AI for Enterprise Finance & How to Do It Right

Every CFO I talk to at a billion-plus company is grappling with the AI landscape on two fronts:

The horizontal assistant: give everyone in your organization Microsoft Copilot or Claude Cowork. The problem with this is you have every employee spinning up 3 agents, none of them work with each other, and in 3 months you've spent $3M in tokens, with 80% of agents either defunct or constantly breaking in production. You're left with a bill for $3M in token spend, and a graveyard of tech debt with no one to own it.

The point solution: bring in one new software solution for AP, another for the close, and another for expenses. This doesn't work because software made for everyone is not made for you. It doesn't understand that your AP process is 7 steps, not 4, and doesn't understand your exception handling logic. As a result, your employees either don't use it, or do use it but the ROI is sub 15%. And worse, your employees are complaining that this new software is different from their old way of doing things, and half of them think it's a downgrade. Nightmare.

My job is to help CFOs understand the right mix of both solutions for them. There's absolutely a need for the horizontal assistant, but it serves a different purpose. Employees will still have work to do, and this assistant is to help one employee do the work of 10. But you're missing the bigger picture: background agents that just do the work without needing to be prompted or used by an employee at all. Imagine an agent that reads every invoice the second it lands, matches it to the right PO, and either clears it or kicks the one weird one over to the single person who needs to decide, all before your team has opened their laptops. An agent that ties out yesterday's bank activity against the ledger every morning, so the close is basically finished before month-end even starts. An agent that chases every vendor for the missing W-9 or the overdue payment on its own, so nobody on your team writes one of those emails ever again. Nobody prompts these. They just run in the background, and the work is already done when you show up.

For context, I run Varick Agents (@varickagents). We embed with enterprise finance teams and deploy AI agents that operate inside the tools they already use. Finance is where we have seen the fastest and most measurable results, because the work is repetitive, the processes are well-defined, and the cost of the manual version is trivially easy to put a number on.

The goal of this article is to show you how we've done this for several companies at scale, pitfalls that we learned to avoid, and how we measure success after all is said and done. Bonus: how we make sure we're not spending millions on tokens every year, and how we bring hallucinations down to near 0. For reference, we took one client's month end close from 12 days to 5. Simultaneously we reduced error rates by 72%. The value capture was upwards of $45M annually, as a combination of revenue uplift, cost savings, and risk reduction. It's the same playbook every time, even if the agents you end up with are wildly different (hence why point solution software doesn't work).

Failure rates for AI implementations in finance are high

Before the how, let's do a quick snapshot of the numbers. The gap between where finance teams are and where they could be is wide, but the AI results so far have been quite poor.

• Gartner surveyed 183 finance leaders and 84% implemented AI or plan to, but only 7% reported high impact.
• MIT's NANDA group looked at 300 deployments and found 95% of enterprise Gen-AI pilots deliver no measurable return to P&L.
• Gartner expects 40+% of agentic AI projects to be canceled by the end of 2027 on cost, unclear value, and weak risk controls.

So when I say most of this fails, these are the stats I'm referencing. And I'm going to tell you why below, but for comparison, 100% of Varick's finance department implementations have successfully been deployed into production, with measurable positive ROI (average is 5.5x).

Now for the work itself:

• 2/3s of invoices still need one or more humans to touch them. Only a third run straight through (Ardent Partners, 2025). At the clients we've worked with, there's often 3 or more people touching every invoice before it's fully processed.
• A manual invoice costs $12.42 to process end to end.
• Half of all finance teams take over a week to close their books (Ledge, 2025), and 94% of them are still living in Excel somewhere inside that close.
• 14% of invoices get flagged as exceptions, and exceptions are the single most-cited headache in AP. This is the stat I want to draw the most attention to. Your exceptions are different from the next company's, meaning no generic SaaS or product can solve this massive headache in the way that you need it. The need for custom software has never been higher for finance functions, and thankfully, AI is the perfect unlock here.

None of this is a technology problem anymore. It is a workflow problem, a human-glue problem, and that distinction is one I'll be highlighting in more detail below.

