Most AI project sponsors walk into board reviews with a 30-slide deck and walk out with the project deferred. The deck loses because it is calibrated to the wrong artifact. Boards in 2026 do not approve AI projects from slide decks; they approve AI projects from one-page investment theses that pre-answer the capital-allocation questions the board would otherwise spend the meeting asking. The 30-slide deck is the work the sponsor did; the one-page thesis is the artifact the board needs. Sponsors who do not produce the thesis perform the capital-allocation conversation in real time during the meeting, which is where projects die; not because the thesis is wrong but because there is no thesis to vote on. This piece specifies the one-page AI project investment thesis: six sections, two to three sentences each, what to include, what to leave out, with a worked example for a real-shape AI project.
It is a spoke under the AI project economics manifesto, which argues AI economics has shifted from feature cost to evaluation cost. The investment thesis is the board-facing expression of that shift; eval-defined success criteria are what convert a budget request into a defensible commitment.
Why the one-page thesis exists
Boards approve capital allocation. Capital allocation requires answers to a small number of questions: what is the problem, what does success look like, how much will it cost, when do we stop, what is the return, and what is the alternative. Those six questions are independent of the domain; they are the same questions a board asks about an acquisition, a product launch, or a factory expansion.
A 30-slide deck reorganizes those six questions into a narrative that is good for storytelling and bad for capital allocation. The board has to extract the answers from the slides during the meeting, which means the meeting is consumed by extraction rather than decision. By the time the board has the answers, the time is up and the decision is “let’s revisit next quarter.”
The one-page thesis inverts the artifact. The six answers are pre-extracted, ranked by capital-allocation salience, and presented at a length the board can read in three minutes. The meeting is then consumed by the decision rather than the extraction. Approval rates on AI projects with a one-page thesis are materially higher than on projects with only a slide deck; not because the underlying work is better, but because the artifact is calibrated to the decision.
The thesis is also a forcing function on the sponsor. Each section is short enough that fluff is impossible; most line has to be defensible. A sponsor who cannot write the thesis crisply does not have the answers, which is itself useful information. Boards that demand the thesis before approving budget filter out projects that are not yet ready, and that filter is what produces the higher hit rate downstream.
The thesis is not a substitute for the slide deck. It is the artifact that the slide deck supports. The deck explains the thesis; the thesis is what the board approves.
Section 1: Problem statement
What to include. Two to three sentences naming the operational or strategic problem the AI project solves, the team or function that owns the problem, and the cost of the problem in measurable terms. The cost is dollars, hours, customer-impact, or competitive position; not “we should do AI.”
What to leave out. Technology choice, model name, architectural approach, project plan. Those are slide-deck content. The thesis names the problem, not the solution.
Length test. A reader who has rarely heard of the project should know after the problem statement what the project is trying to fix and whether they care. A statement that does not pass this test is not a problem statement; it is a project description.
Common failure. “We will deploy AI to improve customer support.” Wrong. That is a solution statement with no problem behind it. The right form: “Customer support handles 12,000 tickets per month with a 26-hour median time-to-resolution; the team is 18% understaffed against the inbound volume; satisfaction scores are 8 points below target on tickets requiring research. The cost is roughly 2.4 million dollars annually in churn risk and team burnout.” Now there is a problem.
Section 2: Eval-defined success criteria
What to include. Two to three sentences naming the specific eval bar that defines success, the threshold the eval bar must clear, and the cadence at which the eval bar is run. The bar is concrete; a measurable property of model output on a defined test set; not “the model works well.”
What to leave out. Vague quality language. “Better than baseline” is not a bar; “exceeds 78% accuracy on the held-out customer-support eval set, run weekly” is a bar. Eval discipline is what distinguishes the AI project from a research demo. The structural argument for eval-defined success is detailed in stop budgeting AI projects in story points, budget them in eval runs.
Length test. The eval bar should be specific enough that an outsider could reproduce it given the test set. If the bar can be paraphrased multiple ways, it is too vague.
Common failure. “We will achieve high accuracy and low latency.” Wrong. The right form: “The model must achieve 78% accuracy on the 2,400-item customer-support eval set, with p95 latency under 1.8 seconds, sustained for two consecutive weeks before promotion to production. The eval set is owned by the customer-support quality team and refreshed quarterly.”
Section 3: Budget cap with TCO bracket
What to include. Two to three sentences naming the 24-month TCO bracket, the major cost lines that compose it, and the volatility reserve assumption. The bracket is honest; a low and a high; not a single point estimate.
