Which calls to start an AI call center with: a selection matrix
Starting with the most frequent call looks attractive, but it often fills the pilot with escalations. This guide shows how to list your call types, a 1–3 scoring scale on five criteria, an illustrative medical-centre example, a rule for ties, the matrix's blind spots, and how to move from the matrix to a pilot plan.
The short answer
An AI call center project should start not with the most frequent call but with the call that will produce the best result. A good starting call has five properties: enough volume, it repeats in the same shape, the risk of a wrong answer is low, the information needed to answer is already written down, and a handoff to a person is rarely needed.
This article turns those five criteria into a simple scoring matrix. You score each call type from 1 to 3, add up the scores and pick the first two or three types for a pilot. The result rests on your own calls, not on gut feeling.
Why "the most frequent call" is the wrong start
The most frequent call is often the most complex one too. In a bank, card problems may be the largest share, but they require authentication, account data and sometimes a complaint process. Start there and the pilot fills up with escalations in its first week, and the team concludes that "AI is not for us".
By contrast, a call with moderate volume that is simple and repetitive — opening hours, address, a list of documents, a booking request — shows results quickly, earns the team's trust and produces learning material for the next stage.
Listing your call types
Before the matrix you need a list. Take a random sample of 100 calls from the last two to four weeks and write a reason next to each one. Agents' notes, CRM outcome codes or call recordings can all be the source. Then merge similar reasons — you usually end up with 8 to 15 call types.
- Information questions: hours, address, price, documents
- Booking and appointment requests
- Order or request status
- Change and cancellation requests
- Technical problems or service faults
- Complaints and dissatisfaction
- Sales interest: new customer, price quote
Five criteria and the scoring scale
- Volume1 — a few calls a month; 2 — dozens a week; 3 — arrives steadily every day.
- Repetition1 — every call is different; 2 — a common shape, details vary; 3 — same question, same answer.
- Risk (reverse score)1 — a wrong answer could cause financial, health or legal harm; 2 — it would annoy the customer but can be corrected; 3 — the impact of an error is small.
- Information availability1 — the answer lives in someone's head; 2 — a document exists but is old or scattered; 3 — an approved, written source exists.
- Escalation need (reverse score)1 — most calls need a human decision; 2 — some do; 3 — rarely.
Illustrative example: a medical centre
Not a real customer case. A medical centre sorted 100 calls by reason and scored five types. The maximum is 15.
- Opening hours and address: volume 3, repetition 3, risk 3, information 3, escalation 3 — total 15
- Appointment request: 3, 2, 3, 2, 2 — total 12
- Price questions: 2, 3, 2, 2, 3 — total 12
- Test results: 3, 2, 1, 1, 1 — total 8
- Complaint: 1, 1, 1, 1, 1 — total 5
The result is clear: the pilot starts with hours and address questions, while appointments and prices are added in the second wave once the knowledge base is in order. Test results and complaints stay with people.
What to do with a tie
If two types have the same total, look at information availability. If the answer is already written down, the pilot starts quickly; if not, knowledge-base work comes first. The second question: for which type is the outcome easier to measure? The result of an appointment request shows up in the CRM; the result of a "general information" call is hard to measure.
Who fills in the matrix and when it is refreshed
One person should not fill in the matrix alone. You need at least three views: the team lead who hears calls every day scores volume and repetition, the knowledge-base owner scores information availability, and the legal or quality lead scores risk. Where two scores differ by two points, listen to that call type together — the gap usually comes from how the call type was defined.
The matrix is not a one-off document. Fill it in again after the pilot's first month: the escalation score can change for the types the agent already answers, and as the knowledge base fills up, second-wave types move up. In a seasonal business, schedule the refresh a month before the peak.
What the matrix does not show
- Customer segment: if the same question comes from a VIP customer, the policy may differ
- Language: if part of the calls are in another language, that language needs its own check
- Seasonality: calls whose share changes during admissions, holidays or campaign weeks
- Hidden agent work: the call is short, but ten minutes of CRM work follow it
Write these points down as notes rather than extra columns. They do not change the score, but they set the pilot's boundary.
From matrix to pilot plan
- 1. Choose the first two or three typesThe highest-scoring types whose outcome can be measured.
- 2. Write a boundary for each typeWhat the agent says, what it does not, and when it hands over.
- 3. Record the baselineToday's volume, answer rate and average wait for that call type.
- 4. Define the fields needed after the callName, number, reason, outcome — what must appear in the CRM.
Common mistakes
- Choosing call types on agents' guesses without looking at a real sample
- Looking only at volume and ignoring risk
- Putting a topic with no knowledge base into the first wave
- Launching every type at once — it becomes impossible to tell which one works
Limits
The matrix is a simplified decision tool; the scores are subjective, and it is best if two people score independently and compare. A 100-call sample may miss rare but important cases. In health, finance and legal topics, the risk score should be set only after a specialist's review.
The selected call types in Vexvon
In Vexvon AI Call Center a scenario is written for each selected call type: what the agent will say and which fields it will extract from the call. Answers come from the knowledge base shared with the chatbot, so topics that score low on "information availability" get their knowledge base filled first. Difficult or out-of-scenario questions are passed to a live operator, and every call's transcript and summary stay in the panel so you can check the pilot's results.
The general framework is in what an AI call center is; to calculate what unanswered calls are worth, see the cost of missed calls.
First step
This week, write down the reason for 100 calls and fill in the five-criterion table. Types scoring 12 or more are pilot candidates. More articles are in the voice agent strategy section; to discuss your table with us, get in touch.