Nearly every enterprise has AI agents on its roadmap. Yet according to The Martech Weekly’s Enterprise Martech Outlook 2026, only around one in ten of those initiatives has been fully scaled in production.
So what’s holding AI implementation back? In a recent interview with Keanu Taylor, Global Head of Research at The Martech Weekly, Celebrus CEO Bill Bruno shared what he sees inside global enterprises every day: the problems that stall AI start with the customer data feeding it.
Here are five lessons from that conversation, and what they mean for teams trying to move AI from roadmap to results.
1. AI Exposes the Work That Used to Happen Behind the Scenes
Picture a quarterly business review. The slides are polished and the dashboards are up. Then an executive asks a question nobody anticipated: Which customers drove the drop? Did the campaign reach people who weren’t logged in? The room goes quiet, because the answer isn’t in any of the reports.
For years, that gap was covered by a familiar cycle:
- The scramble: someone leaves the meeting to find out whether the data even exists.
- The back room: analysts write queries, merge tables, and wait in IT queues.
- The delay: an answer arrives a week later and sparks the next round of questions.
AI collapses that cycle into a single conversation. Anyone can ask a question and expect an answer in seconds. That’s the promise, and it’s also the problem. When the data underneath is incomplete or disconnected, AI doesn’t hide it behind a week of analyst work. The gap shows up in the answer, immediately and in front of everyone.
Where to start: Before scaling AI, list the questions your executives actually ask, and check whether your data can answer them completely, including for customers who never logged in.
2. Siloed Teams Create Siloed Customers
Many enterprises have divisions that each solve their own customer data problem in their own way. Marketing bought one tool, digital bought another, and loyalty built a third. Each solution works locally. Together, they leave the customer in pieces.
That fragmentation shows up in predictable ways:
- Point solutions: they fix one team’s problem while creating a new one somewhere else.
- Best-guess matching: teams fill the gaps with data pools and cookie pools, which offer a guess instead of a recognized customer.
- Missed customers: a loyal customer browsing while logged out looks like a stranger, so the visit never connects to the history the brand already has on them.
Then AI gets asked a simple question, such as how many high-value customers visited this week, and every team’s data returns a different number.
Where to start: Get the teams in a room and solve for the whole customer, not one department at a time. Agree on how to recognize a customer in all three states: anonymous, pre-login, and authenticated. Then decide what AI should do with them.
3. Know Your Data Before You Trust AI with It
Responsible AI implementation starts with data ownership. Every data set carries three things:
- Source: where it came from and how it was captured.
- Biases: which customers and behaviors it over- or under-represents.
- Limitations: what it can’t tell you, such as activity before login.
If you don’t understand all three, combining that data with other sources gets risky fast.
Digital data deserves extra scrutiny, because so many different technologies capture it across websites and apps, each in its own way. Two platforms deployed the same way can return two different numbers. Teams with clean loyalty databases and warehouses are understandably wary of behavioral data that isn’t well-structured. And without knowing a data set’s limits, there’s no way to tell which questions AI shouldn’t be asked.
Where to start: Document the source, biases, and limitations of every data set before you connect it to AI. If you can’t answer all three, that data isn’t ready.
4. Define What Real Time Means for Your Business
“Real time” means something different depending on which vendor you ask. For some, it’s a few minutes. For others, it’s an overnight refresh. For a customer deciding whether to finish a purchase, it means right now.
Instead of chasing the label, answer two questions:
- Today: What does real time need to mean for the decisions your teams make now?
- Tomorrow: What will it need to mean once AI starts acting on customers directly?
A fraud check at login and a monthly churn report have very different timing needs. Matching speed to the use case keeps teams from overbuilding in one place and falling behind in another.
Where to start: Pick the one or two use cases where timing matters most, and build something practical that delivers on them. Trying to solve everything at once is how teams end up overwhelmed and doing nothing.
5. Solve Identity Before You Hand Decisions to Agents
Agentic AI, meaning systems that act on their own by building audiences, sending offers, or flagging risk, raises the stakes. Set-it-and-forget-it agents shouldn’t be on the table until identity is solved. Without it, you have bots acting on bad answers, at scale.
Consider an agent tasked with re-engaging lapsed customers. If it only recognizes people who are logged in, it may send a win-back offer to a loyal customer who was browsing logged out yesterday, while missing the customer who is quietly comparing competitors’ rates.
Before an agent makes a single decision, three things need to be in place:
- Recognition: they recognize the customer across anonymous, pre-login, and authenticated states.
- Timing: they surface what they know while the customer is still engaged.
- Connection: they link live behavior to the data they already trust, such as CRM and loyalty records.
Where to start: Before giving an agent decisions to make, measure how much of your audience you can actually recognize, especially before login.
Where Celebrus Fits in AI Implementation
Every one of these lessons points to the same starting point: AI can only be as good as the customer data it reaches. Celebrus captures first-party behavioral data from the first digital interaction and resolves identity in milliseconds, all within the brand’s own environment. Celebrus AI then runs on that same live, identity-resolved data, so every answer reflects the full customer journey.
The data underneath
- Identity from the first interaction: anonymous, pre-login, and authenticated activity resolves into one profile across devices, domains, and channels.
- Real time measured in milliseconds: identity resolves and data updates while the customer is still engaged, not after a batch refresh.
- Frictionless deployment: deployed once, capturing 100% of behavioral activity with no custom tagging or JavaScript, so nothing goes missing when a site or app changes.
- Your stack, finally complete: loyalty databases, CRM records, and warehouses only see customers after they log in. Native integrations with Teradata, Snowflake, Salesforce, Adobe, Pega, and more are fed live, identity-resolved behavior from every visitor, known or not.
How Celebrus AI puts it to work
- Ask the data directly: teams connect Claude, Microsoft Copilot, or ChatGPT through a standard MCP Server and get answers in one conversation, without waiting on the back room.
- The full audience in every answer: because Celebrus AI queries that same data, answers include the anonymous and logged-out customers most AI tools never see.
- From question to action in one conversation: teams move from what happened, to why it happened, to which customers to act on, without switching tools or waiting on a report.
- Answers you can trust: every response shows which tool produced it, and data and queries stay inside your own private cloud, with GDPR, HIPAA, and CCPA compliance built in.
That difference is measurable. In one case, only 12% of visitors were logged in, yet more than 60% of visits could be recognized, climbing to over 80% for returning visitors. Read the case study here.
Watch the Full Conversation
Bill and Keanu go deeper on identity, data ownership, and why AI is forcing enterprises to confront what’s been happening behind the scenes all along. Watch the full conversation below.
