Start with your data, not your bot

In short. Most AI pilots never start, because the ambition was too big. Others look impressive but cannot change anything, because they depend on data the model could not reach or should not trust. Spend an afternoon on your data sources, then start with a use case that needs no integration and cannot embarrass you.

Usually one of three things is true. You do not have the data, it is not structured, or it lives in the experience of people who have been there for years. That last one is the most common, and the most flattering to ignore.

Where this article sits in the series

This is the third of four connected articles about moving from queue management to customer journey ownership.

By now two things should exist on paper. Part one gave you a week of logged contact causes and an owner with reach beyond the contact centre. Part two turned that into a channel map, plus a written list of the things you will never automate.

This part asks the question those two documents cannot answer on their own: is your data good enough to support the automated parts of that map. It also sets the two goals you will be reporting on for the next two years, one for cost and one for the customer. Part four takes those goals and builds the loop that keeps them honest.

An afternoon with your data sources

List every source that matters: interaction history, IVR and routing data, recordings and transcripts, CRM, knowledge base, and your order or ticket system. For each one, answer three questions. Is it complete. Is it current. Can an AI agent reach it through an API.

Three columns, six rows, done in an afternoon. It is usually enough to find the blocker, and the blocker is usually the knowledge base. Not the technology, but the fact that nobody owns it. Articles that were accurate two years ago are worse than no articles at all, because a bot will quote them with total confidence.

Ask the same three questions of your CRM. Is it complete, and does it give you a real 360-degree view of a customer, or a handful of fragments. If there is not enough in the system, AI will confidently give the wrong answer.

The market has reached the same conclusion. In DMG Consulting's 2026 research, reporting, analytics and business intelligence is the number one planned investment at 41%, ahead of AI infrastructure. Data quality is a joint top concern at 37.1%. Analyst coverage of the CCaaS market in 2026 lands in the same place: securing data pipelines and refining knowledge bases comes before autonomous service, not after it. A general-purpose language model dropped into a messy data landscape is not fit for most customer service work.

"A knowledge base that only exists in the head of someone with fifteen years of service is not a knowledge base."

It is also a single point of failure that nobody has on a risk register. The test is simple. If that person left next month, how much of your service quality leaves with them? That number is your real automation readiness, and it is usually lower than the platform demo suggests.

Start on the inside. Three use cases that need no integration

Our advice is usually to begin where the data problem does not block you. Easy cases, real impact, no customer risk.

Conversation summaries. People notice the time saved. The real win is consistency. You decide the format once, in the prompt. Bullets or prose, always three action points. Now you have records you can analyse instead of forty personal styles.

Understanding what people actually contact you about. Most organisations do not know. What they have is whoever has the loudest voice in the department saying this again, I get this all the time. Automated quality monitoring on every conversation shows the cause, not just the subject, which is the manual exercise from part one running continuously instead of for one week. If a large share of contact traces back to an invoicing problem, that is not a contact centre problem. That is a conversation with the invoicing department, and now you have the evidence to start it.

Coaching. The traditional version is a supervisor and three recorded calls every few weeks, and an agent who can always argue that you picked the wrong three. Listening to three calls every four weeks is a subjective sample. Looking at all of them gives you a trend, and a trend is something you can coach on. The point is to improve the agent, not to win the argument.

Pick a first customer-facing use case that cannot embarrass you

When you do go outside, use clear criteria: high volume, low variation, low emotional weight, data already available, and measurable. That usually points to order and delivery status, opening hours and simple FAQs, or rescheduling an appointment.

Cross-check the shortlist against the never-automate list from part two before anything goes live. If a candidate touches bereavement, vulnerability or safety, it comes off the list regardless of how well it scores on volume.

And use the cause log from part one first. If unclear invoices are your biggest cause of contact, the fastest win may not be automation at all. It may be one sentence changed on the invoice.

It helps to be realistic about where the market actually is. Gartner's headline prediction is that agentic AI will autonomously resolve 80% of common customer service issues by 2029. Measured today, industry roundups put actual self-service resolution closer to 14%, and only around one in five agents reports having generative AI tools available at all. The gap between the forecast and the floor is not a reason to wait. It is a reason to start narrow, prove it, and expand from evidence rather than from a slide.

One customer goal, next to the cost goal

Cost is a legitimate reason to invest. Right now it is almost the only one. In the DMG research, reducing operating costs and increasing scalability is the primary AI objective for 60.5% of respondents. Improving customer experience with personalised, AI-driven interaction comes sixth, at 28.7%.

If cost is your only KPI, in two years the only conversation left is margin. A cost-only business case optimises for containment, and containment is easy to fake. Deflect 40% of contacts and call it a win, while the 60% who escalate wait longer and repeat themselves twice.

So pick one customer metric you would defend in front of your board, and report it next to the savings from day one. First contact resolution on your top three contact reasons. Time to resolution for customers who do reach a human. Satisfaction measured specifically on automated contacts, not blended into the average. One number is enough. It keeps the business case honest when the second year of budget talks arrives.

Do this week

• Build the three-column list of your data sources
• Name an owner for the knowledge base
• Write down your cost goal and one customer goal, on the same page

Checklist, part 3

• Primary objective for AI stated explicitly
• One measurable customer goal alongside the cost goal
• Data sources reviewed for completeness, currency and API access
• Knowledge base has a named owner
• CRM checked for a real customer view, not fragments
• First use case chosen on volume, variation, risk and measurability
• Internal use case considered before anything customer-facing
• Checked whether an upstream fix beats automation

Next in this series

You now have two goals on one page and a first use case that is narrow enough to prove. Part four, Turn it on, measure it, and be willing to turn it off, covers what happens after go-live: the four things worth measuring, the rhythm of one change a month, and why switching something off counts as a success rather than a failure.

Read the full series

These four articles are written to be read in order. Each one produces something the next one uses.

• Part 1. Fix the mandate before you fix the queue. Who owns the problem, and how far does their mandate reach.
• Part 2. Know your customer, then pick your channels. Which contact reason belongs on which channel, and what you will never automate.
• Part 3. Start with your data, not your bot (you are reading this one). Whether your data can support what you want to automate, and where to start.
• Part 4. Turn it on, measure it, and be willing to turn it off. The loop that keeps the first three parts alive after go-live.

Source: DMG Consulting LLC, 2026 CX AI Playbook: Strategic Outlook and Investment Priorities, February 2026, sponsored by Five9. Additional market data from Gartner, Zendesk and public analyst commentary, retrieved August 2026.