Agentic AI in 2026: Companies Accelerate, But Trust and Governance Can't Keep Up

If 2025 was the year of "let's see if this works," then 2026 is the year of "we implement or we'll get left behind." That's how he describes it Informatica, summarizing a study of 600 data leaders.

The most important things in 60 seconds

• GenAI is already in 69% of organizations, and another 25% plan to implement it within 12 months.
• Agentic AI, as systems that not only generate but also "work", 47% of companies already have them, and another 31% plan to do so within a year.
• The scale is growing, but trust can be blind: 65% of organizations declare a high level of employee trust in data used by AI, while at the same time 75% of leaders report gaps in data literacy and 74% in AI literacy.
• The most common agent production blockers: data quality and retrieval (50%), security (43%), lack of competences (42%), agent management tools (39%), observability (39%), guardrails (35%).

What is agentic AI

In short, agentic AI is an approach in which the AI ​​system doesn't stop at generating a response, but instead takes action: invoking tools, executing steps in a process, escalating to a human, and saving in the system. This is the natural "next step" after GenAI—and why companies are accelerating so rapidly.

"Trust paradox": people trust, even though they cannot verify it

One of the most practical observations from the study is the so-called trust paradox: employees trust data and AI results, but often lack the expertise to assess when the AI ​​is right and when it sounds convincing and… wrong. The report found: 65% of employees declare trust vs. need for further training (75% data literacy, 74% AI literacy).

This is a real risk in a production environment – ​​especially where decisions affect: pricing, offers, credit risk, complaints, PII, compliance.

Practical lesson: if you build agents, you simultaneously build verification habits (process) and competencies (people). Otherwise, the agent will "ride on authority."

The 6 Barriers That Most Often Block Agents in Production—And How to Unblock Them

Below is a list of barriers straight from the report (with percentages) and "what to do tomorrow morning" to avoid getting stuck at POC.

1) Data quality and retrieval (50%)

If an agent is to work, it must retrieve the right data, on time, with clear context.

What works in practice:

• data directory + ownership (who is responsible for the source),
• quality rules (what is "correct" for the customer/product/order),
• semantic layer / dictionary of terms (so that "income" means one thing),
• monitoring freshness and completeness for critical data.

2) Security (43%)

The agent is a new type of user: fast, automated, and… sometimes too creative.

Minimum:

• principle of least privilege,
• PII masking/segmentation,
• audit (who/what/why downloaded it),
• tool control: the agent should not have access to "everything".

3) Lack of agentic AI competences (42%)

It's not just about prompting. It's about designing:

• workflow,
• moments of escalation to a human being,
• tests and acceptance criteria.

Quick fix: playbook "how the agent works for us" + training on specific use cases.

4) Poor agent management tools (39%)

Without tools, you create "shadow AI" and then extinguish the fire.

It's worth having:

• agent register (who owns, what purpose, what data sources),
• versioning and change process,
• evaluations after changing data/model/tools.

5) Observability (39%)

If you don't know why the agent did X, then you have a problem in production.

Minimum observability:

• step logs (trace),
• response/action quality metrics,
• alerts on deviations (e.g. sudden increase in "uncertainty", retrieval errors).

6) Guardrails (35%)

Guardrails are not "brakes." They are seat belts.

The simplest and most effective:

• allowlist of tools and actions,
• limits (rate limits, token/share budget),
• escalation thresholds (human-in-the-loop) for irreversible actions.

Pilot → Production: Data is still the bottleneck

In the report, 57% of leaders say it plainly: data reliability still blocks the transition from pilot to production.
And this is interesting: companies know what works because they indicate the specific steps they take most often:

• improved workflow around data/AI (59%),
• investments in data quality (54%) and metadata (54%),
• more frequent data checks (53%),
• upskilling/hiring (47%),
• external support (43%).

This sounds like a "big program," but it can be tackled in stages.

Minimum viable AI governance (no major annual program)

The report also shows that 76% of organizations admit: their visibility and governance do not keep up with how employees use AI.
If you want to get started sensibly and quickly, think of AI governance as three simple questions:

1. Who owns the agent and the data it uses?
2. What powers does it have and how do we audit it?
3. How do we measure quality and risk (and what do we do when metrics decline)?

This is enough to stop acting "on faith".

"Agentic AI Production Ready" Checklist (for use)

1. Defined agent purpose and limits of responsibility
2. Business Owner + Data Owner
3. Data source list + classification (PII/sensitive)
4. Quality rules for critical data
5. Quality monitoring + freshness + alerts
6. Semantic layer / glossary of terms
7. Access control (least privilege)
8. Masking/anonymization where needed
9. Logging agent activity (trace)
10. Metrics: effectiveness, retrieval errors, escalations, time, cost
11. Guardrails (allowlist tools, limits, escalation thresholds)
12. Regression testing (after changing data, model, prompts, tools)
13. Versioning and deployment process (who approves changes)
14. User Training: Data Literacy + AI Literacy
15. What we do when an agent makes a mistake plan (runbook)

FAQ 

Is agentic AI the same as a chatbot?
No. The chatbot primarily responds. In addition to responding, Agentic AI performs actions in the process (e.g., creates a ticket, updates a record, initiates a workflow).

Why does agentic AI so often rely on data?
Because the agent must retrieve and act on data, the study found that the largest production barrier was data quality/retrieval (50%).

How to start AI governance so as not to get bogged down in documents?
Starting with the bare minimum: ownership, authorizations, auditing, and quality/risk metrics. Only then should framework development begin.

Where do these numbers come from (transparent)

These are the conclusions from the report "CDO Insights 2026" prepared by Informatica Based on Wakefield Research survey: 600 data leaders ($500M+ revenue companies), USA/UK+EU/APAC, September 12–25, 2025.

READ MORE OUR BLOG