Solutions in action

Proof you can evaluate.

Every implementation here is reported the same way: the starting problem, what changed, the systems involved, what was measured and the result, with a link to the original source. No testimonials.

Watch our server management solution ↓

Featured solution · Server operations

One dashboard. Faster response.

Server visibility, remote management through ChatGPT, and an agent ready to investigate new alerts.

22-second overview. Illustrative dashboard data. No narration.

Connected tools.
Human oversight.

Bring services, storage, network health and scheduled tasks into one view. Use ChatGPT MCP connectors to inspect and manage operations remotely through conversation.

An integrated agent investigates new alerts using logs and service health. Sensitive actions require approval, with activity and findings available for review.

Discuss your server operations

From alert to investigation.

  1. Detect a new operational alert.
  2. Investigate logs and service status.
  3. Request approval for sensitive actions.
  4. Record activity and findings for review.

A functional demonstration for personal and professional server operations. No measured response-time or savings claim is made.

Download the project JSON ↓
View the structured project write-up
{
  "title": "One Dashboard. Faster Response.",
  "description": "Bring server operations into one dashboard, with visibility into services, storage, network health, and scheduled tasks. ChatGPT MCP connectors enable remote agentic management through conversation, while an integrated agent responds to new alerts by investigating logs and service health. Sensitive actions require approval, and activity and findings remain available for review.",
  "features": [
    "Unified server monitoring and management",
    "Remote management through ChatGPT MCP connectors",
    "Automatic investigation of new alerts",
    "Approval controls for sensitive actions",
    "Recorded activity and maintenance reports"
  ],
  "audience": "Home server administrators and professional server operators"
}

Music: Sascha Ende, ende.app (CC BY 4.0).

Documented implementations.

From other organizations, reported by the company or by the press. They show the approach working. They are not xGold engagements and not outcomes a new client should expect.

Providence Health + IBM · Internal hiring

Orchestrating existing HR systems.

Problem
Hiring managers spent most of an internal transfer on administrative steps spread across HR systems. Transfers were slow and costly to process.
Changed
An AI-driven workflow now coordinates the steps across the existing systems. No system was replaced.
Systems
Oracle HR systems, orchestrated by IBM.
Measured
Hiring-manager admin time, days to start, cost per transfer, process accuracy.
Result
90% less admin work; starts 12 days sooner; 60% lower transfer cost; up to 70% more accurate. Company-reported, October 2026.
Original source

AT&T · Legal operations

Moving outside-counsel work in-house.

Problem
Routine research, drafting and document review were billed by outside law-firm associates.
Changed
A 38-person internal LegalEdge team took on that work with AI-supported research, drafting and review.
Systems
Not disclosed in the reporting.
Measured
Reliance on outside associate time.
Result
Reduced outside-counsel reliance. No company-wide ROI figure reported. Reuters, October 2026.
Original source

Ropes & Gray · Adoption

Paying for the learning curve.

Problem
Billable-hour targets penalized the time lawyers needed to learn and test AI tools.
Changed
Up to 100 hours of AI training and experimentation per lawyer now count toward annual billable targets.
Systems
Firm AI tools; not itemized in the reporting.
Measured
Hours logged on AI work.
Result
More than 35,000 associate hours on AI in 2026 across 1,000+ eligible lawyers. Reuters, October 2026.
Original source

Hiscox + Google Cloud · Underwriting intake

Reading broker submissions in minutes, not days.

Problem
Broker submissions for sabotage-and-terrorism renewals took about three days to turn around.
Changed
A model reads the submission and drafts a priced quote email. An underwriter reviews every quote before it is sent.
Systems
Google Cloud Gemini; Hiscox AI Labs; broker email.
Measured
Quote turnaround time; share of the book on the process; UK productivity on submissions.
Result
Turnaround about 3 days to about 3 minutes; roughly two-thirds of the book on the process; CFO cited a 40% UK productivity gain on submissions. Company-reported, 2024 to 2025.21
Original source

Dow + Microsoft · Freight invoice reconciliation

Auditing the invoices that arrive outside the system.

Problem
Up to 4,000 shipments a day; about 20% of freight invoices, over 100,000 a year, arrived as emailed PDFs outside EDI and were never audited for overbilling.
Changed
An agent pulls PDF invoices from email, structures the data and flags mismatches to a dashboard. Staff query the data in plain language.
Systems
Microsoft Copilot Studio, Microsoft 365 Copilot, Dow's integrated data hub, email, EDI.
Measured
Anomalies found across eight months of 2024 data covering 43,000 shipments; projected savings.
Result
Pilot surfaced billing anomalies such as a $30,000 surcharge against a typical $5,000. First-year savings of "millions" are projected, not realized. Vendor-reported, November 2024.25
Original source

Klarna + OpenAI · Customer service

The case that shows why you keep measuring.

