How To Implement AI In Finance

It's 11:47 PM on the last business day of the month. The controller at a 45-person services firm is still at her desk, matching a stack of vendor invoices against purchase orders, chasing two colleagues for missing receipts, and trying to close the books before the morning standup. She does this every month. It takes three days. None of it requires judgment she couldn't already do in her sleep — it's reconciliation, data entry, and reminder emails.

That's the kind of work AI handles well. Not the judgment. Not the close. The reconciliation, entry, and follow-up that nobody enjoys and everybody needs done correctly. The rest of this article walks through how to actually implement that — starting with mapping the workflow, not the technology.

Where AI Actually Fits in a Finance Operation

Most "AI in finance" conversations start with the tool. They should start with the workflow. McKinsey's research on AI in business functions has consistently found that the organizations getting real value aren't the ones buying the most sophisticated models — they're the ones that picked a specific, repetitive process and rebuilt it around human review points.

In practice, the workflows where AI agents tend to earn their keep share three characteristics:

• High volume, low variance. Hundreds of similar transactions per month, following predictable rules. Invoice coding. Bank reconciliation. Expense categorization. Customer billing adjustments.
• Clear source of truth. The agent can pull data from a defined system (ERP, bank feed, CRM) and compare it against a known reference (PO, contract, rate card).
• Reversible decisions or pre-decision review. Either the action is low-risk enough to automate, or a human sees the agent's recommendation before anything is committed.

What doesn't fit: anything requiring novel judgment on a one-off basis, anything where the data inputs are unreliable, or anything where a single error compounds silently for months. Financial close itself — the actual decision-making and sign-off — stays with humans. The prep work around it is what gets automated.

Before You Touch Any Tool: Map the Workflow

The first thing we do with a finance client is spend a week mapping the workflow before we recommend a single piece of technology. Not because we're conservative — because we don't want to automate a broken process.

Here's what that looks like concretely:

1. Pick one process. Not the whole back office. One process. Often it's accounts payable, since the volume is high and the rules are clear.
2. Document every step. Who receives the invoice. How it gets into the system. What gets matched against what. Who approves. What triggers an exception. Where it lands in the general ledger.
3. Mark every decision point. At each fork in the road, note who's deciding today, what information they're using, and what happens when they're wrong.
4. Identify the bottleneck. Usually it's not the matching — it's the wait. Invoice sits in an inbox. Approval sits in a queue. Exception sits with someone who's out of office.

Only after this map exists do we talk about where an agent fits. The map is also what you'll use to measure success afterward. If you can't point to a specific step that used to take 4 hours and now takes 4 minutes, you don't have a result — you have a tool.

A Concrete Implementation: Invoice Processing for a Services Firm

Here's a real implementation pattern, drawn from a mid-sized professional services client. The firm had roughly 800 vendor invoices a month, a 2-day average cycle time, and a controller doing manual reconciliation for anything over $2,500.

The workflow before AI:

1. Vendor emails invoice to AP inbox.
2. AP clerk downloads, opens ERP, creates a bill.
3. Clerk attempts 3-way match (PO, receipt, invoice). If it matches, files it for batch payment.
4. If it doesn't match, clerk emails the requester for clarification. Waits. Chases. Eventually either resolves or escalates to the controller.
5. Controller manually reviews anything over $2,500 before payment run.

Average time per clean invoice: 6 minutes. Average time per exception: 3 days.

The workflow after:

1. Invoice arrives in AP inbox. An AI agent ingests the email, extracts the PDF, pulls vendor, amount, line items, PO reference, and due date into a structured record.
2. Agent checks the ERP for the matching PO and receipt. Three-way match attempts happen automatically.
3. Clean match: Agent flags the bill for the next payment run and posts a notification to the controller's dashboard. No human action required unless the controller wants to review.
4. Exception (missing PO, amount mismatch, new vendor): Agent doesn't post. It drafts an email to the requester with the specific gap, sends it, and logs the case. If no response in 48 hours, it escalates to the AP lead with the original thread attached.
5. Over-threshold review: Any bill over $2,500 is queued for the controller with a one-click approve/decline view. The agent's matched evidence is shown alongside.

Average time per clean invoice after deployment: 45 seconds of human attention (a glance at the dashboard, batch approval). Average time per exception: 8 hours instead of 3 days, because the chase happens automatically and the right person sees it on day one.

The controller's role didn't go away. It shifted from data entry and chasing to reviewing the 5-10% of cases where judgment actually matters. That's the augmentation pattern, and it's the one that holds up under audit.

Approval Gates, Review Points, and What Happens When Things Break

An AI agent in a finance workflow is not a tool that runs and hopes. It's a system with explicit gates. Every implementation we ship has the same three layers:

1. Confidence thresholds. The agent doesn't act on every prediction equally. A 98% match confidence on a recurring vendor invoice with a known PO? Auto-post. A 71% confidence on a new vendor with no PO? Route to human review, full stop. The thresholds are set with the controller, not by the agent.

