How to Map Your Customer Journey and Find Where AI Helps

Pick one type of customer and list every step from their first search to their repeat visit. At each step, note what the customer wants, who handles it, how long it takes, and where people wait or drop out. Mark steps that are repetitive, text-heavy or out of hours as AI candidates. Keep judgement calls and bad news with people.

For one journey, this takes an afternoon, using evidence you already have: the phone log, the inbox, booking data and reviews. What you end up with is a short list of places where AI would help your customers, not just your staff, and an equally clear list of places it shouldn't go.

Follow me on Instagram@sagnikteaches

Choose one customer and one journey

A map of "our customers" turns into mush, because a first-time enquirer and a ten-year regular go through completely different steps. Pick one customer type and one journey with a clear start and end. Good choices are the journey that brings in the most money, or the one generating the most complaints and phone calls.

Connect on LinkedInSagnik Bhattacharya

For a veterinary practice, that might be "a new puppy owner, from first search to the end of the first year". For a dental practice, "a new patient from enquiry to first hygiene appointment". For a pharmacy, "a customer's repeat prescription, from request to collection". Map one now, and do another next month.

Subscribe on YouTube@codingliquids

Some businesses don't have the volume for the call tally described below. A kitchen-fitting firm might finish 30 jobs a year, each running three months from first visit to final snagging, so a week by the phone would log four calls and teach nothing. Rebuild the journey from the past instead: pull the complete email and message threads for ten finished jobs, lay each one out as a timeline, and mark the gaps. In a firm like that, friction shows up as silence more than volume, such as three weeks between the design visit and the quote, with the customer chasing twice. Quiet gaps where customers chase are good AI candidates: a status update drafted from the job system and checked by a person before it goes.

Gather the evidence before you draw anything

A map drawn from memory shows the journey as it's supposed to work. You want the one customers actually experience. Collect:

• A week of call reasons. Ask reception to tally every call by reason on a sheet by the phone. It takes seconds per call and it's the single most useful input.
• Fifty recent emails and messages, including web-form enquiries and social media messages.
• Booking data: time from first contact to first booking, no-shows, and how many customers come back for the next step.
• Reviews and complaints from the last year.
• Your website's most-visited pages, if you have analytics. They show what people look for before they contact you.
• A mystery shop. Ask a friend to enquire by phone at lunchtime, by web form at 9pm, and by social media message on a Sunday, and to note what happened and how long each reply took.

The tally sheet needs no design. Here is roughly what a week of it might look like on the front desk of a three-surgery dental practice (figures illustrative):

CALL TALLY - reception - week of 8 Sep        Mon Tue Wed Thu Fri  Total
Checking appointment time / is it still on     6   4   5   7   3     25
New patient: taking on? price of check-up?     5   3   4   2   6     20
Change or cancel an appointment                4   6   3   5   4     22
Toothache / pain - need to be seen             3   2   2   4   5     16
Payment or plan question                       2   1   3   1   2      9
Other                                          2   3   1   2   1      9
                                                                     101

Read it as a journey, and two stages jump out. A quarter of all calls are people checking something the practice already told them, which points at the reminder message, not the phone. And 20 new-patient calls a week, several of them asking only for a price, say the website's "find" stage isn't doing its job. The 16 pain calls are the opposite kind of signal: urgent, clinical and anxious, so they stay with a person whatever else changes.

The mystery shop matters because your data has a blind spot: customers who gave up. People who rang, got the engaged tone and booked elsewhere don't appear in your booking system. A friend's notes from mystery-shopping a two-therapist physiotherapy clinic might read:

Tue 12:40, phone:      rang 9 times, voicemail. No call back until
                       10:15 next day.
Tue 21:05, web form:   auto-reply "we'll be in touch". Human reply
                       Wed 16:30 (about 19.5 hours), asked me to ring.
Sun 11:00, Instagram:  reply Mon 09:10, friendly, gave booking link.
                       Booking page showed nothing for 12 days.

None of that would show up in the clinic's own records, because a real customer in the same position simply books somewhere else. Three contact routes, three different waits, and a booking page that ended the enquiry even when the reply was good.

