Back to Guides
Enterprise Software 15 min read

Rethinking the quarterly AI report to the board

Rethinking the quarterly AI report to the board

Most quarterly AI pages are written to survive the meeting rather than to decide anything, and the reason is structural. Grant Thornton’s 2026 AI Impact Survey, fielded with roughly 1,000 US senior leaders in early 2026, found that three in four boards have approved major AI investments while only 52% have set clear AI governance expectations. The money went out. Nobody said what had to come back.

Whoever writes the AI enablement reporting section inherits that vacuum. With no standard to hit, the page fills with what is easy to pull: seats bought, seats active, a session count, a quote from someone who liked the training. It reads fine and decides nothing, which is why the same page gets written again next quarter with different numbers.

A page that earns its slot does four things: states a dated capability baseline, counts named work that changed, puts one output claim on the record with its confound attached, and asks for a specific decision. Everything else belongs in the operating review.

What the current page usually contains

Three failures show up across almost every version of this page, and the rest of this piece replaces all three.

It reports purchasing as progress. Licenses provisioned and percentage activated describe a procurement outcome. A director reading “94% of licenses activated” learns that IT completed a rollout, which was never in doubt. Separating access from capability is the whole subject of our case against treating licenses as capability, and it starts mattering the moment a number goes in a board pack.

It reports a rate the group size cannot support. A company of 50 to 500 people trains 20 to 60 at a time, and a quarter-over-quarter move on a percentage computed across 40 people is mostly sampling error. It gets read as a trend anyway. Counts survive that scrutiny; rates at that size do not.

It reports a gain with no mechanism. “Roughly 20% faster turnaround” with no named deliverable, no measurement window and no competing explanation is an assertion. The first director who asks what else changed last quarter ends the discussion, and the credibility loss carries into the next ask.

Four layers of measurement, one of them for directors

Most of what gets collected during an enablement program is real, useful, and somebody else’s business. Sorting it by audience settles most arguments about what to cut.

Layer Question it answers Typical measures Who reads it
Access Do people have the tools? Seats provisioned, share entitled, tool inventory IT and procurement
Activity Are they opening them? Weekly active use, actions per person, use by team The program owner
Absorption Has the work changed? Named workflows in production, retention at 90 days, high-use against low-use teams Function leads
Value and assurance Did it pay, and can we defend it? One output change, cost to date, incident log, policy coverage, owner Executive team and board

Only the bottom row is board material. The three above it are operating instruments, and they are what the monthly read exists for, which is the split behind our comparison of monthly and quarterly reporting.

That also explains the length of a good board page: once the top three layers are stripped out, one page is the honest size. Pages run long because the writer is nervous, not because directors need the detail.

The four things that belong on the board page

A dated capability baseline, with the re-measure date named in advance

Lead with where capability stood before the program and where it stands now, measured the same way on two dates. A short scored exercise against the team’s own work produces that, and the method matters less than the fact that it was the same method twice. Our fluency assessment framework sets out one version a non-technical team can run itself.

State the next measurement date in the report rather than after the fact. A baseline with a scheduled re-measure is a commitment; the same baseline quoted later, once results are in, is a selected statistic and reads like one.

A count of named workflows, not a share of people

The strongest number available to a mid-market program is a count of specific recurring work now done differently, each with a named owner. Six workflows is six workflows, and a director can ask about any one of them.

Pick work the business already recognizes: the weekly client status update, the first draft of an RFP response, the monthly variance commentary, candidate screening notes. Three of those at one line each outperform any adoption percentage on the same page, because each one is checkable.

One output change, with the confound named in the same sentence

Choose a single deliverable with a cycle time or quality measure that existed before the program, and report the change against the comparison period. One claim, fully sourced, beats five claims that each invite a question.

Then name what else moved. A quarter that brought new licenses, a new intake process, a hiring freeze and a seasonal volume shift offers four candidate explanations for one number.

Writing “two of these three weeks of improvement coincide with the new intake form, so we are treating the attribution as partial” does not weaken the report. It tells the board that the person who wrote it would notice if the program were not working, which is the only basis on which the next number gets believed.

