Reporting AI spend to the board.

Operations guide · July 19, 2026

By the LLM CFO team

Two-thirds of boards now condition further AI funding on proof of return. Yet most finance leaders lack the data to demonstrate it. The difference between a board that continues to fund AI and one that pauses is not the amount of spend, but the reporting: how well you connect the dollars to the business outcome, how honestly you account for gaps, and how credibly you forecast what comes next.

Board reporting is not an internal dashboard

An internal LLM cost dashboard is operational and real-time: token counts, cost per request, spend trends minute by minute. It answers "what is happening right now?" A board report is quarterly and business-focused. It answers a single question: "should we keep funding this?"

This distinction matters because it shapes what goes on each page. An internal dashboard shows token efficiency gains; a board report shows whether you're getting more revenue or lower cost per unit of work. An internal dashboard highlights cache-hit rates; a board report shows whether the AI initiative is cheaper than the people it replaced, or whether it's additive spend on top of existing headcount.

The reporting pack: five sections that matter

Section 1: Spend vs. budget with variance explained

Lead with a single bar chart: budgeted vs. actual for the quarter. Include year-to-date. Spend five sentences max on the variance. No surprises. If spend exceeded budget, name why in the first sentence and quantify the impact.

Example: "Q3 AI spend was $1.8M vs. $1.5M budget, +20% variance. Root cause: search product adoption exceeded forecast (150% of plan), driving LLM requests to 4.2M vs. plan of 2.8M. Mitigation: completed model optimization work in August and rerouted 35% of search queries to cached embeddings. Q4 budget revised to $1.7M with 85% confidence interval."

The board does not need to see the detailed line items. They need to know whether the CFO and product leaders can still predict spend. If you cannot predict it to within 15%, say so and explain why.

Section 2: What we got for it — outcomes measured, honestly caveated

This is the business outcome section. Name the outcome, measure it in revenue or cost terms, show it moved, and caveat any doubt.

The most common mistake is reporting adoption metrics as if they were outcome metrics. "We deployed AI copilots to 4,200 salespeople" is not an outcome. "Salespeople using the copilot close deals 11% faster, adding $2.3M annual pipeline" is an outcome. The first tells you something is in use. The second tells you it matters.

Example outcomes by function:

• Sales: deal cycle time reduction (days to close), win rate lift, average deal size change. Measure vs. a control group of non-users if possible.
• Support: cost per ticket (total support cost / volume), first-response resolution rate, customer satisfaction (CSAT or NPS change). Show the math: if AI handles 20% of volume at 80% the cost of a human, the total support cost per ticket went down by X%.
• Engineering: time to code review, incident triage time, developer satisfaction. Harder to link to revenue, so be specific about cost (e.g., "reduced on-call burden by 12 hours per week per engineer = X engineering hours freed for feature work").
• Finance: close time, audit prep time, forecast accuracy. Measure in days or basis points. Example: "monthly close moved from 8 days to 6.5 days, freeing FP&A team for strategic analysis."

Caveat every number. If the outcome comes from a self-reported survey, say so. If the control group is small, name the confidence interval. If the outcome is modeled rather than measured from actual data, lead with "based on observed metrics and extrapolation, we estimate..." and show your math.

Section 3: Unit economics trend — are we moving in the right direction?

Unit economics show whether the AI is getting more efficient or less. Pick the most relevant metric for your business and show three quarters of trend.

Examples:

• Cost per customer: total quarterly AI spend / active customers. Should trend down or stay flat as you scale.
• Cost per transaction: total AI spend / transactions completed with AI assistance. Should trend down as models improve and efficiency work compounds.
• Cost per business outcome: AI spend / outcome units (revenue generated, support tickets resolved, time saved in hours). Should improve quarter-over-quarter as you optimize prompt, routing, or model selection.
• Revenue per dollar of AI spend: attributed revenue from AI-assisted deals or transactions / total AI spend. Should trend up if the AI is truly driving incremental business.

Show the trend as a line chart, three to four quarters. A flat or improving trend is credibility. A degrading trend is a red flag that either spend is growing faster than outcomes or you're adding unoptimized new workloads. If the trend is degrading, own it and show the recovery plan.

Section 4: Concentration and vendor risk

Name your AI providers and model dependencies. Show what percentage of AI spend goes to each. Call out any single-vendor concentration above 60%.

In 2026, vendor risk is a board-level conversation. Model access can change on a vendor's timetable, not yours — a deprecation, a pricing change, a capacity limit, or an export-control or licensing shift can all land with little notice. A company running 80% of its AI workloads through one provider is one such event away from a material operational problem. A board needs to know whether a policy change or vendor exit would pause your AI roadmap, and how long a switch would take.

What to include:

• Spend mix by provider (pie chart, or table).
• Critical workloads and their provider dependencies (e.g., "customer support copilot runs on OpenAI; no fallback planned").
• Diversification status: "60% OpenAI, 25% Anthropic, 15% in-house." Or if concentrated: "85% OpenAI, diversification roadmap in Q4 2026."
• Key contract terms: Does your contract have a data retention clause? IP indemnification? Exit terms? (These are legal department questions, but summarize for the board.)

This section is brief. The board is asking one question: if this vendor's model becomes unavailable tomorrow, how long can we operate? If the answer is "two weeks," say so and show what you're doing about it.

Section 5: Forward run-rate and the ask

Close with a one-slide projection of AI spend for the next four quarters. Show three scenarios: conservative (current roadmap, no new initiatives), base case (planned launches and model optimizations), and upside (full roadmap including speculative projects). Include a line for "required budget."