Running the mapping session

Map with three or four people, not alone. The owner sees the business from the top; the people at the front desk see what customers actually say. A good group is the owner, one person who answers the phone, one practitioner who sees customers, and whoever handles complaints. Block out two hours:

1. 20 minutes: agree the customer type, the start and end of the journey, and the stage names.
2. 40 minutes: fill in the grid, stage by stage, from the evidence you collected. Where someone says "that doesn't happen", check the tally sheet before arguing.
3. 20 minutes: read out the reviews and mystery-shop notes that relate to each stage. This is where the customer's view comes into the room.
4. 20 minutes: fill in the AI fit column using the signs further down.
5. 20 minutes: pick two or three candidates, and decide who will write each one up.

Do it on a large sheet of paper or a shared spreadsheet on a screen, whichever your team is more comfortable with. The format doesn't matter; the evidence does.

The journey grid

Lay the journey out as rows, one per stage, with these columns. Most service businesses have seven or eight stages: find, enquire, book, prepare, visit, pay, aftercare, return.

Worked example: a new puppy owner at a veterinary practice

Suppose a small-animal practice maps the first year of a new puppy owner. The practice is hypothetical; each figure shows which piece of evidence it would come from.

Three things are worth noticing. First, two of the biggest fixes aren't AI at all: an up-to-date price page and an online registration form. A journey map finds those too, and they should come first because they're cheap. Second, the strongest AI candidates cluster at the edges of the visit, before and after, where customers have routine questions and staff are busy. Third, nothing clinical is on the "yes" list. When an owner messages at 10pm asking whether their puppy's vomiting is normal, the reply has to route them to the out-of-hours service, not attempt an answer.

The practice's shortlist became: after-hours enquiry answers with a booking link, the personalised new-puppy email, and better vaccination reminders, the approach covered in whether AI can handle a vet's vaccination and check-up reminders. The written visit summary went on a "later" list, because it touches clinical content and needs a careful trial; drafting owner-facing explanations is covered in writing pet owner emails with AI.

The new-puppy email shows why "vet-approved" belongs in that cell. Asked to draft it from reception's notes on the nine follow-up calls, an assistant produced a friendly, well-organised email. Three of its sentences (illustrative):

Your puppy can start going out for walks one week after the
first vaccination. Feed four small meals a day until six months.
Most pet insurers cover vaccinations, so check your policy.

Each needed a vet's pen. The walking advice depends on the vaccine and the practice's own protocol, so the vet replaced it with the instruction the practice actually gives. The feeding line was too general for breeds that grow at very different rates, and became "we'll give you a feeding plan for your puppy's breed at the first visit". The insurance sentence was a guess about other companies' policies, which the practice had no business making, and it came out. The email still saved reception the calls. It just couldn't go out on the model's say-so.

The same method works in very different businesses. In an independent pharmacy, mapping the repeat prescription journey (request, ordering, dispensing, ready notice, collection) usually shows the heaviest friction in one place: customers ringing to ask whether a prescription is ready. That step has every sign of a good AI candidate, since it's the same question hundreds of times a week, answered from a status that already exists in the system. The counselling conversation at collection has none of those signs, and stays with the pharmacist.

The signs a step suits AI, and the signs it doesn't

Go down the friction column and score each stage against these. A step with two or more "suits" signs and no "doesn't suit" signs is a candidate.

The last row on the right deserves thought. If regulars love that the receptionist knows their dog's name, automating that conversation costs you something no time saving will show. Where AI does reach customers, the risk to watch is covered in AI mistakes that damage customer trust.

Let AI help you read the evidence

Sorting fifty emails and a week of call notes by journey stage is tedious, and it's a job AI does well. Remove names and contact details first, and use a business plan that doesn't train on your content by default.

Below are [number] customer messages and call notes from a [type of
business], with personal details removed. Our customer journey stages
are: find, enquire, book, prepare, visit, pay, aftercare, return.

For each item, give: the stage it belongs to, the customer's reason
in five words or fewer, and whether it was a routine question, a
problem, or a request needing professional judgement.

Then summarise: for each stage, how many items, the three most common
reasons, and any item that suggests a customer nearly gave up.