A decision ask with a figure attached

Close with one of four asks: continue at current spend, expand to a named group at a named cost, hold while a specific question gets answered, or stop. Directors are comfortable with “hold and re-measure in March” when the question is stated. They are not comfortable with a page ending in a summary.

The sanity check that keeps a claim credible

Run this arithmetic before any time-savings or AI training ROI figure leaves the building. It is the check a numerate director runs silently.

Take a 200-person company that trained 40 people and heard back that each saves about four hours a week. That is 160 hours a week, four full-time equivalents, around 2% of total labor capacity. Set it beside the outside number: across nearly 9,000 executives in seven countries surveyed in 2026, firms reported AI had raised their annual labor productivity growth by 0.2%, with 0.6% expected over the next three years (Yotzov, Barrero, Bloom and co-authors, NBER working paper 34836, revised October 2026).

A claim ten times the aggregate figure is not a win on the page. It is an invitation to doubt the whole document, and it is usually true at the individual level and false at the company level, because the hours were real and never converted into output. Four hours returned to a person with no queue behind them produces no measurable change in what the company shipped.

So report hours redeployed to named work rather than hours saved. Two people absorbing work that would have gone to a contractor is a figure with a cost attached. Forty people feeling faster is not.

Reporting the rework rate beside any gain protects the same credibility. Output that needed human correction before it could be used inflates every gross productivity figure, and stating it unprompted is cheaper than having it found.

Why the governance half now gets read first

Two shifts have moved board attention from “did it pay” to “who is holding this.”

The SEC’s Investor Advisory Committee voted on December 4, 2025 to recommend that the Commission issue guidance on AI disclosure, including whether the board or a board committee is responsible for overseeing aspects of AI deployment. Those recommendations are not a rule, and the figures cited in support show why the direction matters anyway: 15% of the S&P 500 disclose board oversight of AI, while 60% treat AI as a material risk. Private mid-market companies carry no such obligation, but their directors read the same coverage and the question arrives regardless.

The second shift is who holds the file. McKinsey’s state-of-AI research found 17% of respondents at AI-using organizations say their board oversees AI governance and 28% say their CEO does. At this company size that ownership usually sits with whoever got handed the program as a second job, and the board page is where it becomes visible or stays ambiguous.

Three governance lines are enough: the tool and agent inventory with a named owner for each, policy coverage against actual use including anything running outside the approved set, and incidents with their resolution.

State the cadence as a decision rather than a habit. The NIST AI Risk Management Framework asks under GOVERN 1.5 that ongoing monitoring and periodic review be planned with roles defined, including determining the frequency of that review. One line recording that the board reviews this quarterly and the owner reviews it monthly is a control.

Write all of it for an audience that cannot evaluate a technical metric. Diligent Institute and Corporate Board Member surveyed more than 200 US public-company directors for What Directors Think 2026 and found 8% say their board has strong AI expertise. Any metric needing a paragraph of definition is the wrong metric for this page.

Writing the quarter where nothing has landed yet

The first report after a workshop almost always lands before an output change is measurable, and the temptation is to fill the gap with enthusiasm scores. Session satisfaction belongs nowhere near a board page: it measures the room, not the work.

Send the baseline instead, and make the honesty structural:

  1. The dated baseline, with the method named in one line.
  2. The workflows the team committed to, each with an owner and a target date. Commitments are legitimate board content when labeled as commitments.
  3. The date capability gets re-measured, chosen before any results exist.
  4. Cost to date and remaining spend, so the next ask is not a surprise.
  5. The governance lines, which do not depend on results.

That page puts the program on a clock the writer set in public. What happens in the following weeks decides whether the next report has anything in it, the sequence our 30/60/90 behavior-change framework is built around.

If the second report also shows no output change and no workflows in production, that is the finding. Report it as one. A program with a baseline, no movement and a stated decision point is being managed; the same program reported as “strong engagement” for three quarters is how AI enablement budgets quietly die.

A one-page skeleton

Fit this on one page, in this order:

  • Where we stood, where we stand. Two dated capability figures, same method, one line on the method.
  • What changed in the work. Three to six named workflows, one line each, owner named.
  • The one number. A single output change against a stated comparison period, with the competing explanation in the same sentence.
  • Cost. Spent against budget, including internal time.
  • Governance. Inventory and owners, policy coverage, incidents and resolutions, review cadence.
  • The ask. Continue, expand to a named group at a named cost, hold pending a stated question, or stop.

