Field cases · Evidence checked 3 November 2026
What happened after the demonstration?
Four real cases to read beside the book. Look at the work, the result and the information still missing before you borrow a headline for your own plan.
These are public-source cases, not my client results. A company report, a user survey and a controlled experiment answer different questions. I have kept those differences visible. The commercial lessons are my interpretation; they are not claims made by the companies.
These are independent learning cases based on public sources. Chaplabo is not affiliated with, sponsored by or endorsed by the organisations named. Company and product names belong to their respective owners.
01 · Service redesign
IKEA / Ingka
Reported savings and a separate sales channel.
02 · Partner delivery
VOXI + Accenture
What the implementer actually had to build.
03 · Adoption
UK government
What 26 minutes a day can tell us.
04 · A measured limit
BCG consultants
Faster work can still produce a wrong decision.
01 · Customer-reported operating results · 2021–2023
IKEA: connect automation with the next useful job
Ingka Group, the largest IKEA retailer, reported that its Billie chatbot resolved about 47% of the enquiries it received from 2021 to 2023: 3.2 million resolved interactions and nearly €13 million in savings. It also reported reskilling 8,500 call-centre co-workers. Its June 2023 account describes AI and natural language processing; it does not identify Billie as a generative AI deployment. Read Ingka’s report.
Keep the measures separate. The 47% denominator is enquiries received by Billie. It is not every IKEA customer contact. Ingka also reported €1.3 billion of remote customer meeting-point sales in FY22, ending 31 August 2022. This is channel sales, not chatbot revenue, profit or proof that Billie caused those sales. The report provides no full cost reconciliation or supplier fee.
Who does the work? Ingka describes the customer-service change and staff development. It does not name an external implementation partner or technology supplier. I would not add one to the story. For a partner considering a similar offer, the opportunity is to connect routine support, human handover and staff development in one delivery plan.
My commercial reading. A saving becomes useful when somebody decides what to do with the capacity. That may mean serving more customers, improving advice or reducing a real cost. Those choices need different measures. A training service could be valuable, but these figures do not establish its price or margin.
Read beside Chapters 3, 16 and 19. Continue with service economics or the renewal lesson. Explore the manufacturing teaching example.
From this case to your work
Learn the method. Check your decision.
Use these lessons to keep the figures clear and decide what a similar service should do next.
- Keep revenue, contribution and cash separate — Check the base, period and costs behind a figure. Reported savings and channel sales do not tell you a partner’s contribution or cash position.
- Make renewal a service decision — Review the customer’s result, correction effort and available people before deciding whether to continue, change or expand support.
Two questions. Choose one answer for each, then check the reasons. Your answers stay in this page and do not change your eight-lesson progress.
Read answers and explanations
1. Which reading of Ingka’s figures is supported by the case?
A. Why this answer does not fit. The 47% covers enquiries received by Billie. The report does not use every IKEA customer contact as its base.
B. Why this answer does not fit. This was channel sales, not profit. The report does not show that Billie caused those sales.
C. Correct. The figures cover different results. Keep their scope and periods visible; neither provides a complete cost or partner-margin calculation.
2. A partner is reviewing ongoing support for a similar service. What should guide the renewal decision?
A. Why this answer does not fit. A high number of handled enquiries does not show whether customers get the result they need or how much human work remains.
B. Correct. This connects the customer’s result with the work needed to keep delivering it. It can support a decision to renew, improve or narrow the service.
C. Why this answer does not fit. Released time may be used for other work. Check skills and support commitments before promising more delivery.
02 · Customer launch + implementer account · From March 2024
VOXI: the partner’s work sits around the model
Vodafone’s youth brand VOXI announced its chatbot with Accenture on 7 March 2024. The initial release covered a small customer group so issues could be found before wider use. That is a launch decision, not a measured return. Read the customer’s announcement.
