The State of Enterprise AI Assistants
From pilots to production: how the world's largest organisations are deploying AI assistants at scale
Bank of America Erica interactions
Research by Faisal Shariff
A curated analysis of notable enterprise AI deployments across global retail, agentic platforms, and cross-industry implementations. This research highlights significant publicly disclosed deployments and is not intended as an exhaustive market survey.
The Retail AI Landscape
Every major retailer has moved beyond pilot. Here's who launched what, when, and at what scale.
Ahold Delhaize
Steijn (Albert Heijn)
Powered by In-house
2024
Millions of app users[17]
Tesco
Store & Colleague Intelligence
New partnership with Mistral AI
2025-12
Enterprise-wide[22]
Aldi
GenAI Content Platform
Powered by Digital Wave Technology
2025-03
US digital channels[31]
Lidl
AI Operations Suite
Powered by In-house + Cognigy
2024
European operations[111]
Global Grocery AI Landscape
Deployment scale vs. capability sophistication across global grocery retail. Hover for details.
The Search-to-Chat Transformation
Search is fragmenting. Chat is rising. The data tells the story.
Microsoft Copilot
Top User Intents
Analysis of 37.5 million conversations (Jan–Sep 2025): desktop dominated by 'Work and Career' (8am–5pm), mobile dominated by 'Health and Fitness' at all hours
50% of tech leaders unsure $30/user/month is justified. Jefferies projects $11B revenue by CY2026.
Google Gemini
enterprise users of Gemini Pro globally
of Gemini usage from enterprise users
of Fortune 500 using Gemini in at least one department
Workspace Integration
2.3 billion document interactions in H1 2025; Gemini in 73% of enterprise accounts
Agentic Capabilities in Production
The shift from 'assistants that inform' to 'agents that act' accelerated dramatically in 2024–2025.
Inform
Answer questions, summarise documents, search knowledge bases
Create
Generate tickets, records, documents, presentations
Update
Modify data, change statuses, process approvals
Execute
Book meetings, submit forms, trigger workflows, resolve tickets
Orchestrate
Multi-step workflows across multiple systems with autonomous decision-making
Answer questions, summarise documents, search knowledge bases
Generate tickets, records, documents, presentations
Modify data, change statuses, process approvals
Book meetings, submit forms, trigger workflows, resolve tickets
Multi-step workflows across multiple systems with autonomous decision-making
ServiceNow × Orica
IT Service Desk deflection
Doubled fully resolved cases without human intervention[56]
Salesforce Agentforce
GA 2024-10-25Shift from Einstein (recommendations) to Agentforce (autonomous execution)[52, 51, 53]
Customer Results
Key Metrics
- 18,500 deals closed across 39 countries
- Agentforce accounts in production up 70% QoQ
- ARR up 330% YoY, surpassing $500M
Notable Demo
UK Police deployed 'Bobby' agent for non-emergency calls across multiple departments, covering 90+ topics, delivering 20% reduction in human agent demand.
Marc Benioff's vision: one billion agents by end of 2025
Capabilities
Retrieves data, reasons through options, builds action plans, executes them
ServiceNow AI Agents
GA 2025-05AI Agent Fabric for cross-system orchestration across ServiceNow, Microsoft, Google, and Box[56, 57, 58]
Customer Results
Key Metrics
- 18% reduction in escalated incidents
- 33% reduction in incident resolution time
- 15–25 minutes reduction in major incident resolution time
Notable Demo
Stellantis supply chain: AI agent detected 25% battery cost increase, recommended alternate supplier, conducted capability analysis — all from single human alert
Microsoft Copilot Agent Mode
GA 2025-11Agents plan, execute, and validate multi-step tasks directly in Office apps[59, 60, 61]
Capabilities
Creating documents, financial models, table reshaping, presentations from conversational prompts
Grounding
Microsoft Graph for permission-checked retrieval with transparent citations
Cross-Industry Scoreboard
AI assistants at scale across banking, healthcare, telecoms, and manufacturing.