Anything that does not fit those six lines goes in an appendix nobody is asked to read. The arguments a finance reader will press on are in our measurement guide.

Frequently asked questions

What should a quarterly AI report to the board actually contain?

Six blocks on one page: two dated capability measurements taken the same way, three to six named workflows that changed with owners, one output change with its competing explanation stated, cost to date against budget, three governance lines, and one decision ask. The test for each line is whether a director could act on it or challenge it; if neither, it is filler.

How long should the board section on AI be?

One page, or one slide plus an appendix. Once activity and usage metrics are stripped out, one page is the honest size. Long AI sections usually signal padding around a missing result, and experienced directors read length that way.

Is adoption percentage a bad metric for a board?

It is the wrong metric for this audience and this group size. At 40 trained people, a quarter-over-quarter move of 10 to 20 points sits inside normal sampling variation, so the number invites a conclusion the data cannot support. Report counts to the board, keep rates for the operating review, and never show directors a rate without its margin.

What do we report in the first quarter when there are no results?

The baseline, the committed workflows with owners and dates, the date capability gets re-measured, cost to date, and the governance lines. Label commitments as commitments. A board page that says “here is the clock we set ourselves” is stronger than one padded with satisfaction scores, and it makes the second report easier to write honestly.

How do we handle attribution when several things changed at once?

Name the competing explanations in the same sentence as the claim, and say how much of the change you are attributing. A new process, a hiring freeze, seasonal volume and the training all landing in one quarter means the attribution is partial, and saying so is what makes the figure usable. A clean causal claim you cannot defend costs more credibility than a hedged one.

Should AI governance and AI value be in the same report?

Yes, on the same page. Directors are increasingly asked who oversees AI deployment, and the SEC’s Investor Advisory Committee recommended in December 2025 that issuers disclose exactly that. Splitting value and governance into two documents means one of them does not get read, and the governance half is now the half that gets asked about first.

Who should own the quarterly AI report?

The person accountable for the outcome, not the analyst who can pull the data. In most companies of 50 to 500 people that is a COO, VP of operations or head of learning who took the program on as a second job. Name that owner in the report, because ambiguous ownership is itself a governance finding and the board will eventually treat it as one.

What does a board do with this report?

Approve continued spend, approve expansion, hold pending a specific answer, or stop. That is the full range, which is why the page ends in one of those four asks with a figure attached. Reports written to inform rather than decide get skimmed and reappear next quarter.

How do we avoid overstating time savings?

Translate any AI training ROI claim into a company-level figure first. Forty people saving four hours a week is around 2% of a 200-person company’s labor capacity, against an aggregate 0.2% annual productivity gain that executives across nearly 9,000 firms reported in 2026. If your number is an order of magnitude above the outside figure, report hours redeployed to named work instead, and state the rework rate next to any gain.

Key takeaways

  • Boards approved the spending without setting an expectation for what comes back. Write the quarterly page as a decision document, not a status update.
  • Three of the four measurement layers belong to someone other than the board. Removing them makes one page the correct length.
  • Counts survive scrutiny at mid-market group sizes; rates do not. Report named workflows, not adoption percentages.
  • Sanity-check every time-savings claim at company level. A figure an order of magnitude above the published aggregate damages the whole report.
  • Name the competing explanation in the same sentence as the result. Partial attribution stated plainly beats a clean claim you cannot defend.
  • Keep value and governance on one page. Inventory, policy coverage, incidents and review cadence are available from day one.

Building a page that survives a finance review is mostly a measurement problem, and it is easier to solve before a program starts than after. If you are drafting that section now, book a free AI-readiness call and we will go through the baseline you can still capture and the two numbers worth committing to in writing.

Last Updated: Nov 14, 2026

AW

Arthur Wandzel

SFAI Labs helps companies build AI-powered products that work. We focus on practical solutions, not hype.

Make your team fluent in AI — then automate what proves out

  • Hands-on training applied to your own documents and workflows
  • Often reimbursable through state small-business training programs
  • Built for 5–50 person firms with no IT department

Related articles