Use case
Back-office workflows
Reconciliations, approvals and onboarding are prepared by agents with the evidence attached, then routed to the right approver. Nothing is posted or paid without a person signing off.
The problem
How the work goes today
A finance clerk downloads the bank statement, exports the ledger, and matches lines in a spreadsheet. Most match on amount and date. The rest take the afternoon: partial payments, bundled transfers, references that don't quite match.
Approvals arrive by email with attachments. The approver can't easily see the policy limit, the budget remaining or whether the same request came through last week, so they either approve on trust or send it back with questions.
Onboarding a supplier or an employee touches five systems and three teams. Steps are tracked in someone's inbox, and when an auditor asks who approved a bank-detail change, finding the answer takes days.
59%
Share of finance leaders using AI in the finance function in 2025, up only from 58% in 2024, in a Gartner survey of 183 finance leaders. Accounts payable automation and anomaly detection were among the most common uses.[1]
Over 40%
Share of agentic AI projects Gartner expects to be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. This is a forecast, and the reason we keep scopes narrow, measured and behind approval steps.[2]
How it works
Step by step
Highlighted steps are where a person decides. Everything else runs on its own and is logged.
Systems involved
- ERP (SAP, Oracle, Microsoft Dynamics, NetSuite)
- Close and reconciliation (BlackLine, FloQast, Trintech)
- Procurement (Coupa, SAP Ariba)
- HRIS (Workday, SAP SuccessFactors, BambooHR)
- IT service management (ServiceNow, Jira Service Management)
- Identity (Okta, Entra ID)
- Banking portals and statement files
- Workflow and RPA (Temporal, UiPath)
- 01System
Trigger
A bank file, request form, HRIS event or scheduled close task starts the workflow in the workflow engine.
- 02Agent
Gather records
The agent pulls the relevant records from the ERP, bank statements, HRIS or CRM, and extracts data from attachments.
- 03System
Apply rules first
Matching on amounts, IDs, tolerances and policy limits runs as fixed rules in the reconciliation or workflow tool.
- 04Agent
Handle the fuzzy residue
For items the rules can't settle, the agent matches on descriptions, classifies the exception and drafts an explanation.
- 05Agent
Propose
It prepares a proposed match, journal entry, approval recommendation or onboarding checklist with the evidence attached.
- 06System
Route by authority
The request is routed to an approver chosen by amount, risk and segregation-of-duties rules, so the preparer is never the approver.
- 07Person
Approve
The approver sees the evidence and the agent's reasoning and approves, edits or rejects in the workflow tool. GL postings above materiality, write-offs, payments, bank-detail changes and sensitive access grants always stop here.
- 08System
Execute
Approved items are posted or provisioned through the ERP, HRIS or identity provider's API.
- 09System
Log for audit
Inputs, model and prompt versions, outputs and the approver's identity are stored for each item.
- 10System
Monitor
Exception queues, SLA timers and changes in the auto-match rate are tracked on a dashboard the process owner reviews weekly.
Measurement
What we measure
Agreed before anything is built, with a baseline taken from how the work runs today.
- Auto-match rate
- Share of items matched or processed with no human touch, reported separately for the rules stage and the AI stage so you can see which one is doing the work.
- Exception rate and age
- Share of items needing review, and the number of days each exception has been open, by reason.
- First-pass approval rate
- Share of agent-prepared items approved without edits or being sent back.
- Cycle time
- Days to close the books, or hours from onboarding request to fully provisioned, measured against the baseline before launch.
- Post-approval error rate
- Corrections found after approval, in later review or audit, per thousand items processed.
- Audit findings
- Number of internal or external audit findings related to the automated process in each review period.
Pitfalls
What usually goes wrong
- Automating a broken process
- Map the current process, remove the steps nobody needs, and measure a baseline before any automation is built.
- A language model doing exact matching
- Match on amounts and IDs with rules, use the model only for what rules can't settle, and record which component made each decision.
- Agents with too much access
- Give each agent a least-privilege service account, keep development and production separate, and put an approval step in front of every write.
- Auditors can't follow what happened
- Store inputs, model and prompt versions, outputs and approver identity for every item, in a form an auditor can query.
- Fraudulent bank details during onboarding
- Never accept bank details from email alone and verify every change by calling a number already on file.
Build or buy
When a ready-made tool is enough
Custom work isn't always the answer. Here's how we'd decide.
- Close and reconciliation suites (BlackLine, FloQast, Trintech), procurement suites, the AI agents built into ServiceNow or Workday, and RPA platforms (UiPath, Automation Anywhere) are enough when the process sits inside one product.
- Custom work pays off when the workflow crosses several systems with no single suite in charge, which is the usual situation in mid-size companies.
- It also makes sense when the rules are specific to your business, when legacy systems have no API, or when you want one approval and audit layer across several agents.
Industries
Where this applies
Sources
- 1.Gartner, Survey shows finance AI adoption remains steady in 2025 (Nov 2025)
- 2.Gartner, Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (Jun 2025)
Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.
Next step