What to leave out. Single-point precision that the sponsor does not have. Boards prefer a defensible 1.4M to 1.8M bracket over a fake-precise 1.62M number. The decomposition behind the bracket is detailed in decoding AI project TCO.
Length test. The bracket should reflect the sponsor’s actual uncertainty about model price moves and usage scaling over the 24-month window. A tight bracket signals the sponsor has not modeled volatility; a wide bracket signals the sponsor has not done the work to narrow it.
Common failure. “$1.5M annual budget.” Wrong on three counts: not bracketed, not decomposed, not 24-month. The right form: “Budget cap of 1.4M to 1.8M over 24 months. Decomposition: inference 850K to 1.0M with 15% volatility reserve, eval suite 180K capex amortized over 36 months, model deprecation reserve 120K, on-call and platform tooling 170K, engineering 200K. Bracket reflects model-price uncertainty over the 24-month window.”
Section 4: Kill clause with numeric triggers
What to include. Two to three sentences naming the specific numeric triggers that pause or kill the project, tied to the eval bar and the ROI staircase. Each trigger has a number and a checkpoint date.
What to leave out. Vague reassessment language. “We will reassess at year-end if results are not on track” is not a kill clause. The CFO has heard that phrase a hundred times and discounted it accordingly.
Length test. A reader should be able to determine, on any future day, whether the project is approaching a kill trigger. If the trigger is not specific enough to be evaluated against a metric, it is not a trigger.
Common failure. “We will check progress quarterly and adjust as needed.” Wrong. The right form: “Kill triggers: eval-pass-rate below 70% for two consecutive months pauses the project for root-cause review. Stage 1 cost-out target missed by more than 30% at month 9 kills the project and reassigns the team. Vendor-side jailbreak rate above 0.5% at any point pauses for a security review. Triggers are reviewed at most quarterly board checkpoint.”
Section 5: ROI staircase positioning
What to include. Two to three sentences naming which stage of the ROI staircase the project is on now, which stage it is targeting next, and the timeline and dependency for the transition. The staircase has four stages; cost-out, capability earned, revenue-in, moat; detailed in the companion piece on the AI ROI staircase.
What to leave out. Single-point ROI percentages. Boards have a calibrated discount on AI ROI claims; presenting a single number triggers the discount and undermines the credibility of the rest of the thesis.
Length test. The position should be one specific stage with one specific transition target. A project claiming to operate at multiple stages simultaneously is either a portfolio (which is fine, but each project gets its own thesis) or unfocused.
Common failure. “We project 25% ROI over 24 months.” Wrong. The right form: “Stage 1 cost-out: targeting 18% reduction in average ticket handling time, measured against the pre-rollout baseline, producing approximately 280K to 360K in operational efficiency converted to capacity. Targeting Stage 2 capability transition by month 18, dependent on the eval suite reaching the bar specified in Section 2. Stage 3 revenue is not modeled in this thesis; that is a future commitment.”
Section 6: Alternative-cost analysis
What to include. Two to three sentences naming the realistic alternative to the AI project, the cost of the alternative on the same 24-month horizon, and the structural reason the AI project is preferred. The alternative is not “do nothing”; it is the next-best capital allocation.
What to leave out. Strawman alternatives. Comparing AI to a hand-built rule-based system from 2015 is not honest; comparing it to a 2026 hire-more-people approach or a 2026 vendor-SaaS approach is.
Length test. A skeptical board member should agree the alternative is a real option that was honestly priced. If the alternative reads as deliberately weakened, the comparison fails.
Common failure. “There is no alternative.” Wrong. The right form: “The alternative is hiring four additional support agents at fully-loaded 110K each, a 24-month cost of 880K plus management overhead. The hiring path scales linearly with ticket volume; the AI path has a 24-month TCO of 1.4M to 1.8M but scales sublinearly and produces a Stage 2 capability the hiring path cannot. The structural argument for AI is the capability optionality at Stage 2; the cost-out math is roughly comparable; the capability difference is the decision.”
Worked example
Here is the full one-page thesis for the customer-support AI project running through the previous sections, end-to-end, at the length a real one-page artifact runs.
Investment Thesis: AI-Augmented Customer Support, FY26 to FY28
Problem. Customer support handles 12,000 tickets per month with a 26-hour median time-to-resolution; the team is 18% understaffed against inbound volume; satisfaction is 8 points below target on tickets requiring research. The cost is approximately 2.4M annually in churn risk and team burnout, growing at 12% per year as volume scales.