Why horizontal assistants (Claude Cowork, Microsoft Copilot) fail

Even if we ignore the token bill (millions per quarter), the bigger problem is even frontier models get finance work wrong a majority of the time. When the frontier models (Fable, Opus, GPT 5.5, etc) were put through 900+ real finance analyst tasks this year, the best one only hit 52% accuracy (Vals AI). Another study ran 19 models on a real chart of accounts and the highest accuracy was 66% (DualEntry). In a finance function, these accuracy levels are catastrophic. Even Microsoft's own docs say don't use Excel Copilot for numerical calculations or anything with compliance implications, which is hilarious because they put the AI in your spreadsheet in the first place.

Hallucinations don't mean a typo in an email. If your AI hallucinates a vendor or blows an intercompany elimination, that's real money out the door to be found and unwound. Lack of auditability is a massive problem as well. "The AI said so" doesn't fly with a SOX auditor.

Your AI agents need to be guardrailed and permissioned, so it only ever does the exact actions you allow it to, determined as a result of a comprehensive AI audit. Every task further gets boiled down to its most deterministic state, so the model only decides the few steps that need judgement, instead of everything end-to-end. This is how accuracy holds above 97%, with agent traces that can be surfaced to an auditor and leadership.

Why more point solutions make things worse

So you skip the generalist and buy a dozen specialists: an AP agent from Ramp and Brex and Bill, collections from HighRadius, close agents from BlackLine and FloQast, all of it shoved into the ERP by SAP and Workday, plus a new AI-native ERP. Do you see where I'm going with this? AI was supposed to be the reason you finally move away from 20 different software vendors each doing something different. You need a single pane of glass that lives across your existing systems. Those systems already have everything an agent needs to run on top of them, no new platform required. Instead, I see CFOs regrettably introduce more software licenses, more surfaces their team has to log into and keep track of, and in the end, almost no efficiency gain to show for it.

What does work

Every deployment in finance departments that works follows the same philosophy: one layer that sits on top of and in between the software you already run, instead of another tool for your team to log into. It reads from your software like NetSuite, Bill, and Workday, moving data between them and doing the work exactly like your team would. Where it needs assistance, it'll flag edits for your team to adjust.

In doing so, you augment the operator, not the task. Right now your tools each automate a slice of the work, but nobody automates the person in the middle, who is copying a number from one screen into another, checking if two figures match, sending the chase email when they don't, escalating when nobody replies. This person is the glue, and the glue is where all the value lies: cycle time reduction means time is saved and more revenue is generated, faster.

If we take this example back to exceptions: picture an invoice that lands with no purchase order. Right now, an AP analyst has to work out who ordered it, then find the right PO by filtering the inbox, then match it, before finally pushing it through. Exceptions are more common than you think; this happens hundreds of times a month.

However, with a unified agent layer, AI catches this exception the second it lands, then searches the PO system by vendor, amount, date, before finally clearing the clean matches itself, just as your analyst would. When the agent isn't sure, it sends the two likeliest POs to an analyst on Slack and asks them to determine which one is right. 15 minutes of digging time becomes thirty seconds for a yes or no, with all the information surfaced ahead of time. The same shaping occurs in bank reconciliation, intercompany eliminations, W-9 chases, payment-status emails, and the auditor's PBC list.

How to implement this system in practice

We do 5 things every time:

1. Forward deployed engineers embed with your department and map every process end to end. Documented processes and SOPs very rarely capture reality, which is what people actually do. For example: "When something goes wrong I check this spreadsheet first" and "I email Sarah directly because the alerts have been broken for 3 years." A real example: "the SOP says invoices get matched to POs in the system." But in reality, they get matched in the system, except when the PO was never created, in which case Brittany emails the department head for a retroactive one, unless it's under $500, in which case she codes it to the department's general expense line and flags it for later. If you were to just build agents based on the SOP, they’d break the first time it hits Brittany, which is coincidentally Day 1 of production. This is why it's incredibly important to sit with people and watch them work. It's the bridge between services (consulting) and software (development), and yet is also the difference between a successful agent rollout and a shot in the dark that dies immediately.
2. Build inside the tools they already use. The agent runs NetSuite or SAP or BlackLine the way a new hire would, logging in and clicking through the same screens and hitting the same APIs. Nobody on your team needs to learn a new interface, and the only thing people notice is that there's less work piling up, exceptions are clearing faster and the month end close is getting shorter.
3. Build agents that do work instead of dashboards. Most "AI for finance" is an analytics tool masquerading as an agent. Do not fall for this trap. The monitoring and reporting come as a result of the agent actions that drive them. Yes, it's helpful to measure KPIs before the build to see if you're really driving change. But if your artifact is a dashboard or chatbot instead of a background agent, you are leaving efficiency on the table. Don't dedicate months to the equivalent of fancy reporting software.
4. Escalate only when real judgment is required, with a confidence gate in front of it that gets better over time. The goal is to take the 70 to 85% that's pure pattern-matching off of your team's plate, so their time is left for high-leverage, high judgement decisions only. Simultaneously, every time they respond to agent actions (with an approve, edit, or reject), this trains the agent, allowing accuracy to climb each week instead of sitting flat or, worse, regressing. This is where AI engineering is critical; your harness can be the make or break between a system that improves and one that fizzles out.
5. Design for the whole department on day one. This is by far the most overlooked aspect of agent implementations at the enterprise level. Imagine each operator picking up a vibe-coding tool, building an agent for just their corner, but it can't scale past their own work. This misses the bigger picture. Oftentimes, their bottlenecks are located upstream. But then the upstream team builds their own agent, which doesn't talk to the one downstream. Very quickly you have dozens of agents, all siloed to their own activities, and no communication, just tech-debt across the organization. Instead, map out the entire organization, understand who is whose bottleneck, and build with that in mind.

