AI mandate expectations: high.
Margin for error: zero.
A practical hub for Controllers and finance leaders navigating the pressure to implement AI — without sacrificing the accuracy, auditability, or trust their teams depend on.
Introduction
Somewhere in the past 18 months, the question changed. It's no longer "should we be using AI?" It's "why aren't we further along?" Controllers and finance leaders across every sector are receiving the same directive: move faster on AI. The expectations are high. The margin for error is zero.
This playbook exists because most AI mandates fail not from a lack of ambition, but from a lack of sequence. Teams buy tools before their data is clean. They run pilots before they have baselines. They automate workflows that aren't documented.
There is a right order. This is it.
01
The State of AI Mandates
The pressure controllers are feeling is real, and the data confirms it. According to Deloitte's Q4 2025 CFO Signals survey, 87% of CFOs expect AI to be extremely or very important to their finance department in 2026 — a near-consensus view. More than half said integrating AI agents into finance workflows is a top transformation priority this year.
What the data also shows is a significant execution gap. L.E.K. Consulting found that only about 11% of CFOs are using AI meaningfully in finance, while roughly 35% are still at the pilot stage. An RGP survey found just 14% had seen clear, measurable impact — and 86% cited legacy tools and fragmented systems as significant barriers.
While the expectation is nearly universal, the execution is genuinely hard. That gap between what leadership wants and what responsible implementation requires is exactly the terrain this piece is designed to help you navigate.
Why Most AI Mandates Fail Before They Start
The failure modes in AI mandate execution are consistent enough across organizations that they're worth naming directly — not as a warning, but as a map of what to avoid.
| Failure Mode | What It Looks Like |
|---|---|
| The 18-month roadmap | Grand transformation timelines that assume a stable landscape. Iterative, outcome-based sprints work. Grand roadmaps don't. |
| Pain-point thinking | Teams go straight to the technology before they've articulated the value they're trying to create. Fixing pain points optimizes the past. Defining outcomes builds toward the future. |
| The finance silo | AI mandates executed inside finance produce finance-only results. When the initiative lives entirely in finance, you end up with tool proliferation solving narrow problems but creating new coordination overhead. |
| The headcount fear | When a CFO frames an AI mandate as a headcount reduction initiative — even implicitly — adoption stalls because the people who best understood the work stopped contributing. |
"People do not resist technology. They resist change that threatens their identity."— David Fuhriman, CFO, Jewish Federation of San Diego
02
Setting a Foundation
There's a reason most teams receiving an AI mandate find themselves stuck early: they're trying to decide what to build before they've assessed what they have to build on.
As CFO of the Jewish Federation of San Diego, David Fuhriman's framing cuts through the noise: AI is much more a capability to build than a tool to acquire. If AI is a tool, you buy it, install it, and wait for results. If it's a capability, you build the organizational conditions that allow AI to be useful.
David Fuhriman's Five Foundations
| Foundation | What it means in practice |
|---|---|
| Data Readiness | If finding the answer to "How many active customers do we have right now?" requires pulling from multiple systems or debating definitions, the data isn't ready for AI. |
| Process Maturity | Could a new employee follow your workflows from documentation alone? If the answer involves pointing them to a specific person, the process exists as tribal knowledge — and tribal knowledge cannot be automated. |
| Human Capital | AI frees the team to finally get to the work they've been deferring. When that's the frame leadership communicates, the institutional knowledge holders stay engaged. |
| Governance | Good governance requires clear answers: What can AI access? What review applies? Who is accountable when something goes wrong? When do you disclose AI was involved? |
| Technology Infrastructure | This is where most organizations start — but it should actually be the last step. The right technology becomes obvious once the first four foundations are in place. |
"AI doesn't fix bad data. It amplifies it."— David Fuhriman, CFO, Jewish Federation of San Diego
Where to Start Applying AI: Outcomes First
Start with: what does the CFO need to be able to do that they can't today? Then work backward to identify which process constraints are preventing it. Outcome first, process second, tool third.
- Repetitive, rules-based tasks with documented inputs and outputs
- Recurring reconciliations on stable, clean data sources
- Routine AP workflows: invoice coding, payment runs
- Reporting that runs on a fixed cadence with consistent structure
- Exception handling, edge cases, anything that breaks the pattern
- Decisions that require business context AI doesn't have access to
- Variance explanations that require understanding why, not just what
- Controls review and sign-off — the human remains in the loop
Phase 1
Get the Underlying Data Right
Before any AI implementation can succeed, your data needs to be accessible, clean, and consistently defined. Most teams discover during this phase that their data isn't as ready as they assumed — and that the work of fixing it is unglamorous but foundational.
Ask your team: "How many active customers do we have?" If the answer hedges, your data isn't ready.
