AI Automation Examples: 22 Named Organizations, With Sources
AI automation examples from 22 named organizations: what each automated, the tools, the result it reports, the source and what needed work beyond the tool.
Luka Abramovic21 min read
In the 22 AI automation examples below, the clearest results come from automating one narrow, frequent task and passing the rest to a person. Hershey Entertainment & Resorts' phone agent fully handles more than 25% of staff call-offs, eight hours of work a day by the company's count. Definity saves upwards of 3.5 minutes a call, on more than 500 calls a day, by having AI write the call notes. But read the small print. Oper Credits' Google Cloud customer story lists "90% of loan applications complete and compliant on first submission" as a result; its CEO calls 90% a long-term objective.
Each row gives the organization, what it automated and with which tools, the result as the source reports it, and what needed work beyond the AI tool. That last column is our read, not the source's. Inclusion rule: a named organization, one specific task automated in day-to-day work (pilots and plans are labeled), and a public source we read. Checked in September 2026.
A note on sources. Eighteen rows come from Google Cloud's customer stories and its use-case list. Vendor stories are marketing: the numbers are the customer's own, we keep their hedges (estimated, expected, potentially), and listing a story doesn't mean we recommend the vendor. We read each story's body, not just its results summary, which is how we found the gaps listed in the case-study checks. Two rows are our own clients, marked as such. Most detailed stories come from large organizations, so we kept the ones whose workflow a 50-to-500-person business also runs, and we give size wherever the source does.
Back office: invoices, staff calls, insurance data and permits
These rows start from inputs with a predictable shape: a supplier invoice, a shift call-off, a claims spreadsheet. The AI reads the input. The work beyond the tool is writing the result into the system of record.
| Organization | What was automated, and how | Result, as the source reports it | What needed work beyond the tool (our read) | Source |
|---|---|---|---|---|
| Nordell, plastic injection molding, Worthing, West Sussex (UK) | Supplier invoice processing with Demi AP, an AI invoice-processing product, integrated with the company's DELMIAWorks ERP. The £34,000 project had a £17,000 Made Smarter South East grant. | Expected to save more than 60 hours of administration a month. That was a forecast, made when the grant was announced. | The ERP end: matching each invoice to a supplier and purchase order in DELMIAWorks, and a place for the invoices it can't match. Ask for the measured hours a year on. | The Manufacturer, reporting a government-backed adoption program |
| Hershey Entertainment & Resorts, hospitality and entertainment; 10,000+ team members | Staff calling off a shift, reporting they'll be late or asking for their schedule. "Agnes", a voice agent built on Google's Customer Engagement Suite, identifies the caller by caller ID and uses webhooks to read and update internal systems. A web chatbot also sends invoices: it asks for an email address, checks for a stay in the last 60 days and sends the PDF. | More than 25% of call-offs fully automated, which the company puts at eight hours of staff work a day. More than 50% of password-reset calls handled. | The webhooks into scheduling and account systems. The agent is only as useful as what it can look up and change. | Google Cloud story (vendor-published) |
| Carbon Underwriting, Lloyd's of London syndicate (UK); platform built by "a lean team of just three engineers" | Categorizing claim descriptions and occupancy data that arrive in thousands of inconsistent spreadsheets. Gemini models inside the company's Graphene platform on BigQuery. | Categorization pipelines deployed "in a matter of days", work the company says would previously have taken months of manual data science. The summary also reports data processing time cut "from days to hours", without tying it to the AI step. | A common data model that every spreadsheet maps into, and an audit trail for each transformation, which its Graphene Studio tool keeps. | Google Cloud story (vendor-published) |
| Qatari Ministry of Labour, government (Qatar) | Matching expatriate university graduates with private-sector jobs through the "Ouqoul" platform. It links AI candidate matching to the ministry's services for contract authentication and work-permit issuance. | No figures published. The entry says the platform has launched. | The links into the ministry's own contract-authentication and work-permit services. Matching is one step; those are where it changes a record. | Google Cloud use-case list (vendor-published, entry only) |
Hershey's method is worth copying. Before building Agnes, the team ran its contact-center data through Google's Conversational Insights, looking at all calls rather than a sample, and found that staff call-offs, late notices, schedule requests and password resets were frequent reasons for calls. Each is a lookup or a small change to one record, which makes it a good first target.
This week: export 90 days of one back-office inbox or phone line and tag each item with its request type. Our rule of thumb: a type that is at least a fifth of the volume and ends in one record change, such as a shift marked absent or an invoice posted, is a candidate. A type that needs someone's judgment every time isn't one yet.
