When AI and finance disagree, test the decision - not the confidence

The risk is not trust or distrust; it is untested reliance

Board's 2026 Planning Intelligence Report says 31% of surveyed executives would follow an AI recommendation even when it conflicts with their judgment; the reported share rises to 48% among CFOs. The same release says only 39% reported formal governance and escalation for AI-driven decisions. Those figures come from a planning-software vendor's survey and describe reported attitudes, not observed decision quality. They are still a useful control signal: finance teams need a protocol for disagreement before a consequential recommendation arrives.

A separate September survey from Esker, another vendor, asked 338 finance leaders about AI authority, governance, and value. Its headline is almost the mirror image: finance leaders want more proof, core-system integration, and human control before granting more decision authority. The two surveys should not be averaged into a universal adoption statistic. Together they expose the gap between influence and operational governance.

Human review alone does not close that gap. A reviewer can defer to an answer because it is detailed, agrees with the desired outcome, arrived first, or appears quantitative. A senior finance leader can also dismiss a correct signal because it contradicts a forecast they sponsored. Appropriate reliance means accepting correct advice and rejecting incorrect advice case by case. It is not maximum trust, minimum trust, or an instruction to split the difference.

NIST's AI Risk Management Framework treats human-AI configurations as context-specific and calls for differentiated roles, accountability, documented oversight, and tested risk management. Current research likewise shows that AI guidance can alter human judgment and bias, while cognitive forcing such as requiring an independent first view may reduce overreliance at the cost of time and fatigue. Finance should spend that friction where consequences justify it.

Make disagreement observable before making it resolvable

If the human sees the AI recommendation before recording a view, the organization loses evidence about independence. The later “human decision” may be an edited echo. For material decisions, capture the finance owner's first view, assumptions, range, and reasons before revealing the model output. Then freeze the AI recommendation with tool and workflow versions.

Do not ask a vague question such as “who is right?” Break the disagreement into atomic claims: which population is included, which period applies, how a metric is calculated, which assumption drives the forecast, what probability is assigned, what threshold triggers action, which constraint binds, and what consequence follows. Many apparent judgment disputes are really cut-off, unit, sign, or definition errors.

The decision owner should also define a loss function in plain language. Is the larger harm excess inventory, missed revenue, covenant breach, a delayed close, a misleading external statement, or an unfair people decision? Two analyses can use identical facts and prefer different options because they optimize different harms. Making that tradeoff explicit is a governance decision; it should not be smuggled in through a prompt or default threshold.

Use the full challenge only when consequence warrants it. Low-risk, reversible choices can use pre-approved rules and sampled review. Material or hard-to-reverse decisions deserve independent views and specialist escalation. This tiering prevents “human in the loop” from becoming either a universal bottleneck or an empty checkbox.

Use a versioned disagreement record

decision_id: FIN-2026-1047
purpose: approve Q4 demand forecast for inventory plan
owner: CFO-02
authority: board-approved planning policy v7
consequence: high
reversible: partially
human_first_view:
  recommendation: hold baseline at 6.0% growth
  range: [3.5%, 8.0%]
  recorded_at: 2026-10-03T08:10:00Z
ai_view:
  recommendation: raise baseline to 10.2% growth
  model: approved-planning-model@4.3
  workflow: demand-review@2.1
  output_sha256: "9ac4..."
disagreement:
  - claim: pipeline conversion remains at 28%
    status: CONFLICT
    test: cohort conversion by stage and age
  - claim: top-customer expansion is repeatable
    status: UNSUPPORTED
    test: signed orders and concentration scenario
decision_status: PENDING_HUMAN_DECISION

The record keeps provenance without pretending that provenance proves correctness. It should link every material claim to a source, calculation, estimate, policy, or judgment. When a reviewer changes position, record which evidence changed it. A clean audit trail says “the cohort analysis invalidated my initial assumption,” not merely “human approved AI.”

Run the decision challenge in ten steps

1. Classify the decision

Name purpose, owner, authority, amount, affected parties, deadline, reversibility, downstream actions, and consequence. A formatting choice does not need the same gate as liquidity, covenant, hiring, accounting, pricing, or external reporting.

2. Freeze independent positions

Capture the human first view before AI exposure when practical. Freeze the model output, prompt/workflow, source cut-off, tool version, and citations. If independence was already lost, say so; do not reconstruct a fictional first view.

