Most enterprise AI pilots look impressive in a demo and then quietly stall. The agent answers sample questions well, leadership is excited, and six months later almost nobody uses it. The gap between a working pilot and a system people rely on every day is rarely the model. It is everything around the model.
Over the past two years I have designed and run AI agents that answer questions over financial data such as contracts, orders, invoices and employee stock plans. Finance is an unforgiving place to deploy AI, because a fluent but wrong number is worse than no answer at all. These are the five lessons that mattered most in moving from pilot to production.
Lesson 1: Start with a workflow people already hate
The most successful agents I have seen did not start with a clever demo. They started with a task people dreaded: chasing a contract detail across three systems, waiting days for a report request, or answering the same policy question for the hundredth time.
Starting there gives a project two things a demo cannot. It gives a clear measure of success, because everyone already knows how painful the old process is. It also gives a group of users who want the tool to work and will tell you quickly when it does not.
A useful test is to ask the people doing the work what they would stop doing tomorrow if they could. If the answer is not specific, the use case is probably not ready.
Lesson 2: Treat hallucination as an engineering problem
Early in any finance AI project, the model will produce a confident answer that is simply wrong. The instinct is to fix it with better prompts. In my experience, prompting helps at the margins but does not make a system trustworthy.
What works is engineering the context the model receives. Language models are built for human language, while enterprise data lives in tables of identifiers, codes and amounts that carry no business meaning on their own. Giving the model data that has been described in the terms the business actually uses removes a large class of errors before they happen.
For the questions that matter most and are asked most often, it also pays to have domain experts review and approve the underlying logic once, so the system reuses trusted logic instead of improvising each time. Reviewers approve the method, not every individual answer, which keeps the approach practical at scale.
Lesson 3: Build security in from day one
A conversational agent that can see everything will eventually tell someone something they should not know. This risk grows as the user base grows, because one shared system is now answering questions for people with very different entitlements.
The fix is to bind every query the agent makes to the identity and permissions of the person asking, and to enforce that at the moment the question is answered. Single sign-on, role-based access and row-level controls are not features to add after launch. They are the precondition for launching at all, and security teams become allies rather than blockers when they see this designed in from the start.
Lesson 4: Keep humans where judgment matters
Full automation is tempting, but the best production systems are deliberate about where people stay in the loop. Routine lookups can be answered directly. Actions with financial or personal consequences, such as sending a customer correction or approving a request, benefit from a person confirming before anything is committed.
This is not a lack of ambition. It is how trust is earned. Once users see that the agent handles routine work reliably and escalates the rest cleanly, adoption accelerates on its own.
Frameworks such as the NIST AI Risk Management Framework offer a useful vocabulary for deciding where human oversight belongs and for documenting those choices.
Lesson 5: Measure hours saved, not messages sent
Many AI programs report usage in terms of questions asked or conversations started. Those numbers are easy to inflate and say little about value. A tool that people try once and abandon can still post impressive message counts.
Better measures are whether users return, whether answers survive review by the people accountable for them, and how many hours of manual work are genuinely returned to the business. Repeat usage is especially telling, because people only come back to a tool they trust. In my experience, the systems that reached broad adoption were the ones that proved themselves on these measures first.
What this means for leaders
Moving AI agents into production is less about choosing the most capable model and more about disciplined engineering around it. That means picking a painful workflow, engineering context instead of relying on prompts, enforcing access at query time, keeping people in the loop where judgment matters, and measuring real outcomes.
Enterprises that follow this path will find that adoption follows trust, not novelty. The next wave of enterprise AI will be won by the teams that can explain why their agent’s answer is right.
About the author
Arjun Danda Sureshbabu is a Staff Software Engineer and AI architect at a leading enterprise data company, where he designs AI systems used in revenue, finance and employee operations. He previously architected cloud orchestration spanning 100,000+ hosts at Workday and was lead engineer on GE Digital’s Predix Private Cloud. He is a peer reviewer for IEEE Access, a full member of Sigma Xi, The Scientific Research Honor Society, and spoke on reducing hallucinations in enterprise AI at Conf42 Observability 2026 (watch the talk). The views expressed are his own.