Customer service: calls, chats, emails and routing
The customer-service rows split into three jobs. The AI writes something a person used to type, answers a repeat question before it reaches a person, or decides where a message goes.
| Organization | What was automated, and how | Result, as the source reports it | What needed work beyond the tool (our read) | Source |
|---|---|---|---|---|
| Definity, property and casualty insurer (Canada); parent of Economical Insurance and Sonnet Insurance | Call-summary notes, which took agents three to five minutes a call, and caller authentication. An open-source transcription model runs on Dataflow, the transcript goes through Cloud DLP (Google's service for finding and masking sensitive data), an LLM on Vertex AI writes the summary, and the summary is stored in Salesforce. Authentication uses Dialogflow CX. Built with a consulting partner. | "Upwards of 3.5 minutes per call, or 33% of average handle time per interaction", on more than 500 calls a day. The first result came "within about a month". | The sensitive-data step before the summarizing model sees a transcript, and the write-back into Salesforce, where agents already work. | Google Cloud story (vendor-published) |
| Starling, digital bank (UK) | Summarizing call transcripts, answering help-center questions from the bank's customer-facing content, and "Spending Intelligence", where Gemini turns a customer's question into a structured function call to the bank's analytics engine. | 8,000 hours a month saved on transcript summaries. Customer-service referrals down 50%. | The analytics engine produces the figure and the model only translates the question. It's the same principle that fixed wrong numbers in our text-to-SQL case. | Google Cloud story (vendor-published) |
| SIGNAL IDUNA, health insurer (Germany); 10,000+ employees | An agent that reads incoming customer messages and forwards each one to the right department, built by its app team with Gemini Enterprise. A knowledge agent, Co SI Health, covers thousands of policy documents for more than 600 tariffs, with Document AI extracting and structuring them. | Routing error rate down 40%. Employees find answers in the policy documents in under 20 seconds. | Structuring a century of policy documents before any model reasons over them. The routing figure needs a labeled sample of correctly routed messages to mean anything. | Google Cloud story (vendor-published) |
| Sullivan County, New York, county government (US); 78,000+ residents, 40+ departments | Repeat questions on the county website, such as opening hours, paying taxes and renewing a driver's license. A Dialogflow virtual agent, built with a Google Cloud partner in just under three months. | The DMV saw "more than 56% reduction in call volume in the last four months compared to the same time last year". | A maintained answer set for each department. The page's summary turns the DMV figure into a county-wide one (see the case-study checks below). | Google Cloud story (vendor-published) |
| West Sussex County Council, local government (UK); serves 890,000 residents | A Dialogflow chatbot that screens residents' eligibility for adult social care and benefits, plus 14 generative AI agents that help staff search policy guides, legislation and procedures. | Only about 5% of inquiries qualify. The bot helps residents learn their status quickly, "reducing pressure on the Customer Service Center". No figure. | Eligibility rules written precisely enough for a bot to apply, and a direct route to a person for the 5% who qualify. | Google Cloud use-case list (vendor-published, entry only) |
| Sports Basement, outdoor retailer (US) | Customer-service emails drafted with Gemini in Google Workspace. | Time spent writing emails down 30-35%. | Nothing to integrate. Staff still own each reply, so this is the cheapest rung to try. Sending without a person needs order data and exception rules. | Google Cloud use-case list (vendor-published, entry only) |
| Ergo Global, ergonomics services (our client) | Triage and routing across website chat, internal app chat and email inboxes, with answers grounded in approved content. Built by Adamant Code on Chatwoot, with Microsoft Outlook integration and knowledge-base grounding. Pricing questions, bugs, unhappy users, unknown answers, sales opportunities and anyone who asks for a person are handed to staff. | Over five months, the support platform managed 11,782 conversations across seven support inboxes. That is a platform total, not an AI-resolution count. Website chat has a typical first response of 36 seconds. | The handover rules. Before, chat depended on two people noticing each message. | Our case study |
Definity's numbers are easy to check: 3.5 minutes on each of 500 calls is 1,750 minutes, about 29 agent-hours a day, and both inputs are lower bounds. The order of its pipeline is the part to copy: transcribe, screen the transcript for sensitive data, summarize, then save the summary where agents already work.
This week: ask your support lead for last month's tickets or calls with their contact reason and handling time. Add up the after-call work, meaning notes, tagging and forwarding, separately from the conversation itself. At Definity, summaries alone were worth 33% of average handle time. If your after-call work is anywhere near that, our default is to automate summaries and routing before any customer-facing bot.
