AI Customer Journey Mapping

Why most journey maps are fiction

A typical journey map is drawn in a workshop: awareness, consideration, decision, with a persona called something like "Marketing Mary" and guesses about how she feels at each stage. It looks tidy and changes nothing, because nobody checked it against what customers did.

AI changes this by making it cheap to read the evidence you already hold: hundreds of call notes, emails, support tickets and CRM records that no one has time to read in full. Used well, it produces a map built from evidence. Used badly, it produces a more fluent version of the workshop fiction, because a model asked to "create a customer journey" will happily invent one.

The difference is the input. This guide is about giving AI real material and checking what it concludes.

Step 1: gather the evidence you already have

• • CRM records for recent won and lost deals: stage dates, source, notes, loss reasons.
• • Sales call notes or transcripts, where you have them and customers were told.
• • Enquiry emails and form submissions, which show the words people use when they first get in touch.
• • Search queries from your search console data and paid search reports, which show what people look for before they find you.
• • Support tickets from the first 90 days of customers, which show where expectations and reality diverge.
• • Reviews and public comments about you and competitors.

Aim for a balanced sample, for example 20 won and 20 lost deals, rather than everything. Lost deals are as informative as won ones.

Step 2: remove personal data before analysis

Call notes and emails contain personal data. Before putting them into any AI tool, remove names, email addresses, phone numbers and anything identifying, and check that the tool's data terms are acceptable to your business (in particular, whether inputs are used to train models and where data is stored). Under UK GDPR you also need to be comfortable that this analysis fits the purpose for which you collected the data. If in doubt, check with whoever handles data protection for you.

Step 3: ask AI for extraction, not invention

Give the model the redacted material and ask narrow questions, one at a time:

• • What questions did buyers ask, and at what stage (by CRM stage dates)?
• • What triggered the first contact, in the buyer's own words?
• • Where did lost deals stall, and what reasons were recorded?
• • How long did each stage take for won versus lost deals?
• • What did new customers complain about in their first 90 days?

Ask for quotations or record references with each finding, so you can check it. A finding the model cannot point to in the material should be treated as invented.

Step 4: draft the map as a table

A map in table form is easier to check and act on than a diagram. An illustrative template:

Journey map template, with illustrative entries

Stage	What the buyer is trying to do	Evidence	What we provide now	Gap	Owner and measure
Trigger	Work out whether a problem is worth solving	Enquiry emails mention a failed internal attempt	Nothing specific	No content on why internal fixes fail	Marketing: visits to a new explainer page
Search	Find out who solves it and what it costs	Search queries include cost and comparison terms	Service pages without cost guidance	No honest cost explanation	Marketing: enquiries mentioning cost
First contact	Get a quick, credible answer	Won deals had a reply the same day; lost deals often waited longer	Reply when someone is free	Slow first response	Sales: time to first response
Evaluation	Convince colleagues	Loss notes mention finance objections	One generic proposal	No material for the finance director	Sales: proposals with a finance summary
First 90 days	See early value	Support tickets about onboarding steps	An onboarding call	Unclear first steps	Account team: tickets in first 90 days

Step 5: check the map with real customers

Show the draft to five people: two recent customers, one lost prospect if willing, and two of your own salespeople. Ask one question per stage: "Is this what it was like?" Correct the map where they disagree. This step is what separates an evidence-based map from an AI-generated one, and it takes a few hours.

Step 6: fix one gap and measure it

Choose the gap with the clearest evidence and the most commercial weight, give it an owner, and pick a measure you can track before and after. Resist fixing everything; one fixed gap proves the method.

An example of this in AfM's published records is response speed. In the inbound processing record, a 2025 build Jakub delivered before founding AfM, leads had been waiting "1-3+ days" for follow-up while prospects were comparing alternatives; after automated enrichment and priority routing, response became "Same Day". A journey map built from CRM stage dates would surface exactly that kind of gap, because the time between enquiry and first reply is visible in the data.

Keep it alive

Rebuild the map from fresh evidence every six months, or when you change your offer or market. Keep the prompts and the sample definition so each rebuild is comparable. The map is a working document for decisions, not a poster. For how the handoffs on the map are owned, see AI for sales and marketing, and for the conversion rates at each stage, lead generation conversion rates.

FAQ

Do I still need personas?

Short, evidence-based descriptions of the roles in a buying group are useful. Fictional personas with invented hobbies are not. Describe what each role needs to decide and what they worry about, from your evidence.

How much data do I need?

Less than you might think. A balanced sample such as 20 won and 20 lost deals with notes is a sensible start; if the customer check in step 5 contradicts the map, widen the sample. More data helps only if it is of similar quality.

Can AI map the journey from analytics data alone?

Analytics show what people did on your website, not why, and they miss everything that happens in email, calls and conversations with colleagues. Combine them with what buyers said.