Who does the work? Accenture’s implementation account names Microsoft Azure AI Studio and Azure OpenAI. Development took about three months. Accenture built a question filter for VOXI service topics, tested different customer moods, refined the brand voice, and added alerts and a shutdown switch. VOXI supplied the service and brand context. Microsoft supplied the platform. Read the implementer’s case.
What result is public? Accenture reports higher containment and lower agent handling time in the first six months after launch. Here, containment means queries resolved without a human agent. The account gives no before-and-after rates, sample size or cost per resolved query. It therefore supports a direction of change, not a percentage saving. It gives no delivery fee, partner margin or renewal result.
My commercial reading. The three-month build gives us a better discussion than the phrase “add a chatbot”. Someone had to decide which questions belonged in the service, how the answer should sound and when to stop it. Those are pieces of delivery work. A partner can describe, estimate and price them before asking the customer to commit.
Read beside Chapters 6, 11 and 12. Use the paid-pilot lesson and the authority lesson. Explore the Delta Parts teaching example.
From this case to your work
Learn the method. Check your decision.
Use these lessons to define a useful pilot and agree who controls the service when it is running.
- Sell a first step you can finish — Agree a baseline and acceptance checks before wider use. Measure the complete task, including the human effort that remains.
- Give the service only the authority it needs — Enforce the agreed limits, name who can stop the service and plan how to deal with effects that have already happened.
Two questions. Choose one answer for each, then check the reasons. Your answers stay in this page and do not change your eight-lesson progress.
Read answers and explanations
1. Accenture reports higher containment and lower agent handling time, but gives no before-and-after rates. What can you conclude?
A. Why this answer does not fit. The rates, sample size and cost per resolved query are missing. You cannot calculate a percentage saving from the public account.
B. Correct. It gives a direction of change. A buying decision still needs a baseline, measured results and the complete handling effort.
C. Why this answer does not fit. Build time is not a profit calculation. The account gives no delivery fee, partner margin or renewal result.
2. A similar chatbot starts answering outside its agreed service scope. What should the support plan require?
A. Correct. The agreed scope needs controls and an owner. Stopping further answers does not correct answers already given.
B. Why this answer does not fit. A confident answer does not give permission. The application and the agreed service must control what the chatbot may do.
C. Why this answer does not fit. A shutdown stops further work. Earlier answers may still need review, correction or a separate recovery step.
03 · Customer-led trial · September–December 2024
UK government: time saved needs a second question
The Government Digital Service ran a Microsoft 365 Copilot trial from 30 September to 31 December 2024, providing 20,000 employees with licences. Its report, published on 2 June 2025, received 7,115 user survey responses. The headline was an average 26 minutes saved per day. This was self-reported time, calculated using survey range midpoints, not a stopwatch measurement of every participant. Read the findings and method.
Where is the limit? Seventeen per cent reported no clear time saving. The report could not establish how saved time was used. A full-year estimate is therefore an extrapolation from a three-month trial, not a year of observed cash savings.
Who does the work? Microsoft supplied the tool; GDS coordinated the trial and participating organisations collected evidence. The report does not identify one external delivery partner. For a partner planning a similar service, I would turn this into separate jobs: prepare access, choose tasks, help staff practise and review completed work with managers.
My commercial reading. Twenty-six minutes is a reason to investigate the work. It does not tell us how much money a customer can remove from its budget. I would ask what changed in the queue, turnaround time or service quality. Then I would compare that benefit with licences, setup, training, review and support. The public trial does not provide a complete project cost, partner contribution or renewal return.
Read beside Chapters 15, 18 and 20. Continue with finding useful work and measuring the economics. Compare an investment decision.
From this case to your work
Learn the method. Check your decision.
Use these lessons to find a clear task and test whether a time saving becomes a useful customer result.
- Find the work worth paying for — Start with one real task and its current process. Compare options using the same work, complete cost and remaining human effort.
- Keep revenue, contribution and cash separate — Name the measure and its cost boundary. A time estimate does not establish cash savings, contribution or money available in the bank.
