For years, ETL teams have improved reliability by making workflows more predictable: define the sequence, codify the rule, monitor the outcome and intervene when reality departs from the plan.

Agentic AI changes that relationship. Instead of simply executing a prescribed step, a system may now interpret conditions, weigh imperfect evidence, propose a response and, within defined limits, act. That creates a more consequential migration than the move from one data tool or platform to another.

The opportunity is significant, but so is the gap between an impressive demonstration and a dependable production capability. Technology research firm Gartner predicts that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value and inadequate risk controls. The forecast is less an argument against agents than a warning against treating autonomy as a feature that can be added without redesigning the operation around it.

For data teams running mature ETL estates, the pipelines themselves may remain in place while the decision layer around them changes. The central question is therefore not whether an agent can act, but how it earns authority without weakening accountability.

The Ardent perspective: migration starts where the runbook stops

Ardent approaches agentic operations as an engineering and operating-model challenge, not a race to insert autonomous tools into existing workflows. Trusted data, governed access and reliable observability remain essential, but they are only the foundation. A production migration must also define who owns a decision, what evidence can support it, when a human must intervene and how cost, risk and authority are constrained.

That view comes from Ardent’s experience designing, modernising and operating critical data platforms across AWS, Databricks, Snowflake, Redshift and complex enterprise estates. In those environments, the most important knowledge rarely sits neatly in the architecture or formal runbook. It is held by engineers who know when a late source is routine, when a harmless retry masks a wider incident and when a green job status still leaves downstream data incomplete.

Traditional automation can reproduce a documented step; an agent is asked to interpret the conditions around it. Ardent therefore starts migration by drawing out the evidence engineers trust, the exceptions they recognise and the moments when business context overrides system data. This exposes unclear ownership, conflicting signals, actions that are difficult to reverse and technically correct responses that could still create an unacceptable business outcome.

Ardent turns that operational reality into contracts, evidence gates, review rhythms and bounded authority, creating an operating model in which autonomy is earned rather than assumed. That is how anagent moves from appearing capable in a demonstration to becoming dependable in a live data operation.

A practical roadmap for a live ETL environment

Ardent’s roadmap begins not with greater autonomy, but with an operation that is legible enough for autonomy to be earned. Five connected stages move a live pipeline from observation to controlled action.

1. Define the operational contract before selecting the agent

Start with one pipeline and define its operational contract:

• Its accountable owner

• The protected business outcome

• The approved evidence

• The response when a recommendation looks wrong

Measures such as the following show whether the agent is improving the operation:

• Diagnosis time

• Missed delivery windows

• Repeat failures

• Cloud cost

Keep the outcome narrow: for example, reducing time spent diagnosing a recurring failure class. That creates a meaningful baseline and a clear reason to withhold further autonomy if value does not improve.

2. Run the agent in shadow mode and capture disagreement

Run the agent first as a shadow operator. It should:

• Observe the same evidence as the team

• Record its recommendation

• Compare it with what the engineer actually did

Disagreement reveals missing dependencies, exceptions and business priorities.

Include healthy runs as well as incidents so the agent learns acceptable variation and when doing nothing is safest. The aim is not just accuracy, but understanding why two plausible decisions diverge and using that evidence to improve rules and runbooks.

3. Redesign the engineer’s role around review, not duplication

Redesign the engineer’s role so review adds judgement instead of duplicating the agent’s work.Otherwise, teams either repeat every manual check and gain littleefficiency, orallow oversight to fade.

In Ardent’s model:

• The agent gathers and compares evidence, taking on the repetitive evidence gathering

• The engineer tests the recommendation against business context and remains accountable for ambiguity, exceptions and impact

4. Expand authority through evidence gates

Expand autonomy only when a specific action is understood, bounded and reversible. Ardent’s authority ladder moves through four stages:

• Observation

• Recommendation

• Execution with approval

• Automatic execution for narrow, low-risk cases

Each step needs an explicit gate based on incident type, evidence, cost and performance. Authority belongs to a bounded action, not to the agent as a whole, and should return to recommendation mode when conditions change.

5. Scale the operating model before scaling the number of agents

Scale the operating model before multiplying agents. The reusable asset is not the first agent itself, but:

• The ownership charter

• The evidence checklist

• The authority ladder

• Review metrics

• Cost limits

• The process for updating runbooks

Choose the next pipelines by operational fit:

• Repeated manual diagnosis

• Reliable system data

• Understood actions

• Measurable impact

A common vocabulary for ownership, evidence and escalation then supports consistent decisions without forcing every pipeline into the same template.

