← AI Operations collection
AI Operations AI ROI Board Reporting Decision Infrastructure

Your AI ROI Board Report Needs a Baseline Finance Booked Before Launch

A board checks whether finance can find the AI saving in its own books. Seven fields to state, six claims to cut, and the window to measure over.

Christopher Wakare
Updated
13 min read
AI & Operations

The slide says the agent saved the planning team hours. A director who has read a few capital requests asks one question: hours against what? You have a vendor dashboard and a usage count. You have no figure your finance team booked before launch. The room nods, the minute-taker writes "revisit next quarter," and the AI line goes into the budget round with a question mark on it.

The director is not being difficult. The CFO Survey from Duke Fuqua and the Richmond and Atlanta Fed (548 respondents, 11 November to 1 December 2025) reports in its December 2025 wave that "AI investment is not expected to have much effect on the number of employees or produce measurable cost savings in 2026." A working paper from the same institutions, reported by Fortune in March 2026, compared the self-reported 2025 AI productivity gains of nearly 750 executives (1.8% on average, per Fortune) with the gains implied by revenue and employment data, and found the implied gains much smaller. Duke Fuqua's John Graham told Fortune that "ROI often depends on exactly how you calculate it." Caution is rational when the proof on offer is a number nobody outside the team can check. We have read the paper only through Fortune's summary.

Finance-leader surveys read the other way. Grant Thornton's Q3 2026 survey, reported by CPA Practice Advisor, has 84% of nearly 230 US finance leaders saying AI return on investment is meeting or exceeding expectations, and RSM's manufacturing snapshot has 98% satisfied with the business value AI has delivered. Much of that value is productivity, judged through "a broader lens of what value is," which is the capacity line, not a booked saving. A survey answer and a saving in a board paper are different tests.

AI ROI board report, defined

An AI ROI board report is the evidence page a CEO or CFO takes to the board to justify AI spend. It states the baseline finance booked before launch, the interval measured, the deployment type, the named owner, and the cost line that moved, net of AI spend and rework. Decision counts appear as supporting volume, never as the return.

From the work

Written from 5+ manufacturing and distribution engagements, 2025-2026. This is analysis, not a case study: it names no customer and publishes no client figure.

Why a board rejects an AI ROI claim before it reads the number

The AI business cases we review assume a faster return reads as a stronger one. So the deck leads with the 30-day result, the quick win, the first-month dashboard. A board runs a different test. It asks whether the number could have come from somewhere else: seasonality, a price change, a headcount move, a supplier that happened to ship on time. A shorter window gives the number fewer ways to rule those out.

A 30-day result against a baseline rebuilt from invoices after launch loses to a 90-day result against a figure finance booked before go-live. The first has a flattering window and no anchor. The second has an anchor, and a director can check it without calling anyone.

McKinsey's The state of AI in 2026: On the road to ROI (August 2026) surveyed 1,719 respondents between 4 May and 8 June 2026. It counts AI high performers, those with at least 5% of EBIT from AI and significant impact, at about 6% of respondents (92 of 1,719), and reports them about twice as likely as others to report defined processes to measure the impact of AI initiatives. The survey is global and does not list what those processes record, so read it as direction, not as a benchmark for your plant. It does not test the fields below or any other framework. They are our own answer to what a defined process should record.

The cost is a quarter. The next decision on the AI budget waits for the next board cycle, and the spend keeps running while it waits. You can tell you are exposed without opening a dashboard: the result lives only in a vendor portal, or your finance lead first saw the figure the week the deck was built. The fifth root cause in Your AI Pilot Succeeded. Your Operations Didn't., success never defined in operational terms, is the same gap seen from the pilot side.

The AI ROI board report spec: seven fields a CFO can check

Seven fields turn an AI result into something a finance reviewer can accept or reject on its merits. Four of them (baseline, interval, named owner, and the rework inside net position) map to checks in our AI ROI calculator, which is IntelliConnectQ's own scoring rule, not an external standard. The fields also double as the methodology note a CEO needs under any AI figure on a slide: deployment type, measurement interval, baseline source.

The reviewer on the other side of the table often has no standard method either. Protiviti's 2026 Global Finance Trends Survey (press release, 26 August 2026) found that "only 35% of finance organizations say they are effective at measuring AI ROI." The release does not state the sample, roles or geography, so read it as direction. If your CFO has no house method for an AI saving, agreeing these fields with finance before launch gives both of you the same test.

Field What the page states What fails the board
Baseline source The cost line finance booked before launch, the ledger or invoice set it came from, the period, and the date it was recorded. A baseline reconstructed from invoices after launch, or estimated. In the calculator either one caps the claim at Partly defensible.
Measurement interval Long enough for the outcome to show in its own cycle, and at least one full close after go-live, with start and end dates. Adjusted for volume, product mix and raw-material prices. A single month, a pilot window, or any span that skips a close.
Deployment type What the AI does and where: reads and answers, proposes a decision, or writes to the ERP after a named approval. Which system and environment, live or pilot. "AI-enabled" with no scope. A board cannot price a capability it cannot locate.
Named owner One person who approves or overrides each AI decision, and one who signs the saving. "The team" or "operations." A saving nobody signed belongs to nobody.
Decision count and dollar outcome The cost line that moved, in dollars, and the operational KPI that explains it: unplanned downtime, OTIF, first-pass yield, inventory turns, or exception-to-decision time. Decision count sits beside both as supporting volume. Decision count presented as the return, or a dollar figure with no KPI showing how it moved.
Net position The saving after AI spend and after the cost of undoing AI decisions that were overridden or reversed. Selected Business Central agents bill by consumption, and licence and partner fees add to that. Microsoft pools those credits per tenant and says capacity cannot be allocated to specific services, so attribute the use case's share from the credit telemetry the environment sends to Application Insights. Application Insights must be switched on for the environment before the baseline period starts, and the telemetry reports credits, so convert them to dollars at the price on your invoice. Gross saving.
Risk and decision Errors, reversals and control failures in the period, any service impact, and the call the report asks the board to make: scale, fix, keep testing, or stop. A return with no error count, or a report that ends without asking the board for a decision.

The first row carries the most weight. In the calculator a claim grades Defensible only with at least 6 of its 7 checks passed and a baseline recorded before launch. A reconstructed or estimated baseline caps the grade at Partly defensible however many other checks pass. A CFO may weight the fields differently. Copy the fields, not the weights.

The first row assumes the cost sits in its own ledger line. Planner time and expedite freight by cause often sit inside a wider account, and then finance has no booked figure to point at. RSM's 2026 Middle Market AI Survey manufacturing snapshot (129 manufacturing respondents, 21 July 2026) gives "Data quality issues (51%)", "Integration challenges (43%)" and "Unclear ROI (30%)" as reasons for moderate or limited pilot success. When the ledger does not isolate the cost, record a timed or tagged baseline before launch, and label it on the page as timed or tagged, not as a booked figure. The board can then weight it. That beats a baseline rebuilt after launch, which is claim 4 below.

The seven fields hold two scorecards. Deployment type, decision count and the risk line make the AI scorecard: what the system does, how often people use it, how often it is wrong or overridden, and what it costs to run. Baseline, dollar outcome, operational KPI and net position make the business scorecard. The first tells you whether the system works. The second tells the board whether it paid, and only the second decides the case. The two can disagree in either direction. An agent can be accurate, fast and used every day while the plant's costs stay flat. An agent that fires a dozen times a month can carry the case if each alert is an expedite or a stockout that never happened.

Decision count earns its place as volume. A count says the system is used. The dollar line says the use paid. Report the dollars first and put the count beside it. The line under the figure on the slide comes straight from the other six fields:

Methodology note, template

Saving = [cost line] booked in [period] before launch, less [cost line] in [period] after go-live. Site: [site]. Deployment type: [what the AI does, system, live or pilot]. Operational KPI: [KPI] from [baseline value] to [result]. Net of [AI spend] and [rework cost]. Interval: [start date] to [end date], adjusted for [volume, mix, material prices]. Baseline recorded [date] by [finance owner]. Approver of AI decisions: [name]. Saving signed by: [name]. Errors and reversals in period: [count]. Decision requested: [scale, fix, keep testing, stop].

Every bracket is something a director can ask about. A bracket you cannot fill means the slide is not ready.

Six AI return claims a board should reject

Each claim below fails a finance reviewer on sight. Each fix is a field from the table.

  1. Usage as the return. Queries answered, agent actions taken, active users. Each counts activity. A board cannot price a query. Replace it with the cost line that moved.
  2. Hours saved with no baseline. The vendor's tool estimates minutes per task. Finance books overtime, contractor spend and headcount. If none of those moved, report the hours as capacity released, on their own line, never as a saving.
  3. Vendor-reported figures. A vendor dashboard sets the metric, the denominator and the period. Ask for the figure from your own ledger. If the two differ, the ledger wins and the vendor figure moves to the appendix.
  4. A baseline rebuilt after launch. The person rebuilding it already knows the result. Use the figure finance booked before go-live, or say the baseline was reconstructed and let the board discount the claim.
  5. A window shorter than one business cycle. A single month holds one close. Seasonality, a price change or a late supplier can produce the same movement, and a 30-day result cannot rule them out.
  6. Gross savings. Credits consumed, partner fees and the time spent undoing overridden AI decisions all offset the saving. On Business Central, selected agents bill by consumption and the credits pool per tenant, so the invoice does not show one use case's share. That attribution, and the billing itself, is covered in Business Central Copilot Credits billing: a default limit is not a budget. The calculator nets that spend against the saving.

A claim that survives all six has already answered most of what a director will ask. The slide that follows can be short.

Cash, capacity and upside: report AI value on three lines

Claim 2 does not make released hours worthless. Suppose a planner spends four hours each morning reconciling orders, stock, supplier dates and the production schedule, and the AI cuts that to 45 minutes. Salary, headcount and overtime stay flat, so the P&L shows nothing. The planner now catches shortages earlier and handles more orders. That value is real, and it belongs on a different line from a booked saving.

Finance already sorts it this way. In the CFO Survey, CFOs expect AI "to increase worker productivity, enhance decision speed and accuracy" while not expecting much effect on headcount or measurable cost savings from it in 2026. On the plant side, Josh Renicker, Director of Ambassador Programs at the COO Forum and a seated COO since 2013, said on an IndustryWeek podcast in September 2026: "It's all about efficiencies. I haven't met folks that are trying to reduce headcount." That is one operator's account, not a survey. It fits the capacity line: hours freed, no headcount cut, no booked saving.

Gartner's finance practice draws the same line. Its 4 November 2026 release for CFOs says productivity gains in most organizations "are too marginal to materially impact financial performance or consistently lower headcount," and analyst Grace Gstalder calls AI-generated capacity "a resource, not a return." Without active governance, she says, the time savings "will remain dispersed and unmonetized." The survey behind it covered 1,303 senior functional leaders from January to April 2026. Of those, 23% reported a headcount reduction from AI and 17% an increase. Gartner's remedy is an owner and a destination: name the outcomes the freed hours should serve, and hold teams to showing where the time went. The release gives no revenue band for the sample. Hours with no named destination stay on the capacity line.

Value line Manufacturing example Evidence the board sees In the net position?
Cash saved Overtime, contractor spend, scrap or expedite freight that fell Ledger lines before and after go-live Yes
Capacity released Planning, reconciliation or reporting hours freed Hours measured against a timed baseline, plus who redeploys them and to what No, until the hours turn into cash or output
Economic upside Fewer line stoppages, better OTIF, extra saleable output where demand exists Incremental contribution margin, avoided losses, or working capital released Reported separately, with the method stated

Two rules keep the lines apart. Never add them into one total, and never label capacity or upside as realized saving. Count each benefit once: if less downtime produced more saleable output, that output appears as upside and does not come back as a productivity saving.

AI ROI measurement window: report on a business cycle, not a countdown

A 30-day AI return can be true. A board that knows what month-end does to the numbers cannot confirm it in 30 days. Choose the window from the outcome you are measuring. A maintenance fix can show fewer repeat failures within weeks. A procurement change needs several purchase-order and delivery cycles. An inventory change needs months to show that service held while stock came down, and a seasonal line needs a year-on-year comparison. Whatever the KPI needs, the dollar figure spans at least one full close, with start and end dates on the page.

A booked baseline still misleads if the business changed underneath it. Adjust for volume, product mix and raw-material prices. Where you can, compare against a line, site or product family the AI did not touch. Put the adjustments in the methodology note so a director can see them.

Set the window before launch, together with the review date and the stop threshold. Write the AI Pilot Stop Criteria Before the Pilot Starts covers who signs that decision at the review date: a sponsor and a finance controller together. The board report is the document that review produces. If you built the case with the four-variable framework in AI agent ROI in manufacturing, the board report is that case settled against booked figures.

A vendor guarantee window sits beside the reporting period and does not replace it. The guarantee settles who carries the risk. The reporting period settles whether the number holds.

Three questions finance asks before an AI saving counts

The seven fields answer the first question. Two more come up in the same meeting. Each has a free tool you can run before it does. This is IntelliConnectQ's own reading of what reviewers ask, not an external standard.

  1. What did it save, net of spend and rework? The seven-field spec above answers it. Run the claim through the seven checks in the AI ROI calculator.
  2. Who approved the AI decisions, and could they read what they approved? A named approver on the record proves little when the queue is larger than anyone reads. Set the number one approver reads properly in a day, then compare it with your volume in the Approval Budget Calculator. The detection method is in AI approval fatigue: if every agent action needs a yes, nobody reads the approval.
  3. What happens when it is wrong? The risk and decision field reports it after the fact. Owner, stop threshold and undo path get written before launch. The AI Agent Post-Mortem Template holds all three on one page, and the stop threshold is covered in the AI pilot stop criteria article.