Avoid runaway token spend, and agent hallucinations

How you don't spend millions on tokens: a good AI agent is mostly not AI. What we ship is ~85% plain code and ~15% model calls. The models are only used where there's a real need for judgement, like reading a value off of a messy invoice, or sorting an exception into one of your known buckets, or drafting a note for a human to approve. On the other hand, the majority of work is comparisons (math), lookups (filtering), routing (if/then/else statements), and posting (API calls). Compare this to Claude Cowork, where nearly every action is determined stochastically by an LLM. Instead, we have faster, cheaper, more-accurate agents. LLMs were just the unlock.

How you get errors near zero: three layers.

• Deterministic code: it's consistent by design, which is what makes it auditable.
• Evals: a manually created but automatically updated test suite that checks both the answer and the path the agent took, allowing us to catch agents that went somewhere they weren't supposed to, or produced results inconsistent with how we want them to behave.
• Human feedback: every approval and correction your team makes trains the system, and accuracy on a workflow climbs into the high 90s inside a couple of months. We watch GL coding go from about 85% to 97% and up as the corrections stack. And because it's code and evals instead of a black box, you can actually answer the question "why did the agent do this," any time a stakeholder or auditor asks. Horizontal agents can't do this.

How is this measured

Fortunately, when you have agents that live on your systems of record across every single workflow and every slice of software, you now have the ability to track data at the most granular and real-time level. It becomes very apparent that 80% of exceptions are being handled by agents, and the time to reconcile exceptions dropped from 4 days to 2 hours. Some real results:

• Close dropped from 12 days to 5
• Exception handling went from 130 hours a month to 20
• Invoice processing went from 20 minutes each to under 1 minute, on average

There are only 3 buckets of value capture that matter for any AI implementation. Am I saving time/money? Am I increasing revenue? Am I reducing risk? It helps to bucket everything you're measuring into these 3 categories, and measure accordingly for value capture and KPI purposes.

Where to start

Find process owners at your organization, and start with them. Understand at a deep level what their current process is (be prepared to talk to their sub-process owners, analysts, ICs, etc for more information). Get to the bottom of:

• How do things run today, what is the workflow genome, so to speak
• What is the data volume and throughput across each task
• What are the error rates and what is the cost of an error today?
• How are exceptions handled and across what formats?

From there, take your learnings, and begin to map out the following:

• What would AI do vs not do for each workflow? What would a post-AI world look like for each process?
• Across the 3 buckets of value capture, what is the quantifiable amount for each?
• How much time and effort would each build take? What are the risks to each one?

Compare value capture to investment, and you now have your prioritization list.

But in summary, don't buy a platform, and don't stand up a data science team. This entire process doesn't even need to take a full year. You instead need to find people who'll sit with your team, learn the real workflow, and build an agent inside the systems you already run, measuring it every step of the way. If you want to see the end state first, we built a five-minute walkthrough of the process here.

This is exactly what we do at Varick Agents. We've embedded with finance, sales, and operations teams at companies from $1B in revenue all the way up to Fortune 500 giants doing over $50B, and we build agents that run their tools inside their existing systems. We only take on a handful of new engagements a quarter, and we're scoping the fall cohort now. If your close is still two weeks and your best people are still doing data entry, come find us at varickagents.com.