How CFOs Build an AI Roadmap for the Finance Function

By Jane Ha · Updated Oct 11, 2026

A finance AI roadmap is a sequenced plan that says which finance processes get AI, in what order, on which data, under whose controls, and with which skills in the team. It's different from a company-wide AI strategy. The CFO is the one person who owns the close, the forecast, the controls and the budget at the same time, so the roadmap has to respect all four.

This guide is for CFOs, controllers, FP&A leads and finance transformation managers who've moved past "should we try AI?" and now face a longer list of ideas than they have people to run. It covers a five-part method: inventory the processes, score the use cases, do the data work, sequence the releases and plan the skills. For the generic company-level steps, such as maturity stages and the build-versus-buy question, it points to the AI transformation strategy pages instead of repeating them.

Key Facts

• In Deloitte's Q2 2026 CFO Signals survey of 200 North American CFOs, 43% said they feel confident in their organization's current AI governance, and 59% named balancing speed against risk as the biggest challenge to an enterprisewide framework (Deloitte).
• In the same survey, 33% said the chief information security officer owns AI governance, while 19% said the CFO carries the greatest responsibility (Deloitte).
• KPMG's 2026 survey of 1,013 senior finance leaders found the strongest AI gains in decision-making quality, forecast accuracy and responsiveness, which are judgment-heavy areas rather than transactional ones (KPMG).
• In that survey, 36% named data quality as the most cited opportunity and vulnerability, 38% were upskilling existing finance teams on AI, and 28% were rethinking the types of talent they need (KPMG).
• KPMG also reported that organizations formally tracking AI-related KPIs saw a 10-point lift in the share reporting moderate or significant ROI improvement over those that didn't (KPMG).

Why Finance Needs Its Own Roadmap

A company-wide AI plan usually optimizes for adoption and spread. Finance can't. A wrong answer in sales content costs a bad email. A wrong answer in a reconciliation or an accrual can reach the financial statements, and that puts the work under internal controls from day one.

Three things make finance different.

The calendar is fixed. You can't pilot a new tool in the middle of quarter-end. Close, audit and budget season set the windows when change is safe, and a roadmap that ignores them gets pushed out by the first busy month.

Evidence matters as much as output. A tool that saves ten hours but can't show what it did, who checked it and which version ran creates work for your auditors. The control design is part of the use case, not a step after it.

The data is yours to fix, and nobody else will. Chart-of-accounts mappings, vendor masters and entity structures sit in finance's own systems. KPMG's survey puts data quality at the top of the list of barriers, and no vendor can clean it for you.

The KPMG finding about where gains concentrate is useful for ambition. Forecast accuracy and decision quality are where respondents saw the strongest improvement, so a roadmap made only of back-office automation may be aiming low. A balanced one mixes quick transactional wins with a few judgment-heavy bets.

Step 1: Build the Process Inventory

Start with what the function does, not with tools. List every recurring finance process and, for each one, note five things: volume, cycle time, who does it, how many systems it touches and where errors usually appear. A spreadsheet is enough.

Group the list by the cycle your team already uses.

Two habits make the inventory worth the effort. First, interview the people who do the work, not just the owners. The preparer knows that the "automatic" matching still needs a spreadsheet on the side. Second, record the workarounds. A process with three offline spreadsheets is a data problem before it's an AI problem, and you'll want that flagged when you score it.

Also capture shadow use. Someone on the AP team may already paste invoice text into a public chatbot. The inventory is the right place to find out, and AI and finance team roles can help you see who is likely to be doing what.

Step 2: Score the Use Cases

Now match the inventory against what AI can do. For ideas, the finance use cases overview lists the common candidates by function. Don't try to evaluate all of them. Pick the processes where the inventory shows real pain, then score each candidate on the same five criteria.

Here's an illustration of how the scoring plays out. The scores below are made up to show the mechanics, not benchmarks.

Read the table the way a CFO would. Bank reconciliation scores highest not because it's the most valuable but because the data is ready, the review is straightforward and the risk is contained. Cash forecasting has the biggest value but weak data, so it moves to a later release rather than being dropped. Contract review scores low on control risk because errors there reach revenue and disclosures, which is a reason to start it after your controls mature.

Three cautions on scoring. Weight the criteria if your situation demands it, for example doubling control risk if you're preparing for your first SOX year. Don't average away a zero: if a use case has an unacceptable control risk, it waits no matter how high the rest score. And rescore every quarter, because data readiness and team skill change faster than you think.

Step 3: Do the Data Work First

The data column in your scoring table is where most roadmaps quietly fail. A forecast, a match or a flux explanation inherits every inconsistency in the records beneath it. So treat data work as roadmap items in their own right, with owners and dates.

