How Do You Test an AI Agent Before Putting It in Production in a Bank?

7 min readAbleneo AI transformation team

Short answer. Test an AI agent in a bank against a versioned evaluation harness that scores the full task, every tool call, and the end system state, not just the final answer. A workable path takes about 90 days: 30 days to build the harness and a baseline, 30 days of shadow mode against live traffic with no customer impact, and 30 days of staged rollout behind written approval gates. Gartner projects more than 40% of agentic AI projects will be cancelled by the end of 2027, and weak evaluation is a leading cause.

1What This Means in Practice

An AI agent is not a chatbot. It reads a request, decides on steps, calls tools, and changes system state: it opens a case, updates a record, moves a payment flag, or escalates to a human. Testing it like a single question-and-answer model misses the part that carries risk. A polished final message can hide a wrong tool call, an invented parameter, or a skipped approval step. In a bank, that gap is the difference between a demo and a production incident.

Testing an agent before production means building a harness that replays a fixed set of real tasks and scores each run on several dimensions at once. The dimensions that matter are task completion, tool-call accuracy, the resulting system state, latency, and cost per task. The harness runs in continuous integration and blocks any change that regresses a score below a defined threshold.

Ableneo runs roughly 4 of 5 AI projects into production, and that rate holds because the test gate comes before the go-live decision, not after. In 2025, 34 production AI projects shipped across 20 clients, and 94% of them used large language models, so the harness pattern is the same one applied repeatedly rather than a one-off.

2Why This Matters for Regulated Industries

A credit-scoring or fraud-decision agent in an EU bank is a high-risk AI system under Annex III of the EU AI Act (Regulation (EU) 2024/1689). Article 9 requires that high-risk systems be tested to identify the most appropriate risk-management measures, and that testing be carried out against prior defined metrics and probabilistic thresholds before the system is placed on the market or put into service. Article 15 requires a declared, measurable level of accuracy, robustness, and cybersecurity that holds across the system’s lifecycle. Testing is not optional polish. It is a legal precondition for deployment, with the Annex III obligations due by 2 December 2027.

DORA adds operational resilience testing and third-party risk obligations for financial entities, in force since January 2025. The ECB has also asked significant institutions to deliver AI-cyber action plans by 31 October 2026. For a bank, the practical effect is that an agent’s test evidence, its harness results, its approval gates, and its audit trail, become documentation a supervisor can ask to see. Building that evidence during testing is cheaper than reconstructing it during an audit.

Score the full task, every tool call, and the end system state, not just the final answer. A good-looking reply can hide a wrong action.

3What Should an AI Agent Eval Harness Actually Measure?

A harness for a banking agent measures five things on every run. Task completion asks whether the requested outcome was reached. Tool-call accuracy checks that each tool was called with correct parameters and in a valid order. System-state verification confirms the end state is what the task required, because a correct-looking answer can still leave a record wrong. Latency and cost per task close the set, since an agent that is accurate but slow or expensive fails a different gate. Scoring all five at once stops a single good-looking metric from masking a real fault.

The golden tasks come from real production traces, not invented examples, so the harness tests the distribution the agent will actually see. A bank should hold out a separate test split that the team never tunes against, to catch overfitting to the eval set.

4How Do You Test Tool Calls and Actions, Not Just the Final Answer?

The highest-risk failures in a banking agent are in the actions, not the prose. The harness records and replays every tool call, then scores whether the call was correct, whether its parameters were valid, and whether any required approval was requested before a sensitive action. A payment-hold agent that writes a fluent summary but releases a hold without approval has failed, even if the summary reads well.

Checking the resulting system state is the companion test. After each run, the harness verifies the record, the flag, or the case status against the expected end state. This catches the class of fault where the agent reports success but leaves the system in the wrong place, which is the fault type that turns into a customer complaint or a reconciliation break.

5What Does Shadow Mode Testing Look Like in a Bank?

Shadow mode is the safest way to test an agent against real conditions without real exposure. The agent runs on live traffic and logs the action it would take, while the existing process or a human still makes the binding decision. The team then compares the agent’s proposed actions against the real outcomes for 30 days. Agreement rate, the pattern of disagreements, and the cost profile all become evidence for the go-live decision.

Shadow mode also produces the baseline the staged rollout needs. Once agreement is high and the disagreements are understood, the agent moves to a small share of live decisions behind an approval gate, then widens as the evidence holds. This staged path is how a stalled pilot reaches production without betting the whole process on a single switch.

6How Do You Keep the Test Suite Honest Over Time?

A test suite decays if it is not maintained. Three disciplines keep it honest. First, feed new production traces back into the golden task set so the harness tracks real drift in inputs. Second, when a model or prompt changes, re-run the full suite in continuous integration and block any regression, because a model swap can quietly break a tool-call pattern that passed last week. Third, if the harness uses an LLM to judge outputs, calibrate that judge against human review on a regular sample, since an uncalibrated judge drifts. Reproducibility matters too: run at temperature 0 or repeat each task and take the median, so a score means the same thing twice.

7The Ableneo Perspective

Ableneo treats the evaluation harness as part of the build, not a step bolted on at the end, which is why roughly 4 of 5 projects reach production across a portfolio of 34 production AI projects in 2025. The same discipline shows in named work: at T-Mobile Czech Republic, a document-validation system cut processing from 7 hours to 30 minutes, 14 times faster, at over 90% accuracy, after earlier attempts had not reached production. The lesson from FS&I clients such as ČSOB, Erste, and UNIQA is consistent. An agent that cannot be measured cannot be signed off, and an agent that cannot be signed off never ships. For the runtime side of this, see our guide on how to keep an AI agent reliable in production in a bank.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo