Circles powers telco personalization with OpenAI technology

Mid-market companies adopting AI for customer-facing operations often focus on the model itself—its accuracy, its capabilities, its vendor’s reputation. But the model is rarely what determines success or failure. What determines outcomes is the decision framework you use before you commit: how you scope the problem, how you define “working,” and how you structure the path from proof-of-concept to production.

This piece is for operations leaders, IT executives, and RevOps teams at companies with 100–2,000 employees who are evaluating AI investments for customer personalization, service automation, or revenue optimization. If you have been through one AI pilot that stalled, or you are about to green-light your first, this framework will save you months.

The Three Questions That Predict Outcomes

Before evaluating vendors or scoping a pilot, disciplined organizations answer three questions. Getting these wrong costs more than the project budget—it costs the political capital needed to try again.

What Problem Are We Actually Solving?

This sounds obvious, but most AI initiatives start with a solution (“We need AI-powered personalization”) rather than a problem (“Our churn rate is 9% higher than competitors because we cannot respond to usage pattern changes within the billing cycle”). A recent case study from a telco operator illustrates this: their measurable outcomes—22% increase in average revenue per user, 9% reduction in churn—came not from deploying AI broadly but from targeting AI at specific, well-defined friction points in the customer lifecycle.

The test: can you state the problem in terms a finance team would recognize? If the answer is “better customer experience” or “more personalization,” you are not ready. If the answer is “reduce time-to-resolution on billing disputes from 4.2 days to under 24 hours,” you have something you can actually measure.

Who Owns the Outcome?

AI projects cross organizational boundaries by nature. A personalization engine touches marketing, product, customer success, and IT. Without a single accountable owner—someone whose compensation or career advancement depends on the project’s success—decisions stall at every handoff.

The warning signs:

• Steering committees with no decision-making authority
• IT owns the implementation but marketing owns the use case
• Success metrics defined differently by each stakeholder
• No one can fire the vendor or kill the project without consensus

In most engagements, projects with a single accountable executive reach production 2–3x faster than those governed by committee.

What Does “Production” Actually Mean?

A proof-of-concept that works on sample data is not a production system. Production means: integrated with live customer data, handling real transaction volumes, monitored for drift and failure, and operated by people who did not build it. The gap between POC and production typically costs 3–5x what the POC cost, and takes 3–4x as long.

If your vendor or internal team cannot articulate what production looks like—including SLAs, monitoring, rollback procedures, and ownership handoff—you do not have a production plan. You have a demo.

The Integration Reality

Mid-market companies rarely have clean data infrastructure. Your CRM has duplicates. Your ERP has legacy fields no one understands. Your customer data lives in three systems that do not agree on what constitutes an “account.” AI does not fix this—it amplifies it.

The most common integration failures follow a pattern:

• Data quality issues surface only after the model is trained, requiring expensive retraining
• Real-time data feeds that worked in testing fail under production load
• The model expects data in a format that requires manual transformation for every batch
• Security and compliance reviews take 8–12 weeks longer than planned because the AI vendor’s data handling was not vetted upfront

Successful projects budget 40–60% of total effort for integration and data preparation. Projects that treat integration as a phase to “figure out later” rarely reach production.

Measuring What Matters

AI vendors love to report accuracy metrics. “Our model achieves 94% accuracy on intent classification.” That number is meaningless until you know: accuracy compared to what? Your current process? A human doing the same task? A random baseline?

The metrics that matter for business outcomes are different:

Before signing a contract, define the business metric that justifies the investment. Then work backward to determine what technical performance is required to move that metric. If the vendor cannot explain the connection, they are selling technology, not outcomes.

The Baseline Problem

You cannot measure improvement without a baseline, and most organizations do not have one. What is your current churn rate by customer segment? What is the average handling time for the specific support queries you want to automate? How many hours does your team spend on the manual process the AI will replace?

Establishing a baseline typically takes 4–8 weeks and requires instrumentation you probably do not have. That work should happen before the AI pilot starts, not alongside it. Projects that skip this step often succeed technically but cannot prove ROI, which means they get defunded at renewal.

The Adoption Failure Mode

Assume the model works. Assume the integration succeeds. You still have to get people to use it.

Frontline employees who have been doing their jobs for years do not automatically trust AI recommendations. Sales reps who built relationships by intuition resist “next-best-action” prompts. Customer service agents who take pride in problem-solving ignore suggested responses. This is not irrational—it is human.

The successful 20% of projects address adoption from day one:

• Involve end users in defining use cases and success criteria
• Start with augmentation (AI assists humans) before automation (AI replaces steps)
• Make it easy to override or ignore the AI, and track when people do—those overrides contain signal about where the model is wrong
• Celebrate the first win publicly, with specific numbers, attributed to specific people

Change management is not a phase after go-live. It is the work that determines whether go-live means anything.

When to Walk Away

Not every AI project should proceed. The disciplined organization knows when to say no:

• When the problem is not painful enough to justify the organizational disruption
• When the data does not exist and creating it would take longer than the expected payback period
• When no single executive is willing to own the outcome
• When the vendor cannot show reference customers in your industry at your scale
• When “success” cannot be defined in terms finance will accept

Killing a project before it starts costs nothing. Killing it 18 months in, after you have spent budget and political capital, costs everything.

The companies that succeed with AI in customer operations are not the ones with the biggest budgets or the most sophisticated models. They are the ones that answer the hard questions before they start building. They define the problem in business terms. They assign a single owner. They budget for integration. They establish baselines. They plan for adoption. And they have the discipline to walk away when the answers do not support the investment.

The model will work. The question is whether your organization is ready to absorb it—and that question has nothing to do with AI.