3. Validate the source population

Reconcile entities, periods, currencies, scenarios, versions, joins, exclusions, and cut-offs. Put conflicts in an exception register. An AI can be directionally clever while reasoning over the wrong population.

4. Reproduce material calculations

Use approved spreadsheets, planning systems, SQL, or code. Tie to control totals and preserve formulas or code. Separate a calculation error from a disagreement about assumptions.

5. Atomize the dispute

Create one row per claim. Compare positions, evidence, uncertainty, test, owner, and status. This stops a correct observation in one part of an answer from laundering unsupported conclusions elsewhere.

6. Run sensitivity and break-even tests

Vary one assumption at a time. Show where the preferred option changes, what downside violates policy, and which input deserves more evidence. A forecast without sensitivity is a point estimate wearing a suit.

7. Challenge both sides symmetrically

Ask for the strongest disconfirming evidence, missing alternative, selection effect, incentive, and failure mode for both positions. Do not reserve skepticism for the AI or the person.

8. Route specialist authority

Accounting policy, tax, treasury, legal, HR, security, compliance, internal audit, or model risk may own part of the decision. The CFO cannot convert specialist uncertainty into approval by confidence.

9. Build bounded options

Include accept, reject, modify, defer, limited pilot, or reversible test where applicable. For each, name evidence, consequences, monitor, exit threshold, and owner.

10. Approve the exact version

Bind approval to one recommendation, amount, period, destination, model/workflow version, and action. Preserve dissent and unresolved uncertainty. Monitor the assumption that distinguished the winning option and define correction before execution.

Worked example: AI recommends a higher demand forecast

An FP&A model recommends 10.2% Q4 growth. The CFO's independent view is 6.0%, with a 3.5%-8.0% range. The model cites a larger pipeline, improving conversion, and an expansion order from the largest customer. The difference could change inventory purchases and cash needs.

Reconciliation finds that the model counted renewed opportunities and included one unsigned expansion. Deterministic cohort analysis shows early-stage pipeline grew, but aged opportunities convert below the assumed 28%. The largest customer's expansion would lift growth above 10%, yet concentration risk and contract timing make it unsuitable as the base case.

The correct result is not “the CFO beat the model.” The challenge changed both positions. Finance adopts 7.1% as a central case, retains a customer-expansion upside scenario, limits the first inventory release, and monitors signed order and cohort conversion thresholds. The decision record identifies which evidence moved the range.

Failure modes the workflow must block

Pilot for 30 days in shadow mode

Week 1: select one recurring decision class, such as forecast override or spend variance escalation. Define consequence tiers, authority, required sources, deterministic calculations, specialist routes, and what the AI may draft.

Week 2: create the decision record, source register, atomic disagreement matrix, sensitivity template, and approval manifest. Seed stale data, duplicate populations, wrong signs, unsupported causality, fabricated citations, and policy conflicts.

Week 3: run historic cases without changing prior decisions. Measure whether the process detects known errors, which evidence changes views, false challenges, reviewer time, and whether staff preserve genuine dissent.

Week 4: run live shadow decisions beside the existing process. Do not execute AI-shaped actions automatically. Approve one bounded decision only if source reconciliation, deterministic calculations, specialist review, and version binding work.

At the end of the pilot, examine disagreements that changed no decision as carefully as those that did. Some challenge steps will expose missing evidence yet confirm the original choice; others will create noise without changing risk. Remove ritual checks that add no information, strengthen tests that reveal material assumptions, and preserve a sampling plan so low-frequency failures remain visible.

Final human decision gate

FAQ

Should a CFO follow AI when it disagrees with their judgment?

Not because it disagrees, and not because it sounds confident. Preserve the independent human view, test both positions against approved evidence and reproducible calculations, then let the named owner decide within policy.

Does human review prevent automation bias?

No. Review can become a rubber stamp. Independent first views, visible disagreement, counterfactuals, consequence tiers, and evidence-linked approval create meaningful friction.

Can AI make the final finance decision?

Only where policy explicitly delegates a bounded, tested, low-consequence class. Material planning, reporting, capital, accounting, people, or external actions require the named authority and any specialist gates.

What if the human and AI reach the same answer?

Agreement is not proof. For material decisions, still validate sources, calculations, assumptions, constraints, and authority. Shared errors can produce fast consensus.

What should be retained?

Retain scope, independent views, source versions, calculations, challenge record, edits, dissent, approval, exact released artifact or action, monitoring, and correction route under applicable policy.