Documents: verification, search and review
The document rows differ in where the person sits: only on the documents that fail the check (Covered California), on the decision the evidence supports (TCE's engineers, the banks that receive Oper Credits' files), or in the middle of the pipeline (Consult).
| Organization | What was automated, and how | Result, as the source reports it | What needed work beyond the tool (our read) | Source |
|---|---|---|---|---|
| Covered California, state health insurance marketplace (US) (proof of concept) | Verifying residents' identity and income documents: 50,000 a month in 56 document classes, 71% of them validated by hand. Document AI, with a consulting partner. | In the proof of concept, used unconfigured, 80-96% of documents were verified automatically depending on type, 84% on average. The legacy tool managed 28-30%. The goal is above 95%. The story says residents get instant verification "once this solution goes live". | Results by document class, because an average hides the classes at the bottom of the range. The documents carry personal data, so compliance controls (Assured Workloads) were part of the design. | Google Cloud story (vendor-published) |
| Oper Credits, mortgage process platform (Belgium); 45 employees, integrated with about twenty banks in six countries | Checking borrowers' bank statements and pay slips in mortgage files with Vertex AI models. | Documents are "analyzed almost instantly", a job that "previously required several hours of manual work". In Belgium, 60-70% of applications are returned by bank validation teams for missing or incorrect information. The goal is 90% complete on first submission "in the long term". | Each bank's definition of a complete file. That is policy, not extraction, and it differs by lender. | Google Cloud story (vendor-published) |
| TCE, infrastructure and engineering (US) (our client) | Searching bid-document packages that can reach 5,000-50,000 pages. Answers carry source citations, page references, exact paragraphs and highlighted source text, with project, team and role permissions inside TCE's Azure environment. Built by Adamant Code. | In production and still used. Estimated search time fell from about 60 to about 10 minutes per engineer per working day, including time spent reading the AI's answers: about 83%. The estimate is based on engineer interviews and client estimates, not formal usage logs or a controlled pilot. | Permissions and citations an engineer can check. How to test this kind of search on your own files is in our guide to AI document search for engineering bids. | Our case study |
| Consult, from i.AI, a UK government AI team (pilot) | Analyzing public consultation responses. AI proposes themes, officials select, edit or add themes, and then AI assigns the final themes to each response. Open source. | No figures in the repository. The README calls it an incubation project in a research stage, trialed across the Civil Service and not yet recommended for critical use. | Review is a step in the workflow, not bolted on after it. Input workbooks are validated before the pipeline runs, and each component has its own evaluation pipeline. | GitHub repository (open source) |
Covered California's first move is the one to copy. The team ran Document AI out of the box, with no configuration, to get a baseline for each document type before any tuning. The arithmetic shows what the result would mean for staff. Before, 71% of 50,000 documents, 35,500 a month, were checked by hand. At 84% automatic, 16% still reach a person: 8,000 documents a month. That is 77% fewer hand checks, not zero.
This week: pick your three most common document types and pull 20 real examples of each, including poor scans. Run whichever tool you're considering on them with no configuration and record the result per type, not overall. Our integration guide covers where the documents that fail should go and who works them.
Sales and orders: quotes, order entry and order questions
Order rows touch money and the ERP, so they succeed or fail on the connectors and on the rules for mismatches.
| Organization | What was automated, and how | Result, as the source reports it | What needed work beyond the tool (our read) | Source |
|---|---|---|---|---|
| Danfoss, manufacturer of energy-efficient solutions (Denmark); operates in more than 100 countries | Entering B2B orders that arrive by email, as free text, tables and PDF, image or spreadsheet attachments. Go Autonomous's Autonomous Commerce platform, with ERP and CRM connectors, triggers the actions in the ERP. | Its head of digital business says "above 80%" of these transactional decisions are now made by an AI agent; the results summary says "up to 80%". The team estimates about five minutes saved per order. Average response time improved "potentially" from 42 hours to near real time. Staff use one interface instead of up to five systems. | The ERP and CRM connectors, and the cases the story names: pricing mismatches, missing part numbers and company-specific workflows. Pricing exceptions still go to staff. | Google Cloud story (vendor-published) |
| Enpal, solar panel supplier | Generating quotes for prospective customers, including assessing roof size and the number of panels, with a Google Cloud partner. | Time per quote down 87.5%, from 120 minutes to 15. | Roof assessment is the hard input. The quote still has to reach the sales system with current pricing. | Google Cloud use-case list (vendor-published, entry only) |
| Elanco, animal health; 14,000+ users across 60+ locations | Sales reps' questions about customer orders, which meant 17,000 contacts with service centers in 2023. The company estimates those cost $250,000 in time. Elanco.ai, a Gemini-based tool, answers from SAP order data in BigQuery through Google Cloud Cortex Framework. | Answers in about 10 seconds. Across Elanco.ai, an estimated 70% less time on routine tasks and an estimated $2.3M in first-year efficiencies. It cost less than $300 a month to run at first. | Getting SAP order data into a warehouse the AI can query. The model sits on top of that plumbing. | Google Cloud story (vendor-published) |
| Bunnings, home-improvement retailer (Australia) | Helping online customers plan a project and pick the components. "Buddy" is built on Gemini Enterprise for CX, with safety guardrails codified so it keeps customers from attempting dangerous jobs themselves. | More than double the online conversion rate for transactions with Buddy, against a baseline the story doesn't give. Built in six weeks, after earlier groundwork with Vertex AI Search (now Agent Search). | The guardrails are business rules someone had to write down: a list of jobs the assistant must not coach a customer through. | Google Cloud story (vendor-published) |
Danfoss explains why it chose a platform built for order handling rather than prompting a general model. Customer emails mix free text, tables and attachments, and handling them means spotting pricing mismatches, missing part numbers and Danfoss-specific workflows. Our read: the same holds if you build. Most of the work is the checks against your price list and item master, not reading the email.
