An AI support case can go wrong before the first calculation. Ask how much the chatbot can automate, and you may end up maximising the number of conversations handled by software. The customer may still need to call a person to resolve the same problem.
I would start with the outcome the supplier wants to improve, then define what counts as a resolved enquiry. That gives the cost analysis something meaningful to measure.
Start with the client’s decision
Here is the fictional prompt:
A UK energy supplier is considering an AI chatbot for customer-service enquiries. It wants to know whether to roll it out.
You know the proposed action. You do not yet know what would justify it. Lower cost, shorter queues and better service could lead to different rollout choices.
A useful first question is:
“What would the supplier need the chatbot to improve for this rollout to be successful?”
The interviewer answers:
“Reduce support cost per resolved enquiry by 20% within twelve months, while maintaining customer satisfaction.”
You now have a commercial test and a service condition. A cheaper conversation that leaves the underlying enquiry unresolved will not satisfy that test.
Establish which enquiries belong in the pilot
Ask:
“Are there enquiries that must stay with a human adviser?”
In this case, billing disputes and customers needing extra support remain with people. Treat that as a client requirement when estimating how much work can move to the chatbot.
Define which customers and requests still need human support or another service channel. Those boundaries belong in the business decision. The exclusions in this worked example are assumptions of the fictional case.
A useful follow-up concerns the unit of measurement:
“If a customer uses the chatbot and then calls an adviser about the same issue, do both contacts count towards one resolved enquiry?”
Suppose the interviewer says yes. Include the cost of both contacts, but count the eventual resolution once. Otherwise the chatbot can look more efficient simply by creating additional interactions.
Quick Case Opening Drill
A regional water company is launching a chatbot pilot. You are scoping it before estimating any savings.
- Pilot target
- Chatbot resolves 60% of enquiries
- Must stay human
- Billing disputes and payment-plan requests
| Enquiry type | Share of enquiries |
|---|---|
| Meter readings and moving home | 45% |
| Payment-plan requests | 25% |
| Billing disputes | 30% |
What is the largest share of enquiries this pilot could take?
Translate the answers into a plan
You can now say:
“I’ll assess the total cost of resolving the same mix of enquiries, including repeat contacts and escalation. I’ll keep the excluded enquiries with human advisers and test customer satisfaction alongside the cost reduction.”
The next analysis should establish the current cost per resolved enquiry, the share suitable for automation and the cost of completing each route. You can then assess implementation costs and the time needed to reach the target.
This is a better starting point than a long discussion of model providers. A technically impressive tool still needs to improve the specified outcome.
Count the second contact
Assume the supplier currently resolves 10,000 distinct enquiries per month at a total cost of £60,000. The baseline is £6 per resolution, so the 20% reduction requires £4.80 or less at the same enquiry mix.
A proposed monthly model looks attractive:
| Route or cost | Assumption | Monthly cost |
|---|---|---|
| Chatbot-first enquiries | 5,000 at £1 each | £5,000 |
| Human-first enquiries | 5,000 at £6 each | £30,000 |
| Additional platform and oversight costs | Fixed monthly cost | £5,000 |
| Total before chatbot escalations | 10,000 enquiries | £40,000 |
The £1 chatbot cost includes the software interaction, but it does not include a subsequent human adviser. Suppose 2,000 of the chatbot-first enquiries need that additional help, costing £6 each.
The monthly cost becomes £40,000 + £12,000 = £52,000, or £5.20 per resolved enquiry. All 10,000 enquiries eventually resolve, so there is no reason to increase the denominator to 12,000. The proposal saves 13.3%, missing the 20% target.
This does not automatically rule out the tool. It identifies what must improve. At the stated costs, the model has room for £8,000 of additional human handling before reaching the £48,000 target. That allows about 1,333 escalations, or 26.7% of the 5,000 chatbot-first enquiries. The proposed 40% escalation rate is too high.
You have converted a vague promise of automation into a testable operational requirement.
Quick Case Math Drill
An online travel agency now sends every booking enquiry to a chatbot first. Customers it cannot help phone an adviser, who resolves the enquiry on that call.
- Chatbot cost
- £1 per conversation
- Adviser cost
- £8 per call
- Callers after chatbot
- 40% of chatbot-first enquiries
What is the average cost per resolved chatbot-first enquiry, in pounds?
Check when the target applies
“Within twelve months” can mean the cost rate achieved in month twelve or the average cost over the entire first year. Clarify that distinction before counting savings.
A system that reaches £4.80 per enquiry in its final month may still cost more over the year once setup, training and the ramp-up period are included. Ask which outcome management uses for approval, then show the other consequence if it is material.
