Painted scene of an analyst reviewing an evidence binder and invoices at dusk while an unused brass approval stamp waits beside a blank signature line.

AI Reconciliation Agents: Human Approval and Audit Controls

November 17, 2026

Painted scene of an analyst reviewing an evidence binder and invoices at dusk while an unused brass approval stamp waits beside a blank signature line.

AI Reconciliation Agents: Build an Evidence-First Approval Workflow

By Erin Harris, AI Strategy & Business Transformation Consultant

An AI reconciliation agent should investigate mismatched records, assemble evidence and recommend next steps before it is allowed to change financial data. The right starting point is not model selection. It is a controlled workflow that separates reading, reasoning, recommending and approving.

AWS published a November 16, 2026 customer case describing Duco's implementation of Amazon Bedrock and AgentCore for financial operations. Duco uses agents for reconciliation onboarding and exception investigation, with analysts reviewing decision-ready cases. The post reports reconciliation deployments ten times faster and a 70 percent reduction in onboarding costs for one global bank. These are attributed case-study outcomes, not independently validated savings estimates for other organizations. Source: https://aws.amazon.com/blogs/machine-learning/how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore/

Why the handoff matters

Financial discrepancies are rarely just arithmetic. A missing payment can reflect timing, inconsistent identifiers, stale reference data or a genuine exception. A system that immediately posts an adjustment may make a clean-looking dashboard while obscuring the underlying cause. A better agent gathers the original records, explains its reasoning and identifies what remains unresolved.

Duco's design also offers an important infrastructure lesson. Its team reports reducing main agent context from about 200,000 to 20,000 tokens by replacing a large static tool catalog with on-demand skills discovery. That reduction does not mean every business can cut costs by 90 percent. It means an implementation can avoid loading irrelevant tool instructions into each task. The case also describes tenant-scoped identity and tool authorization, not simply a prompt asking the model to respect boundaries. Source: https://aws.amazon.com/blogs/machine-learning/how-duco-transforms-financial-data-automation-with-amazon-bedrock-agentcore/

A practical four-stage control design

Stage 1: Read. Define which ledgers, invoices, bank feeds and policies the assistant can access. Start with read-only permissions and record the identity under which each source is retrieved.

Stage 2: Investigate. Require the agent to compare source records, identify possible timing differences, preserve document identifiers and flag missing evidence. Do not let a model-generated explanation substitute for a source citation.

Stage 3: Recommend. Ask for a case packet with discrepancy amount, likely causes, alternative explanations, confidence limitations, proposed action and supporting evidence. Require an explicit 'insufficient evidence' outcome.

Stage 4: Approve. Route material adjustments, rule changes and customer-impacting decisions to a named employee. Record the approval, any override and the eventual result so the team can evaluate mistakes later.

Worked hypothetical example

A service business records a $4,800 invoice, but its bank feed shows a $4,650 deposit. A poorly controlled agent might create a $150 adjustment to make the accounts reconcile. An evidence-first agent instead retrieves the invoice, payment record and processor statement, notices that a fee could explain the difference, and asks the bookkeeper to confirm the fee treatment. If the processor statement is missing, it reports the unresolved gap rather than inventing an explanation. This example is hypothetical and is not a Duco result.

What to test before deployment

Use at least three categories of test cases: ordinary timing differences, genuinely missing transactions and deliberately misleading records. Measure the proportion of cases with correct source citations, the rate of unsupported recommendations, the time an analyst spends reviewing a case, and whether the system can be prevented from modifying records without approval. Test access with two users who should see different records. Confirm what happens when an upstream service is unavailable.

A successful pilot should improve the quality or speed of the human decision, not merely the number of automated responses. Some teams will discover that simpler rules and exception dashboards outperform an agent for routine cases. That is a valid result.

Key Takeaways

  • Separate investigation permissions from transaction-changing permissions.
  • Require evidence, alternatives and unresolved questions in each case packet.
  • Validate user authorization in tools and data systems, not only in prompts.
  • Evaluate total review effort and error rates, not only agent response speed.

Next step

EH Synergy's AI Opportunity & Transformation Assessment helps identify which workflows are suitable for AI assistance, which controls are required and where conventional automation may be a better choice. Inquire about the assessment through EH Synergy's established contact channel.

blog author avatar

Erin Harris

Founder of EH Synergy. M.Ed. in Educational Technology & Instructional Design and MBA in Marketing. Erin helps everyday people and small businesses use AI and automation to build real, useful tools.

Back to Blog
Erin Harris, Founder of EH Synergy

About the Author

Erin Harris, Founder of EH Synergy

M.Ed. in Educational Technology & Instructional Design and MBA in Marketing. Erin writes about how businesses can use AI and automation in ways that actually work.

Book a Call →

New to AI? Start with the free course.

The Clear Build Starter: five short lessons to build your first planner or tracker with AI. No coding needed.

Start the Free Course
EH Synergy

EH Synergy LLC · Dallas, Texas

Book a Call
© 2026 EH Synergy LLC