JPMorgan
LLM Suite
Scale: 250K employees, ~half daily
5-page IB deck in 30 seconds[76, 77, 83]
Tech: OpenAI + Anthropic models (multi-model strategy)
American Banker 2025 Innovation of the Year
Deep Dive
Launched as secure corporate ChatGPT in 2023, scaled to 250K in 8 months. Now layering agentic AI for regulatory reports. Staff may fall 10% over 5 years.
HSBC
Multiple AI tools
Scale: 600+ AI use cases; 20K developers with coding assistants
15% coding efficiency gain; 3M client interactions/yr via GenAI (88% ease)[81, 82]
Tech: Google partnership; LLM productivity tool for all colleagues
Group-wide mandatory AI training plus 'AI Academy' for beginner-to-advanced skills
Morgan Stanley
AI @ MS Assistant + AskResearchGPT
Scale: 16K advisors (WM, 98% adoption) + Institutional Securities
98% advisor adoption; document access 20% → 80%[84, 85, 87, 90, 92, 95]
Tech: OpenAI GPT-4 (enterprise arrangement)
Two flagship AI assistants: AI @ MS Assistant serves 16,000 Wealth Management advisors with instant access to proprietary research and policy content — 98% of advisor teams are active users. AskResearchGPT serves Institutional Securities, synthesising 70,000+ research reports annually with one-click email-draft export and inline citations. Both grounded in proprietary content only, with human-in-the-loop for all client-facing outputs. Bank aims to double client assets to $10 trillion partly through AI-freed advisor capacity.
The ROI Evidence Wall
Published, cited figures from production deployments. No vendor hype.
Bot Sprawl & Federation
The average enterprise has 12+ separate AI entry points. Fragmentation kills adoption.
Federation Solutions
Meta-Bot Architecture
Single entry point that ingests user intent, determines which backend system should handle it, and routes accordingly[105]
Microsoft Agent 365
Control plane for managing and securing agents with enterprise-wide visibility[60]
Academic: Intent-Based Orchestration
Centralised intent analysers, orchestration engines, and reasoning systems creating a single access point with domain collaboration[107]
Governance Best Practices
Central agent registry tracking all agents, owners, purposes, permissions
Standardised templates and approved configurations to reduce duplication
Single entry point for users with intelligent routing to backend systems
Human-in-the-loop for high-stakes actions with approval workflows
Cross-team coordination to prevent shadow AI and duplicate development
What Failed (and Why It Matters)
The lessons that don't make it into vendor pitch decks.
McDonald's
AI Drive-Thru Cancelled
IBM partnership from 2021, tested at 100+ US locations. Accuracy was only 80% — below the 95% target.
TikTok videos showed AI adding 260 McNuggets ($211), bacon to ice cream, and multiple sundaes instead of a vanilla ice cream
Outcome
Instructed franchisees to remove technology by 26 July 2024[75]
Lesson Learned
Voice AI in noisy, high-variance environments requires far higher accuracy than currently achievable
Air Canada
Chatbot Legal Liability Ruling
Chatbot provided incorrect bereavement fare information to a grieving passenger.
Air Canada argued the chatbot was a 'separate legal entity' — the BC Civil Resolution Tribunal called this a 'remarkable submission' and rejected it
Outcome
Ordered to pay $812.02 in damages. Chatbot removed from website.[106]
Lesson Learned
Companies are legally responsible for all information on their websites, regardless of whether it comes from a static page or a chatbot
AI Overviews Launch Errors
Within days of US rollout, AI Overviews produced viral errors from unvetted sources.
Suggested non-toxic glue for pizza (from 11-year-old Reddit joke) and eating rocks (from The Onion)
Lesson Learned
AI-generated content drawing from unvetted sources at scale creates massive reputational risk
Systemic Challenges
$67.4 billion
Hallucination Costs
Estimated enterprise losses globally in 2024. 82% of AI bugs stem from hallucinations — 'silent misinformation' that looks perfectly normal.[106]
$14,200 annually for mitigation
Sources & Methodology
Every claim in this report is traceable to a named source.
This research draws on primary sources (company announcements, earnings calls, corporate blogs), analyst reports (Forrester, Gartner), trade press, and vendor case studies. All data points are from 2022–2025. Claims are cited to their original source.