Problem
Chat resolution averaged 11 minutes. Non-engineering hiring was frozen.
Changed
An assistant handling refunds, returns, payments and disputes launched in the app across 23 markets in January 2024.
Systems
Klarna app; OpenAI models.
Measured
Conversation volume, resolution time, repeat inquiries, satisfaction, headcount equivalent.
Result
2.3 million conversations in the first month; resolution 11 minutes to under 2; repeat inquiries down 25%. In May 2025 the CEO said quality had fallen and Klarna began rehiring human agents. Company-reported; use both halves.23
Original source

Octopus Energy · Customer email

A scoped trial with the hard cases routed to people.

Problem
Routine customer emails to the UK supplier consumed advisor time. No baseline count was published.
Changed
A three-month trial of an assistant answering routine emails end to end. Vulnerable customers, sensitive cases and complex complaints went to humans.
Systems
Kraken operating platform, owned by Octopus.
Measured
Emails handled per week, share of total email, satisfaction against comparable human replies.
Result
About 8,000 emails a week, 4% of all UK customer email; satisfaction 76% against 72% for human advisors. Company-reported, July 2026.26
Original source

UK Government Digital Service · 12 departments

Twenty thousand licences, honestly measured.

Problem
Drafting, summarizing and record updates across 12 organizations. No baseline beyond self-report.
Changed
20,000 Microsoft 365 Copilot licences for three months with change management and training.
Systems
Microsoft 365 Copilot in Teams, Word, Outlook and PowerPoint.
Measured
Self-reported daily time saved from 7,115 responses, adoption rate, satisfaction. No control group.
Result
26 minutes a day on average; 17% saw no saving; adoption peaked at 83%. The report says it could not identify how the saved time was used. Government-authored, June 2025.27
Original source

Morgan Stanley + OpenAI · Advisor tooling

An assistant that writes into the system of record.

Problem
Advisors could not efficiently search about 100,000 internal research documents, and meeting notes were written by hand.
Changed
An assistant over internal research, then a meeting tool that, with client consent, drafts notes, surfaces actions, emails a summary and saves the note to the CRM.
Systems
OpenAI models; Salesforce.
Measured
Adoption by advisor team. No time or cost figures published.
Result
98% of advisor teams using the assistant. Adoption only; no savings figure. Company statements via InvestmentNews, June 2024.28
Original source

DBS Bank · Contact-centre co-pilot

Giving agents the answer and the notes.

Problem
About 500 customer service officers searched knowledge bases and wrote post-call notes by hand during and after every call.
Changed
A co-pilot transcribes the call live, searches the knowledge base for suggested answers, writes the summary and pre-fills the service request.
Systems
DBS knowledge base and service-request system; governed under the bank's AI protocol.
Measured
Transcription accuracy, time per service request, officer sentiment, query volume.
Result
About 20% less time per service request across 500 staff handling 250,000+ queries a month. "Up to" wording; no published baseline. Company-reported, 2024.29
Original source

What a pilot report looks like.

This is the acceptance scorecard a process owner receives at the end of a build-and-pilot phase. The figures below are illustrative, not client data.

Illustrative scorecard · Invoice intake pilot · 8 weeks

MeasureBaselinePilotTargetStatus
Cost per invoice$11.20$4.10≤ $5.00Met
Days to post9.82.6≤ 3.0Met
Exception rate24%11%≤ 10%Not met
Field accuracy97.1%99.4%≥ 99.0%Met
Review minutes per exception—4.2≤ 5.0Met
Unauthorized actions—00Met
Run cost per invoice—$0.31≤ $0.50Met

The decision

Expand, improve or stop.

Six of seven measures met. The exception rate missed because one vendor's statements arrive as scanned images; the fix is scoped at two weeks. Recommendation: expand to the remaining entities, with the scan fix in the first sprint.

Signed by the process owner, the finance contact and xGold.

When we publish our own implementations they will appear above in the same five-field format, including the measures that did not move.

Start with one workflow.

A 30-minute call. You leave with a one-page map of the workflow you picked, what it costs you today and whether it is worth a fixed-fee assessment.

Start an assessment How we work