The data readiness gap: why AI agents stall before launch

The demo went great. The agent answered pipeline questions in plain English, the founders were impressed, and someone said "ship it" in the meeting. Three weeks later the same agent told a customer their invoice was paid when it wasn't, quoted a revenue number the CFO didn't recognize, and got quietly unplugged. The model didn't get dumber between the demo and production. It met your data.

That story is 2026 in miniature. Gartner predicted in August 2025 that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5% in 2025. At the same time, data is what leaders flag first: in Informatica's CDO Insights 2026 survey of 600 data leaders (January 2026), about half named data quality as the top challenge in deploying agentic AI. Informatica sells data management software, so it has a stake in that answer. The bottleneck moved, and most budgets haven't noticed.

What does "AI-ready data" mean?

It doesn't mean perfect data. It means known data. Four properties, in practice:

• One definition per metric. "Active customer," "MRR," and "churn" each mean exactly one thing, written down, with an owner. If your sales dashboard and your finance sheet disagree on revenue, an agent picks one at random and answers confidently.
• Documented lineage. You can trace any number the agent might quote back to the table and transformation that produced it. When the agent is wrong, you can find out why in minutes, not in a week of archaeology.
• Quality checks where it counts. Freshness, nulls, duplicates, and volume anomalies monitored on the tables the agent reads. Not on all 400 tables. On the 20 that feed answers.
• Access rules the agent inherits. The agent can see what the person asking is allowed to see, and nothing else. "We'll add permissions later" is how a support bot ends up reciting salaries.

Why do agents fail when dashboards were fine?

Because dashboards forgive and agents don't. A human reading a slightly stale chart squints, cross-checks, and interpolates. Twelve years of BI culture is built on that quiet human error-correction. An agent has no squint. It takes the data at face value and then acts: sends the email, files the ticket, gives the discount, answers the customer.

That's the uncomfortable part of the data readiness gap. Your data was never as good as the dashboards made it look. The agent didn't create the quality problem. It removed the human buffer that was absorbing it.

How much governance does a mid-sized company actually need?

Less than the enterprise playbooks say, and more than you have. The full governance stack (councils, stewardship programs, quarterly committees) is built for 5,000-person companies and dies of starvation at 100. What a 20-to-500-person company needs is minimum viable governance:

1. A metrics layer with named owners. Every metric an agent can quote has one definition and one accountable human. This is the semantic layer's real job in 2026: it stopped being a BI accessory and became the contract between your data and your AI.
2. Contracts on the critical tables. The 15 to 25 tables that feed agents and executive reporting get schema checks, freshness SLAs, and an alert channel someone actually reads.
3. An access model, not access exceptions. Roles mapped once, inherited by every tool, agents included.
4. An evaluation set. Fifty real questions with verified answers, re-run on every change. If the agent's accuracy drops, you know before your customers do.

We'd scope that as a 30-to-60-day project for a mid-market team, not a transformation program. It's also the highest-ROI AI spend available to you, because it lowers the cost of every subsequent AI project, which is the argument we ran through in what AI implementation actually costs.

Fix the data first, then buy the agent

Our read: the order matters more than the vendor. Launch an agent on unready data and you'll end up in incident response, with the organization's trust in AI taking the hit. Spend six weeks on readiness first and the agent has a real chance of holding up in front of customers. Same model, same budget, different order. Gartner predicted in June 2025 that more than four in ten agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. We'd bet a good share of that wave comes from teams that launched first and fixed the data later.

This is the exact seam where our Data practice and AI practice work as one team: readiness, metrics layer, and evals on one side, agent design on the other, one contract instead of two vendors pointing at each other. And if you want an honest read on where your data operation stands next to the other seven dimensions, that's what the DGTL Readiness Index maps.

Related: The modern data stack for mid-market → · What AI implementation actually costs → · Data infrastructure for B2B →