Two questions. Choose one answer for each, then check the reasons. Your answers stay in this page and do not change your eight-lesson progress.
Read answers and explanations
1. What does the average 26 minutes saved per day mean in this trial?
A. Correct. Users reported time savings, and the calculation used survey ranges. The report could not establish how the saved time was used.
B. Why this answer does not fit. This was not a stopwatch study of every participant. The average also does not mean that every person had the same result.
C. Why this answer does not fit. The trial lasted three months. A full-year estimate is an extrapolation, and the report does not provide a complete project cost.
2. A team is considering a similar tool. Which first test would give useful buying evidence?
A. Why this answer does not fit. Time released does not automatically remove a cost or put cash in the bank. You need to know what changes in the work.
B. Why this answer does not fit. A selected demonstration does not give a fair comparison with the current process or show the work still needed.
C. Correct. Keeping the task and cost boundary the same helps the team see whether the tool improves a result worth paying for.
04 · Randomised experiment · 2023 work, 2026 journal publication
BCG: test the decision as well as the draft
A study with 758 BCG consultants compared no AI, GPT-4, and GPT-4 with a prompting overview after a baseline task. On 18 tasks selected within the model’s capability, AI users completed 12.2% more tasks and finished 25.1% more quickly on average. On a separate business problem, the combined AI groups were about 19 percentage points less likely to give the correct recommendation. These are experimental comparisons, not customer revenue results. Read the published study, especially pages 4–5 and 11.
What failed? The difficult task required a recommendation using interviews and financial information. Correctness meant choosing the right answer, not writing a persuasive response. This was one deliberately selected problem using a 2023 model. It does not establish the failure rate of today’s models or every consulting task.
Who does the work? BCG and the researchers designed and evaluated the exercise; consultants used GPT-4. There was no paying end-customer deployment in this experiment. The result is relevant to an advisory or delivery partner because its own staff can make the same mistake while preparing customer work.
My commercial reading. Faster drafting creates room for a better service only if the final decision meets the required standard. I would price independent checks where an error matters. The study gives no implementation contract, support cost, partner profit or renewal rate. We cannot turn its productivity percentages into a margin forecast without our own cost and quality evidence.
Read beside Chapters 11, 15 and 20. Work through pilot acceptance and practical authority. Review the partner portfolio decision.
From this case to your work
Learn the method. Check your decision.
Use these lessons to test important failure cases and keep actions under the right control.
- Sell a first step you can finish — Use representative tasks and important failure cases as acceptance checks. Faster drafting alone does not prove that the complete task succeeds.
- Give the service only the authority it needs — Separate advice from permission to act. A convincing recommendation still needs the agreed checks and application controls before an action is allowed.
Two questions. Choose one answer for each, then check the reasons. Your answers stay in this page and do not change your eight-lesson progress.
Read answers and explanations
1. Which conclusion fits the BCG experiment?
A. Why this answer does not fit. Speed and correctness are different measures. The AI groups were less likely to give the correct recommendation on the separate business problem.
B. Why this answer does not fit. The experiment used a 2023 model and selected tasks. It cannot establish the failure rate of today’s models across all work.
C. Correct. The result shows why task choice and checks for wrong decisions matter. The productivity gains do not remove that limit.
2. A partner is testing an advisory service that could change customer records. Which plan best follows this case?
A. Correct. Failure cases test the decision, not just the writing. A reviewer checks the evidence, and application controls enforce the agreed authority.
B. Why this answer does not fit. A convincing answer may still be wrong. Its wording cannot give the service permission to act.
C. Why this answer does not fit. The complete task includes review, correction and exceptions. Faster drafting may still produce a poor decision or move work to another person.
Bring one question back to your plan
Which part of the result would you need to prove for your own customer? Write the baseline, the person who accepts the result and the continuing cost. Then use the cash decision tool to work out whether the offer is worth repeating.
Sources checked 3 November 2026. Historical periods stay attached to their figures. Open the source registry for dates, definitions and limits.