The same discipline applies to satisfaction. Use comparable enquiries and include customers who were escalated or abandoned the chatbot. A survey of only the people who successfully finished the automated flow would miss the experience most likely to undermine the proposal.
Keep the opening focused
You do not need every escalation rate and cost before presenting your approach. In the opening, establish the outcome, the service boundary and the meaning of resolution. Request the detailed numbers when you reach the economics.
If the prompt already supplies those definitions, use them and move on. A question earns its place by resolving a gap.
Practise with the customer-support case-opening drill. After each answer, state one concrete change to your analysis. For a related denominator problem, try AI cost per completed task.
Case setup drill
How would an MBB interviewer judge your case opening?
Managed IT Services Provider Helpdesk Cost Overrun
Summarise the client’s brief and ask our AI interviewer clarifying questions. Get feedback on your opening, with examples of stronger questions.
Frequently asked questions
Opening the case and defining success
What should I ask first in an AI customer-support chatbot case?
Ask what the supplier would need to see to justify rolling the chatbot out. The prompt gives you the proposed action but not the reason for it. Lower cost, shorter queues and better service could each support a different rollout choice, so pin down the outcome before you estimate anything.
Why is “how much can the chatbot automate?” a weak opening question?
It can lead you to maximise the number of conversations software handles, which is not the same as solving problems. A customer who chats with the bot and then still calls an adviser about the same issue has not been served more cheaply. Measuring resolved enquiries keeps the analysis tied to the supplier's actual goal.
What counts as a resolved enquiry?
A resolved enquiry is a distinct customer problem that has been fully dealt with, whichever channel finishes it. Count enquiries, not chatbot sessions. If a customer starts with the bot and is then escalated to an adviser, that is one resolution with two contacts, and both costs belong to it.
Do I need every escalation rate and cost before presenting my approach?
No. In the opening, establish the outcome, the service boundary and the meaning of resolution, then request detailed figures when you reach the economics. If the prompt already supplies those definitions, use them and move on. Each clarifying question should earn its place by resolving a genuine gap, not by gathering numbers early.
Scope and service boundaries
Which enquiries should stay with human advisers?
Ask the interviewer rather than assuming. In this fictional case, billing disputes and customers needing extra support remain with people. Treat that as a client requirement, so those enquiries are taken out of scope before you estimate how much work could realistically move to the chatbot.
Can I treat the excluded enquiries as a footnote?
No, they are part of the business decision. Define which customers and requests still need human support or another service channel, and carry those boundaries into your estimate. Leaving them in would overstate the chatbot's potential share of work and could produce a recommendation that breaks the client's own requirements.
How do I turn the clarifications into a stated plan?
Restate the measure, boundary and test in one sentence. For example, say you will assess the total cost of resolving the same mix of enquiries, including repeat contacts and escalation, keep excluded enquiries with human advisers, and test customer satisfaction alongside cost. The interviewer can then see how each answer shaped your approach.
What analysis should follow the opening?
Establish the current cost per resolved enquiry, the share of enquiries suitable for automation and the cost of completing each route. Comparing routes on a completed-resolution basis shows whether the chatbot genuinely lowers cost once handovers to advisers are included, rather than simply looking cheap per conversation.
Should customer satisfaction matter if the brief is mainly about cost?
Yes, test it alongside the cost reduction. Lower cost, shorter queues and better service are different outcomes, and the supplier may care about more than one. Keeping satisfaction in the plan stops a cost saving from hiding a service problem, such as customers needing repeat contacts to get an answer.
The cost maths and the twelve-month target
How do I work out the target cost per resolution?
Divide the current total cost by the number of distinct resolved enquiries, then apply the reduction. In the example, £60,000 across 10,000 enquiries gives £6 per resolution, so a 20% reduction means £4.80 or less. Hold the enquiry mix constant, or a lower figure may just reflect easier enquiries.
Why can a chatbot-first model look cheaper than it really is?
Because a low cost per chatbot conversation ignores the customers who then contact an adviser. Counting only the first route makes the chatbot's share look inexpensive. Add the cost of the second contact for escalated enquiries, then compare the total cost per resolved enquiry with the £4.80 target before judging the model.
What does “within twelve months” actually mean?
It is ambiguous, so clarify it before counting savings. It could mean the cost rate achieved in month twelve or the average cost across the whole first year. Ask which outcome management uses for approval, lead with that measure, and then show the other so decision-makers see both views.
Could the chatbot reach £4.80 and still miss the target?
Yes, if the target is the first-year average. A system that reaches £4.80 per enquiry in its final month may still cost more over the year once setup, training and the ramp-up period are included. Those one-off and transitional costs are easy to miss if you only model the steady state.