The usual finance candidates:

• Chart of accounts cleanup. Retire dormant accounts, document what "Other" means, and move department and region into dimensions rather than account names.
• Master data. Deduplicate vendors and customers, add persistent IDs and require approval for new records.
• Mappings. Move entity-to-group mappings out of personal spreadsheets and into a controlled place with change records.
• History. Check that the periods a model will learn from are consistent. A system migration three years ago can break a forecast model's trend.

The finance data readiness guide walks through a checklist for each. A useful rule: if a use case scores below 3 on data readiness, schedule the data fix in the release before it, and have the use case wait.

Step 4: Sequence the Releases

Sequence is where the scores, the calendar and the dependencies come together. Most finance functions do well with three releases, not a single multi-year plan.

Release 1: prove and contain (first six months). Pick two or three high-scoring, low-control-risk use cases, such as reconciliation matching, invoice coding assistance or commentary drafts. Run each in parallel with the existing process for a few cycles. Build the evidence habit now: logs, reviewer sign-offs, override records. In parallel, start the data fixes that later releases need. Mirroring the company-level path, this is the move from experiments to a governed pilot described in the five stages of AI maturity.

Release 2: extend and connect (next six to twelve months). Move proven tools into more entities or volumes, and add one judgment-heavy use case where the data is now ready, such as forecasting or scenario work. This is also when your month-end close starts to benefit from several tools working in the same calendar.

Release 3: redesign (after that). With controls, data and skills in place, you can rework the process itself instead of bolting AI onto it. Maybe the close is restructured around continuous reconciliation, or approvals are rewritten around exception handling.

For each release, set a go or no-go gate: accuracy within tolerance, controls documented, owner named, fallback tested. If a gate isn't met, the release slips and the next one doesn't start. Don't schedule anything new inside the quarter-end and audit windows.

On the build-or-buy question, don't invent your own logic. The build versus buy versus integrate decision covers it at company level. In finance, add one test: can the vendor give you the audit evidence your controls need, such as logs, version history and access reports?

Step 5: Plan the Skills and the Operating Model

A roadmap with no people plan stalls at release one. KPMG's survey is a good picture of the common gap: 38% of organizations were upskilling existing finance teams, but only 28% were rethinking the types of talent they need (KPMG). Training people on a tool is one thing. Deciding what the function does differently is another.

Practical moves:

1. Name an owner for every tool and every control. One person is accountable for scope, changes and monitoring.
2. Teach data judgment. KPMG's report singles out the ability to assess data quality and interpret outputs as the most critical capability need. Preparers and reviewers should learn to spot a plausible but wrong output.
3. Change the reviewer's job. As preparation shrinks, review grows. Write down what a reviewer checks and what makes them look closer.
4. Decide who owns AI governance. Deloitte's survey shows ownership scattered: 33% point to the CISO and 19% to the CFO. Settle it for your function early. AI and internal controls covers the SOX side of that conversation.
5. Plan the career path. If entry-level reconciliation work shrinks, say how junior staff will learn the judgment they used to pick up by doing it.

Measure the Roadmap, Not Just the Tools

Pick a handful of measures before release one starts, so you have a baseline. Good starting points are cycle time for the processes in scope, exception and override rates, forecast error, hours of manual work moved to review, and the share of controls with complete evidence. KPMG found that organizations formally tracking AI-related KPIs were ahead of those that didn't, by a 10-point margin on ROI improvement, so the discipline itself seems to help. The measuring AI in finance guide goes deeper on metrics.

Report to the board in the same three buckets each quarter: what shipped, what it delivered against the baseline, and what risk or control issue came up.

Common Pitfalls

• Starting with a tool, not the inventory. You end up with a clever product and a process it doesn't fit.
• Choosing the highest-value use case first. The biggest prize usually has the weakest data. Win something contained first.
• Treating data cleanup as a side project. If it has no owner and no date, it won't happen, and release two slips.
• Piloting in the wrong month. A pilot launched into quarter-end gets judged on close pressure, not on merit.
• No baseline. If you didn't measure cycle time before, you can't prove a gain after.
• Letting governance trail the rollout. Retrofitting controls after go-live is more expensive than designing them in.
• A roadmap for tools, not for people. Reviewers who haven't been trained or re-scoped become the bottleneck.
• Setting it and forgetting it. Rescore the use cases every quarter. Models, vendors and your own data all move.

Where to Go Next

This week, build the process inventory for one cycle, such as record to report, and score its top five candidates on the five criteria. Then pick the two you'd run first, name an owner for each, and list the data fixes they depend on. That small document is the seed of your roadmap, and it will tell you quickly how ready the function really is.