What Data Readiness Actually Requires
| Requirement | What it means |
|---|---|
| Accessibility | Data behind a manual CSV export isn't accessible for automation. API access to source systems is the baseline. |
| Consistency | Single sources of truth per entity. Consistent definitions across systems. |
| Documentation | Definitions that exist only in people's heads are not definitions. AI can only work with what is explicitly stated. |
Phase 2
Calibrate Risk to Your Company Complexity
Not all accounting processes carry the same risk profile. Phase 2 is about honestly mapping which parts of your operation can tolerate experimentation and which require near-zero error tolerance — before you commit to any implementation path.
| Old Approach | New Approach | Why It Matters |
|---|---|---|
| 18-month transformation roadmap | 30-90 day outcome-based sprints | The AI landscape shifts faster than long plans can accommodate |
| Requirements fixed upfront | Iterative requirements that sharpen over time | You don't know what AI can do in your environment until you run it |
| Pain point as the starting question | Business outcome as the starting question | Fixing a pain point optimizes the past |
| Finance-led, finance-owned | Cross-functional steering committee | Finance-only AI produces finance-only results |
Phase 3
Get Proof at Your Scale
This is the phase most teams skip. They move from pilot to production before establishing genuine benchmarks. Real proof means having before/after data that holds up to scrutiny, not just a vendor case study or an impressive demo.
"Ask questions of the business, not the system. When I evaluate an AI tool for a finance team, I want to know: can it answer a question a CFO would actually ask — across all the systems that question touches?"— Dean Quiambao, Armanino
What Good Vendor Evaluation Looks Like
| Ask the Vendor | What the Answer Tells You |
|---|---|
| Can you show us value within 30-60 days? | Whether their delivery model is genuinely iterative |
| Can we speak with a reference at our scale, in our ERP? | Whether their success stories represent your environment |
| What broke during implementation for that reference? | Whether they're honest about friction |
| Is the team still using it 3 months after go-live? | Whether adoption held |
Phase 4
Build an Experimentation Practice
With clean data, calibrated risk, and genuine proof points in hand, you're ready to move fast — but only in the right environment. Phase 4 is about maximizing what you learn before anything touches your general ledger.
The word "experimentation" can imply a lack of structure. What works is the opposite: a deliberate practice of small, sandboxed automation work that produces real output, builds real capability, and generates the internal evidence base that makes Phase 5 decisions defensible.
"The thing about coding with the purpose of finding efficiencies is that it's very contagious. Once you start, you cannot stop."— Francisco Meyo, Abridge
What to Experiment On
- CSV-to-journal-entry automation for a recurring workflow
- Data transformation between systems with known mapping logic
- Custom report built on a clean data export
- Automations replicating a manual process you understand end to end
- Anything that posts to the ledger without human review
- Automations touching data you'd need to explain to an auditor
- Full integrations requiring security or infrastructure oversight
- Workflows where edge cases aren't documented yet
Phase 5
Implement Carefully in the GL
Production implementation in the GL is a different discipline than experimentation. Phase 5 is about moving from sandbox to real workflows with the rigor those workflows require — rollout plans, fallback procedures, and the human change management that determines whether any of it sticks.
"Auditability, observability, repeatability — when you're building agents, those three things are really, really important."— Tom Alexander, CrossCountry Consulting
Cross-Functional Governance
Once automations move to the GL, they stop being finance projects. They become company infrastructure — touching systems IT never reviewed, data legal hasn't approved, controls that external auditors have opinions on.
| Steering Committee Member | What They Need to Own |
|---|---|
| CFO | Strategic outcome definition, cross-functional authority |
| Controller | Implementation plan, phased rollout, surfacing blockers |
| CIO / IT | Vendor access approval, data security, infrastructure |
| Legal | Data handling, third-party AI contracts, regulatory exposure |
| Internal Audit | Control design review, audit trail requirements |
Build vs. Buy
Build vs. Buy — What to Consider
AI has made building dramatically cheaper and faster. But the shift doesn't simplify the decision. It moves the hard question from feasibility to maintenance: how costly will this be to maintain?
The teams that will look smart in two years aren't necessarily the ones that built the most. They're the ones that built the right things — and bought the rest from vendors who've thought harder about the infrastructure, security, and audit requirements than any accounting team could reasonably sustain on its own.
- Data transformation scripts
- Process automations with known inputs and outputs
- Commissions calculators
- Internal workflow automations
- Tools that enhance data quality
- Payroll tooling
- AI-powered reporting & flux analysis
- Anything with severe legal exposure
"We're not in the business of creating software. We're in the business of allowing decision makers to have financial information on a timely, accurate basis."— Francisco Meyo, Abridge
Closing
What to Bring Back to Your CFO
At some point, you have to close the loop. The CFO or board that issued the mandate is going to ask for an update, and what you bring to that conversation matters more than most controllers plan for in advance.
The stronger move is to show judgment. Here's what we evaluated. Here's what we ruled out and why. Here's the risk framework we're applying. Here's what our first deployment will prove — and here's what it won't.
The controller who walks into that CFO conversation with a phased plan, a risk framework, and a 90-day proof point isn't responding to a mandate. They're leading one.
"The question is not whether your organization can build these foundations. The question is whether you will."— David Fuhriman, CFO, Jewish Federation of San Diego
The work is infinite — which means the opportunity is too. Start where you are. Use what's relevant now, and know that the rest will be here when you need it.