This week: count how often your sales team asks someone else about an order, whether by call, email or chat message, over one month. Elanco's 17,000 contacts a year is the kind of number that justifies connecting order data to a search or AI tool before automating order entry itself.
Operations: inspection, store ordering and the till
The operations rows run on data from the physical world: sensors, cameras and sales history. Each needed a data pipeline or a training set built first.
| Organization | What was automated, and how | Result, as the source reports it | What needed work beyond the tool (our read) | Source |
|---|---|---|---|---|
| Talgo, train manufacturer and maintainer (Spain); the story lists 80 trains in Spain and 36 in Saudi Arabia | Monitoring trains from 2,000 on-board sensors, about 2 GB per train per day, feeding machine-learning models for predictive maintenance, some still in development. "TALVI" camera inspection checks for loose or broken parts as a train passes. The Saudi maintenance team and the Spanish headquarters watch the same real-time data. | "Potential savings of over 200 hours a year per train in maintenance inspection time." This is an older story: it describes deliveries due by 2021. | The pipeline from moving trains to one data store, before any model. | Google Cloud story (vendor-published) |
| Fortenova Group, food production and retail (Croatia); 45,000+ employees | A daily forecast for fruit and vegetables from historical sales, promotions and inventory data, which store managers use to place orders. Built by the group's DataLab, set up with a data consultancy, on BigQuery and Vertex AI. | In a pilot at 65 Konzum stores, food waste in the category fell 7% and profit rose 8%. Since then, 500 Konzum stores use the recommendations. | A group-wide data platform came first: building it was one of the DataLab's first tasks. The model recommends; the store manager still places the order. | Google Cloud story (vendor-published) |
| Lush, cosmetics maker and retailer; Lush Lens expanded from Japan across the UK and Ireland | Identifying unpackaged products at the till by camera. Lush Lens runs on mobile point-of-sale devices, with models trained on Vertex AI. | About 800 products can be scanned. At the Glasgow store's Christmas peak, queues went from "out-the-door" to about 3 minutes. | The training set: 550,000 images, hundreds per product from several angles, and new products added as the range changes. | Google Cloud story (vendor-published) |
Talgo's inspection number can be checked from its own inputs. A high-speed train inspection takes six hours, and camera inspection would move depot visits from weekly to every three weeks. That is 52 × 6 = 312 hours a year, down to (52 ÷ 3) × 6 = 104, a saving of 208 hours per train, which matches "more than 200". The quote is in the future tense ("will mean", "would save"), and the summary calls the saving "potential".
This week: for any operations candidate, write down where the input data lives today and how old it is when someone acts on it. If the answer is a spreadsheet updated weekly, the first project is the data feed, as at Fortenova and Talgo, not the model.