Cost has to be designed into the workflow

Agentic AI changes the economics of data operations because reasoning recurs. An agent may inspect logs, retrieve documentation, compare incidents, call tools and reassess the result before recommending one action. Costs escalate when agents are used where predictable automation would suffice, when every eventtriggersmodel reasoning or when nobody has defined the expected improvement.

Ardent therefore separates routine execution from genuine judgement. Existing orchestration should continue to handle predictable steps, rules should manage known thresholds and agents should be introduced only where context, comparison and decision-making create additional value. The goal is not an entirely agentic ETL estate, but targeted agentic behaviour where fixed logic is no longer enough.

That division of labour makes cost an architectural decision rather than a post-launch surprise. Teams can reserve model calls for incidents where several imperfect signals, and the kind of contextual judgement once held only in engineers’ experience, must be reconciled. Simpler cases can stay with rules, while limits on investigation time prevent an agent from consuming resources without improving the outcome.

That discipline does not end once a pipeline goes live. As agents take on more incident types, spend can drift upward even when each individual decision looks reasonable, so cost needs the same review schedule as accuracy and authority. Ardent tracks cost per resolved incident alongside diagnosis time and repeat-failure rates, so a pipeline earning more autonomy is also earning it at a sustainable price.

Trust is built through visible review

Enterprise adoption reinforces the same point: global consultancy McKinsey found that 62 per cent of surveyed organisations are at least experimenting with AI agents, yet nearly two-thirds have not begun scaling AI across the enterprise. The strongest performers are more likely to redesign workflows and establish clear leadership ownership: conditions that matter as much as model capability.

Trust grows when teams can see the evidence used, understand why an action was proposed, challenge the reasoning and trace the outcome. Ardent builds review into deployment from the outset so that the informal knowledge surfaced during migration does not disappear into a model. Accuracy matters, but so do effort removed, business impact, proportional cost, how clearly the reasoning can be explained and whether engineers can fulfil their revised role.

Visible review also creates the evidence leaders need to decide whether the operating model is maturing. Acceptance rates alone are not enough: a high rate may reflect genuine quality, passive approval or recommendations that are too cautious to matter. Ardent looks for a broader pattern: less avoidable effort, faster diagnosis, stable outcomes, proportionate cost and clear records of why authority was granted or withdrawn.

Sustaining that trust means the audience for review has to widen as adoption grows. Early on, review sits mostly with the engineers closest to the pipeline; as authority expands, business owners and leadership need the same visibility into evidence and outcomes, not just a summary after the fact. That is often whatseparates organisations that scale agentic AI responsibly from those that stall after the pilot: review becomes a shared, ongoing habit rather than a gate one team owns alone.

One decision boundary at a time

Theorganisationsthat make this transition successfully will not give agents the widest authority first. They will make accountability explicit, test performance against real operating conditions and redesign human review so that it adds context rather than duplicating machine work. Most importantly, they will treat autonomy as something that can be narrowed or withdrawn when the evidence changes, not as a one-way destination.

For data leaders, the first move is contained and measurable: choose one pipeline where manual interpretation creates genuine friction, define its operational contract, run the agent alongside the team and use the evidence to decide what authority should change next. That is how an impressive pilot becomes a dependable part of a live data operation.

Ardent brings that discipline together with the platforms and pipelines already supporting the business. By combining data engineering, governance, SRE practice and operational ownership, we help clients move beyond the question of whether an agent can act and answer the harder question: under what conditions should it be allowed to act in production?

Speak to Ardenttodayabout building the operating model behind your next agentic AI migration, not just the platform beneath it.

Sources and further reading

• Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027” (25 June 2025).

• McKinsey & Company, “The State of AI in 2025: Agents, Innovation, and Transformation” (5 November 2025).

• Ardent, “Building a Migration Path for Agentic AI-Ready Data Pipelines” (2026).

• Ardent, “Intelligent ETL: From Data Preparation to Decision Intelligence” (2026).

• Ardent, “Runbooks, Incident Management and Alert Triage: The Practical Side of AI in SRE” (2026).

• Ardent, “Robust, Scalable Data Pipelines with AWS Infrastructure” (26 August 2022).