2. Approval gates for irreversible actions. Posting a bill is reversible. Releasing a payment is not. Anything that moves money out of the bank account requires a human click, even if the agent has done all the preparation. According to Gartner's research on finance transformation, the highest-performing finance organizations are the ones that automate aggressively internally while keeping tight controls at the external boundary — payments, customer-facing communications, regulatory filings.

3. Fallback behavior. Every agent has a defined failure mode. If the bank feed is down, the agent doesn't guess — it stops the reconciliation step and notifies the AP lead. If the email parsing fails on a non-standard invoice format, the invoice goes to a human with a flag, not into a "best guess" bucket. If the ERP is unreachable, the agent queues work locally and retries with backoff, alerting only after sustained failure.

These aren't edge cases. They're the system. We design the fallback before we design the happy path.

Measuring Whether It's Working

"Implementing AI" isn't an outcome. It's a means to one. The metrics that matter are operational, and they should be set during the workflow mapping step so you can measure the before and the after honestly.

For invoice processing, the standard set is:

• Cycle time per invoice (clean and exception paths separately)
• Exception rate — and within that, how many exceptions are resolved without human intervention
• Straight-through processing rate — the percentage of invoices that post without a human touching them
• Cost per invoice processed — the IMA and APQC benchmarks are widely cited; track your own trend, not a vendor's number
• Error rate at month-end — specifically, post-implementation reconciliation adjustments. This should not increase. If it does, the agent's confidence thresholds are wrong.

Harvard Business Review has written about the gap between AI pilots and operationalized AI, and the consistent finding is that the organizations that succeed are the ones that define a small number of operational KPIs upfront and tie the project to them — not the ones that track "time saved" in vague aggregates. HBR's coverage of AI in business functions repeatedly makes the same point: the value is in the specific, measured workflow, not in the platform.

Plan to review these numbers monthly for the first quarter, then quarterly. The first 30 days will surface the threshold tuning that needs to happen. After that, you should see the exception rate drop as the agent learns your vendor patterns and your team learns where the agent is reliable.

Common Pitfalls When Implementing AI in Finance

A few patterns we see repeatedly when implementations go sideways:

Starting with the model instead of the workflow. If someone pitches you a finance AI and the first slide is about the model architecture, leave. The model is the easy part. The workflow mapping, the ERP integration, the exception handling — that's the work.

Underestimating integration. A finance AI agent is only as good as its connection to the systems of record. If pulling a PO from the ERP takes 4 seconds and you do it 800 times a month, that's an hour of latency that has to be designed for. Plan the integration before you sign the statement of work.

Ignoring the human review load. If the agent routes 50% of invoices to a human for review, you've moved the work, not removed it. The goal is to reduce the human's time, not the system's. Tune the confidence thresholds until the human review queue is the genuinely ambiguous 5-10%, not the agent's hedge against being wrong.

Treating it as an IT project. It's an operations project. The controller, the AP lead, and the people doing the work today need to own the implementation. IT supports it. The operations team runs it.

Frequently Asked Questions

How long does a typical finance AI implementation take?

For a single process like invoice processing, expect 4-8 weeks from kickoff to production. Week 1-2 is workflow mapping and integration scoping. Week 3-5 is agent build and ERP integration. Week 6-7 is shadow mode — the agent runs alongside the human, no actions taken. Week 8 is controlled rollout with approval gates. The timeline that should make you suspicious is any vendor quoting a finance AI in production in under two weeks; they are skipping steps.

What ERP systems do you integrate with?

We work with the common mid-market ERPs — QuickBooks Online, NetSuite, Sage Intacct, Microsoft Dynamics Business Central — and have built connectors for each. For systems we haven't integrated before, the timeline extends by 2-3 weeks depending on the API surface. We don't recommend screen-scraping or file-drop automations for finance workflows; the audit trail matters too much.

What happens to the AP team when an agent takes over the routine work?

In every implementation we've done, the AP team has been redeployed, not reduced. The 6 minutes per clean invoice frees up roughly 60-70 hours a month at the volume described above. That time gets reallocated to vendor relationship work, accruals, audit prep, and the kind of analysis the controller has never had bandwidth for. McKinsey's research on AI in finance functions has found this pattern consistently — the work changes, it doesn't disappear.

How do you handle the audit trail?

Every agent action is logged with the input, the model's confidence, the action taken, and the timestamp. Every approval gate records who clicked and when. This log is exportable and is designed to satisfy SOX and external audit requirements from day one. We treat the audit trail as a feature, not a retrofit.

What's the cost range for an implementation like this?

It depends on volume, ERP complexity, and how many approval gates the client wants. Single-process implementations for mid-market services firms typically run in the low five figures for build plus a monthly operations fee. We'd rather quote against a mapped workflow than give you a number that becomes wrong the moment we look at your actual data.

What to Do Next

If you're a finance leader at a services firm and any of this sounds like your operation, the next step isn't a demo. It's a workflow audit. We sit down with your team, map one specific process end to end, and tell you honestly whether it's a good fit for an agent, what the build would look like, and what the operational metrics would be. If it's not a fit, we'll say that too.

Book a free AI automation audit and we'll start with the workflow, not the technology.

Related Resources

• About Omni Studio
• AI for Real Estate
• Omni vs Jobber