind it.

Enterprise search is only as good as what made it into the index, and conventional ingestion loses precisely the content people search for — the tables, the charts, the figures inside the deck. Agentic Document Intelligence turns every file in a client’s repository into fully-understood, Glean-native knowledge: a Connector VM running inside the client’s own cloud orchestrates Glean’s Chat agents to read each document and writes the result straight into the customer’s tenant through the Indexing SDK. No summarization, no lossy parsing, and no content ever leaves the tenant.

A watermark-driven checkpoint fetch picks up only what is new or updated since the last run, multi-process extraction calls Glean Chat in parallel for verbatim text, markdown tables and complete descriptions of every chart and infographic, and a bulk indexer activates the assembled documents as native Glean content — searchable, filterable and answerable across Chat, Agents and Search, access-aware and respecting Glean ACLs throughout. It is validated at 160,000+ document scale, auto-retries transient failures, resumes from any point, and tunes to a new repository through a single config file — connector VM to first indexed document measured in hours rather than weeks.

Transport operators lose revenue when the people or cargo actually carried don’t match what was ticketed or manifested, and without automated video analysis that leakage goes undetected. Smart Green Transport is an AI-powered audit and intelligence engine built on a YOLOv8 computer-vision pipeline: it counts passengers or cargo from onboard camera feeds and reconciles the count against tickets sold or manifests logged, flagging over-boarding and under-reporting as it happens.

The case study sets out the full architecture — YOLOv8 detection, a FastAPI backend, a PostgreSQL evidence store and a React dashboard for auditors — running at 95%+ detection accuracy. Because the engine uses open-weight models on-premise or in a private cloud, video, image and passenger data never leave the operator’s infrastructure, and the same pipeline extends directly to supply-chain tracking.

A sustainability risk score has to survive scrutiny, which means the number needs a defensible derivation rather than a black box. This white paper sets out a model that reads company-specific information and documents instead of relying wholly on statistical analysis of financial reporting: the company divides into the three ESG categories, each category into risk factors, and each factor into sub-factors — “waste”, for instance, drawing on both corporate offices and manufacturing facilities.

Qualitative inputs are converted to quantitative ones and normalized by type — log normalization across varying magnitudes, encoding for categorical data, scaling for ratios — then combined in a weighted linear pre-score that transforms into five brackets on a 0–10 scale. Deeper analysis adds rules the score function must obey: rising emissions intensity has to raise risk, and an established company takes a heavier penalty than a young startup for the same employee turnover. Machine learning then fits candidate score functions against the pre-scored data — and because the fit runs both ways, a function that disagrees with the pre-scores can prompt a revision of the rules rather than being tuned away.