How to check an AI case study before you repeat its number
Several stories above say one thing in their summary and something narrower in their text. Run these six checks before a case study's figure goes into your business case.
| Check | Where to look | Example from this list |
|---|---|---|
| Is the headline figure achieved, or a target? | Compare the results summary at the top with the quotes in the body. | Oper Credits: the summary lists "90% of loan applications complete and compliant on first submission"; the CEO's quote calls 90% the objective "in the long term". Danfoss: the summary says "up to 80%" of decisions are automated, the quote "above 80%". |
| How much of the business does it cover? | Find the sentence the number came from. | Sullivan County: the summary says citizen inbound calls fell 56% year over year; the body says the DMV, over four months. Lush's 3-minute queue is one store at Christmas. Carbon Underwriting's "days to hours" sits in the summary with nothing tying it to the AI step. |
| Measured, estimated or forecast? | Hedge words: estimated, expected, potentially, projected, up to. | Elanco's page title says "saving $2.3M"; its body calls it an estimated $2.3M in efficiencies. Danfoss "potentially", Nordell "expected", Talgo "potential". |
| A share of what, compared with what? | The denominator and the baseline. | Hershey's 25% is of call-offs, one call type. Ergo's 11,782 conversations were managed by the platform, not resolved by AI. Bunnings' doubled conversion is for transactions with Buddy against an unstated baseline, and shoppers who choose to use an assistant may be readier to buy anyway. |
| Pilot or production? | Stage words: proof of concept, pilot, trial, rollout, "once live". | Covered California: a results bullet says it "automatically validates 84% of documents in real time"; the body says the 84% came from an unconfigured proof of concept and that instant verification comes "once this solution goes live". Fortenova's 7% came from a 65-store pilot before the 500-store rollout. |
| Who measured it, and how? | Method: system logs, a time study, interviews. | TCE's 83% is based on engineer interviews and client estimates, not formal usage logs or a controlled pilot, and we say so wherever we use it. |
If a story matters to your decision, send the vendor this before you rely on it:
Subject: Questions about your [customer] case study before we rely on it
Hello [name],
We're assessing [task] and read your story about [customer].
Before we use its figures in our own business case, could you confirm:
1. Is the headline figure measured in production, or a pilot result,
estimate or target? If measured, over what period and how
(system logs, time study, interviews)?
2. What does it cover: all [orders / documents / calls], or one type,
one department or one site?
3. Of the items the system handled, what share went through with no
person touching them, and what share was reviewed or corrected?
4. What happens to an item the system can't handle? Who sees it,
in which tool, and how quickly?
5. Which systems does it read from and write to, and who built and
maintains those connections?
6. Could we speak to the person at [customer] who runs the process
day to day, not only the project sponsor?
Thanks,
[name]Our rule: a vendor that answers questions 3 and 4 with numbers has usually run the thing in production. If the reply is vague on both, treat the case study as a demo.
Scope your own first automation this week
The rows with the clearest results started from one request type with a known volume and a clear place for the result to land. Fill this in for your best candidate before you talk to any vendor, us included.
- Pick one request type from a 90-day export like the one described in the back-office section, named the way the person asking would name it.
- Pull 50 real examples of it, including ones that went wrong.
- Fill in the worksheet with the person who handles them today.
- Work out the hours that stay with people using the formula below, with a straight-through share you have measured on your own examples, not the vendor's.
| Question | Your answer | Example from this list |
|---|---|---|
| The one request type | Staff calling off a shift (Hershey) | |
| Monthly volume, and its share of all requests | 50,000 documents a month (Covered California) | |
| Minutes per item by hand today, and who does it | 3 to 5 minutes of notes per call (Definity) | |
| Where the input arrives, and in what form | Email with free text, tables and PDF or spreadsheet attachments (Danfoss) | |
| What the output changes, and in which system | Summary saved to the Salesforce record (Definity) | |
| What counts as an exception, and who takes it | Pricing questions, bugs, unhappy users, unknown answers, sales opportunities and requests for a person go to staff (Ergo) | |
| Data to remove or protect before a model sees it | Transcripts pass through Cloud DLP before summarizing (Definity) | |
| How you'll measure the result, and over what period | Four months against the same months a year earlier (Sullivan County DMV) |
The hours that stay with people depend on the share of items that go straight through:
People-hours per month = volume × (1 − straight-through share) × minutes per exception ÷ 60
In a sheet, with volume in B1, straight-through share in B2
and minutes per exception in B3:
=B1*(1-B2)*B3/60With Covered California's proof-of-concept figures and an illustrative four minutes per document (the story gives no handling time), 50,000 × 0.16 × 4 ÷ 60 is about 533 hours a month, against about 2,367 hours for the 35,500 documents checked by hand before.
Count the straight-through share in whole documents. A document goes straight through only if every field on it is right, so its rate is never higher than the per-field accuracy a vendor quotes, and usually lower.
If you have a candidate, bring the filled-in worksheet and your 50 examples, including the ones that went wrong. We'd start with the exceptions and the system the result has to be written back to, because that is where most rows above needed work beyond the tool. Our internal AI systems work starts there.