Key Takeaways
- Data warehouses, data lakes, BI dashboards, automation, and AI each solve different problems. Choosing the wrong tool is the most common and most expensive mistake.
- Rule-based automation works best for repeatable, structured tasks. AI works best where judgment and unstructured text are involved, always with a human reviewing.
- Most “AI failures” are really data and process failures, so a strong foundation comes first.
- A focused 13-week pilot can prove value without a large budget.
Imagine a public records request that starts working the moment it arrives. An AI agent reads the request, follows your plain-language instructions, pulls and redacts the right documents with a full audit trail, and hands the result to your team for review. That’s not science fiction. Organizations are doing versions of this today.
For government finance teams, the hard part is cutting through the vocabulary to figure out what’s useful.
The Vocabulary, in Plain English
- Data warehouse: A well-run records room. Data is cleaned and organized before it’s stored, so two people pulling the same report get the same answer. Great for trusted, repeatable reporting.
- Data lake: Stores everything in its raw form, including PDFs, emails, GIS files, and meter readings, at low cost. Without governance, it can become a “data swamp.”
- Lakehouse: The best of both, combining low-cost storage with structure for reporting. Many organizations are moving this direction.
- Business intelligence (BI): Tools like Power BI, Tableau, and Domo turn stored data into dashboards for council packets, budget reviews, and monthly close. BI won’t fix bad source data, but it makes good data visible.
- Robotic process automation (RPA): A very fast, very literal temp that never gets bored and never improvises. Ideal for high-volume, rule-based work.
- Large language models (LLMs): Excellent at summarizing, drafting, and classifying text. They aren’t built for arithmetic and can state wrong answers confidently.
- Agentic AI: Instead of answering one question, an agent works toward a goal. It plans steps, uses tools, checks its own work, and hands results to a person to approve.
Why tools matter. Tools are where agents make their biggest impact. A chatbot can explain how a bank reconciliation works, but an agent with the right tools can actually perform one: pulling the bank and general ledger files, matching transactions, and investigating exceptions. It can also enter transactions into your ERP or assemble the documents for a records request. That’s work people currently do by hand. With clear permissions and a human signing off at the end, agents can take real work off your team’s plate.
Which Tool for Which Job?
A simple rule of thumb: if you can write the rule down completely, then you should use traditional, cost-effective automation techniques to automate it. If you can’t, you need judgment from a model with a human checking, or from a person.
Rule-based automation fits work such as:
- bank reconciliations
- invoice and PO matching
- report distribution
- syncing systems that don’t talk to each other
- recurring journal entries
- compliance packets
AI fits work such as:
- summarizing a 400-page ACFR for your board
- triaging records requests
- reviewing contracts and grants
- drafting budget narrative
- explaining variances
Before automating anything, review the process. Many steps exist only because a person has always done them that way.
Is the Hype Real?
Yes, and also no. The capability is real, and it continues to improve. But adoption trails capability. According to McKinsey’s 2025 State of AI survey, most organizations report using AI somewhere, yet fewer than 10% have scaled AI agents within any single business function.
The hardest part was never the model. It’s your data, your definitions, and your processes. These systems also fail fluently, so unreviewed output is a liability.
Build the Foundation First
Getting to AI is a sequence of building blocks:
- Core systems implemented correctly
- Data governance
- A managed data layer
- A reporting layer that defines what terms like “fund” and “encumbered” mean
- Only then, agentic AI with clear limits
An agent inherits your data quality.
Start With the Problem, Not the Tool
A practical 13-week pilot keeps things manageable:
- Weeks 1–3, find friction: Inventory time-consuming admin tasks and the numbers teams debate.
- Weeks 4–8, pilot one: Map the process, clean the data, match the need to AI, automation, or a dashboard, and build something simple.
- Weeks 9–13, prove and scale: Run old and new processes side by side, measure hours saved, and report honestly on what worked.
The goal isn’t headcount. It’s making sure experienced finance professionals spend their time on judgment, not keying invoices.
Curious where to start? Rea’s data and analytics advisors help Ohio governments find the right first project and build a foundation that lasts. Connect with an advisor.
About the Authors
Chris Holmes, CPA, CMA, MBA, is a Principal, Data Analytics & Engineering at Rea. He focuses on data engineering, analytics platforms, and practical automation that changes how organizations operate. You can reach Chris at Chris.Holmes@reaadvisory.com.
Andrew James is a Principal, Data Analytics at Rea. He helps teams turn data into strategic insight through reporting and visualization, and get more value from the systems they already own. You can reach Andrew James at Andrew.James@reaadvisory.com.