[paste messages]

For the veterinary practice in the worked example, the first lines of the reply might look like this (illustrative):

#  Stage      Reason (5 words max)          Type
1  enquire    new puppy, are you registering routine
2  prepare    what to bring first visit      routine
3  prepare    puppy chewed chicken bones      routine
4  aftercare  second jab date                routine
5  pay        insurance claim form help       routine

PREPARE: 14 items. Top reasons: what to bring (5), feeding (4),
insurance (3).

Item 3 is the one to catch. A puppy that has eaten cooked bones is a question for a vet, not a routine one, and it belongs under "needs professional judgement". If that label went uncorrected, the "prepare" stage would look more automatable than it is, and a later auto-reply might treat bone questions as feeding advice. Correct the category, add a line to the prompt ("anything about symptoms, injuries, or something the animal has eaten is professional judgement"), and rerun the batch. Then compare the stage counts with the tally sheet. If they disagree badly, one of your two sources is wrong, and it's worth finding out which before you draw conclusions.

From map to shortlist

1. Take the stages marked "yes" and write each as a separate idea. "After-hours enquiry answers" and "new-puppy email" are two ideas, not one.
2. Put each through the one-page AI use case template to score volume, time, error cost and data readiness.
3. Decide whether to start on the customer-facing side or behind the scenes. The trade-offs are in customer-facing or back-office AI.
4. Before changing anything, record the numbers you'll judge it by: reply times, follow-up calls, second-vaccination bookings.
5. After launch, ask customers. The methods in measuring customer reaction to AI tell you whether the journey actually got better from their side.

A rough sum helps you choose which "yes" to start with. At the vet practice, the nine follow-up calls a week took about four minutes each: roughly 36 minutes of reception time, plus owners' irritation at having to ring. Late second vaccinations matter more. If the practice registers around ten puppies a month (illustrative) and one in five books late, two puppies a month are left less protected for longer, and those owners are the likeliest to drift away before the health-plan conversation. The after-hours enquiries matter most, because an enquiry left 20 hours is often one that registered elsewhere, and a new puppy means years of visits. So the order became enquiries, then reminders, then the email, even though the email was the easiest to build.

Then check each change against the map. Six weeks after the new-puppy email started going out, the practice reran the call tally for a week. In an illustrative rerun, follow-up calls from new owners fell from nine to three, but a new question came up twice: "The email says bring the vaccination card, and the breeder didn't give us one." That isn't a failure. It's the map working at the next level of detail, and it became one extra sentence in the email telling owners to bring whatever paperwork the breeder gave them, or nothing at all.

Where journey maps mislead

• Mapping the manual, not reality. "We reply to web forms within two hours" may be the policy. The mystery shop says otherwise.
• Only staff in the room. Staff know the friction they feel. Customers feel different friction. Reviews and the mystery shop bring their view in.
• Averaging different customers. A journey that mixes new and long-standing clients produces an average nobody experiences. A small gym that mapped "members" in one grid concluded that onboarding was fine, because most messages came from long-time members asking about class times. Split out, the first-month journey told a different story: new joiners who never booked an induction made up most of the cancellations in month two.
• Too many stages. Twenty rows means you've mapped your internal process instead of the customer's experience. Keep it to seven to ten.
• Stopping at the map. A beautiful grid that doesn't produce a shortlist with owners and dates was an afternoon spent drawing.

Further reads

• How to Map a Business Process Before You Automate It — Map the internal process behind a journey step.
• How to Analyse Customer Feedback Surveys With AI — Go further with customer feedback before you map.
• When Not to Use AI in Your Business: 9 Tasks to Keep Human — The steps that should stay human.
• AI for Veterinary Practice Owners: Costs, Tools, and First Steps — Costs and first steps for vet practices.
• How to Reduce No-Shows With AI Reminders and Automatic Rebooking — A common fix for aftercare drop-off.
• Should You Tell Customers You Use AI? A Disclosure Guide — Disclosure once AI appears in the journey.
• AI Use Cases by Department for Small Businesses (With Examples) — 21 practical AI use cases across seven departments, each shown in a small insurance brokerage, with how to start and what to watch.
• Automating a Broken Process: Why It Backfires and What to Fix — Why automating a broken process backfires, the five-pass fix to do first, and how to split the cleaned-up process between rules, AI and people.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.