Deloitte asked finance leaders about AI this year. 63% said they have fully deployed it. Of the teams actively using it, 21% say the investment has delivered clear, measurable value.
Read that again in budget terms. Most finance functions are now paying for AI. About one in five can show what came back.
The 21% are not better at AI than everyone else. They made three decisions before they scaled, and each one is a decision you can still make.
1. They picked one task, not a category
The projects that cannot prove value started with "AI for finance." A tool arrived, the team was told to use it, and no single outcome had an owner. There was no baseline, so there was nothing to measure at the end of the quarter.
The 21% started at task level. Not "automate the close," but the multi-entity bank reconciliation. The marketing accrual. The intercompany tie-out. One task has a named preparer, a known duration, an error rate, and someone on the team who knows exactly what good output looks like.
That scope discipline is what makes the return provable. You can put a number on 40 hours of reconciliation preparation moving from build to review. You cannot put a number on "our team uses AI now."
2. They put the work inside the close, not beside it
The second pattern is more subtle. A lot of finance AI lives in a separate window. Someone opens it, asks it something, copies the answer into the spreadsheet they were already building. The spreadsheet still gets built. The data still gets pulled by hand. The close is exactly as long as it was, and now there is one more tab open.
That is the rebuild with an assistant bolted on. Every month the team pulls the same data from the same systems, reconstructs the same spreadsheets with the same formulas, drafts the same entries, and writes the same explanations. Context resets. The cycle repeats. An assistant sitting next to that process does not shorten it.
The 21% put the work inside the close itself. The reconciliation is tied before anyone opens the task. The working paper is built, with live formulas, in the task where the working paper belongs. The journal entry is drafted from source data and waiting for approval. The team opens to reviewed output instead of a blank sheet.
That is the difference between a tool your team uses during close and a close that arrives prepared.
3. They set the controls before they scaled, not after
The most common objection to AI in accounting is the one a finance leader put to us directly: same prompt, same data, ten different outputs.
That objection is correct, and it is fatal. A close cannot run on a system that answers differently on Tuesday. Neither can an audit.
So the 21% resolved trust first. The teams that got past pilot had four things in place before they expanded scope:
- Output that reproduces exactly, period after period.
- Reasoning they can open and inspect rather than take on faith.
- A person approving every entry before it posts.
- An audit trail that builds as the close runs rather than as a project after it.
The teams that scaled first and added controls later are the ones now explaining a number they cannot trace.
The fourth thing, and it decides more pilots than the other three
Cost predictability. Most AI tools bill by usage, so the invoice moves every month. For a finance leader who has to forecast the line, an unpredictable invoice is not a pricing detail. It is the reason the pilot never becomes a rollout, because nobody commits to a number they cannot state in advance.
The 21% knew their annual cost on day one.
What this looks like in a real close
Ledge is the platform where the close arrives prepared. Accounting Agents do the preparation work inside your close checklist. Reconciliations tied. Working papers built with live formulas your auditors can trace to source. Journal entries drafted from source data, formatted for NetSuite, posted when you approve. Flux explained down to the transaction.
Against the four decisions above:
One task at a time
You pick the task. An Accounting Agent gets built for it, runs it every period, and gets sharper as your team corrects it. Then you pick the next one. Teams typically have their first agent running inside two weeks and the close prepared within 30 days.
Inside the close, not beside it
See, manage, and automate on one platform. The agents execute inside the same checklist that tracks the close, which is why the status you see is real rather than reported. Data comes from NetSuite and 150+ integrations, with live bank feeds over the Swift network and direct APIs.
Controls first
Every Accounting Agent runs bespoke, deterministic code. The AI does the design work once, writing the logic for your reconciliation or your accrual, and that code then runs the same way every period. Same input, same output. Glass-box reasoning, so every step is inspectable. Nothing posts without a person approving it.
Predictable cost
A flat subscription. Unlimited agents, no usage meter, no implementation fee. The number is the same in January and December.
"Scale reconciliation without scaling headcount. We went live quickly without R&D or costly implementers and saw very fast time to value."
SVP Finance, global payroll and payments platform
"Taking out the long tail of 60%+ of the more menial tasks is SUPER valuable, and not just because it will save me money. It will allow those high-performing members of my team that have been trapped down the chain to be unlocked."
CFO, enterprise software company
That second quote is the part the ROI conversation usually misses. The measurable return is close days and preparation hours. The return that changes the function is the team you hired for analysis finally doing analysis.
Where to start
Pick the task your team dreads most and put a number on it. Hours, preparer, error rate, how late it runs. That number is your baseline, and it is the reason the 21% can prove what they got.
Source: Deloitte, Finance Trends survey, 2026. Customer quotes are anonymized to protect confidentiality.