Example:

• Q4 2026: Conservative $1.6M, Base $1.8M, Upside $2.1M.
• Q1 2027: Conservative $1.7M, Base $2.2M, Upside $2.8M (includes new recommendation engine).
• Q2 2027: Conservative $1.8M, Base $2.0M, Upside $3.1M (upside assumes engineering AI assistant launched).

Then state the ask clearly: "We are requesting $8.2M budget for H1 2027 (base case). Conservative case allows us to maintain current scope at $3.3M. Upside requires $5.9M and delivers an estimated $4.8M return. We recommend the base case."

The metrics directors actually ask about

They ask:

• "Is it cheaper than the alternative?" Can you quantify the cost per outcome (e.g., cost per support ticket) vs. the cost if a person did it? If not, you don't yet have a business case.
• "What happens if we stop?" What operational impact would killing the AI initiative have? Revenue lost? Headcount needed to replace it? Time-to-market delay?
• "Are we locked in?" Can we exit the vendor or model and still operate? Do we have contractual lock-in or architectural dependency lock-in?
• "What are peers spending?" Is $1.8M for search AI a lot or a little? Directors often compare to peer spend as a sanity check. Be ready with rough peer benchmarks (you don't need exact numbers, order of magnitude is enough).
• "When does this pay back?" The board wants a timeline. "This AI initiative required $500k in Q1-Q3. Based on current outcome trajectory, payback occurs in Q2 2027. If outcomes improve 15% via model optimization (underway), payback occurs in Q1 2027."

They don't ask (resist reporting these as wins):

• Adoption metrics. "We deployed AI to 2,000 salespeople" is input, not outcome. The board assumes you deploy capabilities; they measure whether the deployment worked.
• Pilots and time-to-value. "We launched an AI pilot in support" does not tell a board whether the pilot succeeded. Report outcomes, not activity.
• Technology metrics. Token counts, model inference latency, cache-hit rates are operational KPIs. They matter to the engineering team and internal dashboards. They do not belong in a board report unless they directly explain a cost variance or outcome miss.
• Vague productivity gains. "Engineering teams report time savings" without quantification is theater. If you claim time savings, convert to hours freed and show what the team did with those hours. If they did nothing extra, the time savings are phantom.

How to handle the awkward quarters

Awkward quarters happen. Spend grew but outcomes lagged. A model upgrade cost more than expected. A rollout underperformed forecast. Credibility compounds. If you overclaim once, the board discounts your next three projections by 30%. Here's how to handle it.

Own the gap immediately

Do not hide it in footnotes or hope it goes unnoticed. Lead with it. "Q3 AI spend exceeded target by 20% while return lagged forecast. Root cause: [specific reason]. Mitigation: [actions taken]. Updated forecast: [revised outlook]."

Explain the root cause in one sentence

Not ten slides. One sentence. "We selected model A expecting 12 requests per second; it delivered 8, forcing us to run two instances instead of one."

Show the mitigation or recovery plan with dates

"We switched to model B in August. September data shows throughput at 14 requests/sec and cost per outcome improved 18%. Q4 forecast updated accordingly."

Update the forward forecast

Do not pretend the gap will magically close. Show the revised numbers. "Q3 underperformance cost $200k. Q4 optimizations recover $140k of that. We now expect full payback by Q1 2027 instead of Q4."

Reduce confidence if warranted

If your forecast was wrong once, the board is now watching more carefully. You can say: "Q4 forecast: $1.8M, 70% confidence interval (was 90% in Q3 due to execution risk on model change)." Transparency builds credibility even if the confidence window widens.

Pre-empt the questions: what boards will ask before you answer

Q: "What happens if we stop?"
Come with a 30-second answer ready. Not a strategic essay. Example: "Stopping the support copilot means support cost per ticket returns to 2024 baseline ($85 vs. current $71). Annual support cost impact: +$840k. Pipeline impact: none currently; support AI is internal efficiency. Strategic impact: we lose a feature in competitive RFPs."

Q: "Are we locked in?"
Answer the contract question and the architecture question. "OpenAI contract: 12-month term, exits with 30 days' notice as of March 2027 (no exit penalty). Architectural lock-in: medium. Search embeddings are specific to OpenAI's API; migration would cost 4 weeks of engineering time but is possible. Customer support copilot is model-agnostic; we can switch models in 2 days."

Q: "What are peers spending?"
This is hard because vendor data is private. But you can give order-of-magnitude ranges. "Based on public commentary from [peer A] and [peer B], AI spend for similar-sized companies in our space is 0.5% to 1.5% of revenue. We are at 0.8%, so in the middle of the range. We expect to move to 1.2% by Q2 2027 as new initiatives scale."

Q: "When does this pay back?"
Give a specific quarter and show your math, even if rough. "Sales copilot payback: Q1 2027. We spent $650k in setup and Q1-Q3 run costs. Expected incremental revenue from copilot: $1.8M annually based on 11% deal cycle acceleration and 80 active salespeople. Payback calculation: 650 / (1,800 / 4 quarters) = 1.4 quarters, rounding up to Q1 2027."

One-page board summary template

Use this template to assemble your quarterly board report in five sections. Adapt the metrics to your business.

Related

• LLM cost dashboards: the five views that matter
• AI governance for finance leaders: policy, allocation, accountability
• LLM cost monitoring: what to track and how to control it
• AI FinOps: the operating model for LLM spend