Frequently asked questions

What should an AI ROI board report include?

Seven fields: the baseline source (a cost line finance booked before launch), the measurement interval (long enough for the outcome to show, and at least one full close), the deployment type, a named owner for the AI decisions and for the saving, the dollar outcome with the operational KPI that explains it and decision count as supporting volume, the net position after AI spend and rework, and a risk and decision line (errors, reversals and control failures, plus whether to scale, fix, keep testing or stop). Four of them map to checks in IntelliConnectQ's AI ROI calculator, which is our own scoring rule and not an external standard.

How long should you measure AI ROI before reporting to the board?

Long enough for the outcome to show in that process's own cycle, and never less than one full close after go-live. A maintenance fix can show fewer repeat failures within weeks, procurement needs several purchase-order and delivery cycles, and a seasonal line needs a year-on-year comparison. State the start and end dates, and adjust for volume, product mix and raw-material prices. No external standard sets a number of days. A vendor guarantee window is a separate commitment about who carries the risk.

Is the number of AI decisions a valid ROI metric?

No. A decision count shows the system is used. It does not show the use paid. Report the cost line that moved, in dollars, first, and put the decision count beside it as supporting volume. Take the count from a decision log that records each decision and its approver.

What baseline will a CFO accept for AI savings?

A figure finance booked before launch, with the ledger or invoice set it came from, the period, and the date it was recorded. A baseline reconstructed from invoices after launch, or estimated, weakens the claim: in the calculator it caps the grade at Partly defensible. If you have not launched yet, record the baseline now. If you launched without one, say so on the page and grade the claim down yourself before a director does.

What to get in writing before you sign an AI ROI guarantee

A guarantee moves risk. It creates no evidence. OpsGrid, IntelliConnectQ's decision infrastructure for Dynamics 365 Business Central, is in active beta with a 3-week implementation, and the offer reads "Guaranteed ROI on one agreed use case within 90 days of production, or 50% refund." The OpsGrid page states the terms as "Measurable ROI within 90 days of moving to production on one use case we agree with you after the discovery phase, or we refund 50% of the implementation cost." Business Central is required. The guarantee covers one use case, agreed with you in the discovery phase before kickoff, not every workflow OpsGrid supports. Most companies don't carry every problem on the decision list on the OpsGrid page, so the use case is picked from the decisions listed there that you actually have. It is worth evaluating if you run Business Central and can name a slow operational decision that costs you money, such as an expedite approved after the truck should have left. If you cannot name one yet, start with the Decision Latency Diagnostic.

A refund clause answers what happens if the number misses. It leaves open which number, measured against what, and signed by whom. Get answers in writing before kickoff, from any vendor, OpsGrid included:

  1. The measurement. The baseline source, the interval, and who signs the result. Make it your finance controller, working from source data you both agreed.
  2. The threshold. Whether any measurable gain counts, or the result must cover implementation and running costs.
  3. The costs inside the calculation. Software, integration, your own staff time, training, AI usage, support and rework.
  4. The counterfactual. How the result separates the product's effect from changes in demand, prices, staffing or supplier performance.
  5. The remedy. What qualifies for the refund, what is excluded, and the process to claim it.
  6. The demo. Whether you see it run on one of your own workflows, with realistic data from your environment, before you sign. Sample data shows the interface, not the fit.
  7. Acceptance criteria. The error, override and service-impact limits the pilot must stay inside, written next to the financial threshold.
  8. Day 91. Whether anyone still measures the benefit after the window, and what it costs to keep running.

OpsGrid is in beta and has no standard ROI definition yet, so we set the baseline, interval and signer for each engagement in the discovery phase, against the one use case the guarantee covers. Ask us the rest of the list too. With the answers, the guarantee and your board report describe the same number. Without them, the vendor and your finance lead each measure something different and each can be right.

OpsGrid logs every decision, override and outcome to an audit trail, which gives the decision-volume line of your report a source. The dollar line still comes from your ledger, and the baseline still comes from finance. Run your own claim through the seven checks in the AI ROI calculator first, then apply for OpsGrid beta access with the baseline already in hand.

Would finance accept your AI saving in a board paper?

The AI ROI calculator nets your saving against Copilot Credit spend and rework, then grades the claim on seven checks, starting with whether the baseline was recorded before launch. Free, indicative, nothing leaves your browser. OpsGrid carries Guaranteed ROI on one agreed use case within 90 days of production, or 50% refund.

Open the calculator →

The Execution Edge

Monthly. For operations leaders building faster on AI. Real case studies, system blueprints, and tools, no fluff.

Your subscription could not be saved. Please try again.
Your subscription has been successful.