Eval-defined success. The model must achieve 78% accuracy on the 2,400-item customer-support eval set, with p95 latency under 1.8 seconds, sustained for two consecutive weeks before promotion to production. The eval set is owned by the customer-support quality team and refreshed quarterly.
Budget cap. 1.4M to 1.8M over 24 months. Decomposition: inference 850K to 1.0M with 15% volatility reserve, eval suite 180K capex amortized over 36 months, model deprecation reserve 120K, on-call and platform tooling 170K, engineering 200K. Bracket reflects model-price uncertainty.
Kill clause. Eval-pass-rate below 70% for two consecutive months pauses for root-cause review. Stage 1 cost-out target missed by more than 30% at month 9 kills the project. Vendor jailbreak rate above 0.5% at any point pauses for security review.
ROI staircase. Stage 1 cost-out: targeting 18% reduction in median time-to-resolution, producing 280K to 360K in operational efficiency converted to expanded capacity. Targeting Stage 2 capability transition (multilingual support and tier-2 ticket handling) by month 18, dependent on the eval bar.
Alternative-cost. The realistic alternative is hiring four additional support agents at fully-loaded 110K each; 24-month cost of 880K plus management overhead. The AI path has a higher 24-month TCO but scales sublinearly with volume and produces a Stage 2 capability the hiring path cannot. The decision is the capability optionality, not the cost-out math.
The full thesis is 280 words. It pre-answers the six capital-allocation questions, leaves room for board questions on any of them, and converts the meeting from defense to sign-off. The slide deck behind it can be 30 slides; the artifact the board votes on is the one page.
Frequently asked questions
Why does a one-page thesis convert better than a 30-slide deck?
Boards approve capital allocation, which requires answers to six questions: problem, success, cost, kill condition, return, alternative. A 30-slide deck reorganizes those answers into narrative that is good for storytelling and bad for capital allocation, so the board spends the meeting extracting answers rather than making decisions. The one-page thesis pre-extracts the answers, ranks them by salience, and converts the meeting from extraction to decision.
What is the right length for each section?
Two to three sentences. The brevity is a forcing function; most line has to be defensible because there is no room for fluff. Sections that need more than three sentences either contain content that belongs in the slide deck or signal that the sponsor does not have a crisp answer yet.
How should the budget cap be presented?
As a 24-month bracket with decomposed lines, not a single-point annual number. A defensible 1.4M to 1.8M bracket beats a fake-precise 1.62M number with the board because the bracket reflects honest uncertainty about model prices and usage scaling. Single-point precision the sponsor does not have is read as overconfidence and discounted accordingly.
Why must kill triggers be numeric?
Because vague reassessment language has been used too many times for boards to trust it. CFOs have approved “we will reassess at year-end” projects that turned into “let’s give it one more quarter” three cycles in a row. Numeric triggers; eval-pass-rate, cost-out target percentage, jailbreak rate; are evaluable on any future day, which is what makes the kill clause a real decision rule rather than a deferral.
How does the thesis handle ROI?
By positioning the project on the ROI staircase rather than reporting a single ROI percentage. The staircase has four stages; cost-out, capability earned, revenue-in, moat; and the thesis names the current stage, the next stage, the transition timeline, and the dependency. The structure replaces the unwinnable single-number debate with a transition-discipline conversation.
What makes an alternative-cost analysis honest?
A realistic alternative that is honestly priced on the same 24-month horizon. Comparisons to strawman alternatives (hand-built systems from 2015, “do nothing”) fail because skeptical board members detect the strawman and discount the entire thesis. The honest alternative is the next-best 2026 capital allocation; typically hiring, vendor SaaS, or a different AI project; priced and named.
Should the thesis mention specific vendors or models?
Only when vendor concentration is a material risk that the board would otherwise discover from the slide deck. The default is to keep vendor specifics out of the one-page thesis and into the slide deck behind it, because vendor names age fast and the thesis should outlive any specific frontier model release.
How often should the thesis be refreshed?
Quarterly at minimum, with each refresh tracking against the kill triggers and the staircase transition plan. The thesis is a living artifact, not a one-time approval document. The refresh is short; usually a paragraph per section noting what has changed; and is the artifact the board reviews at the quarterly checkpoint.
Can the thesis be used for AI projects led by an external agency?
Yes, with the same six sections. The agency’s role appears in Section 3 (budget) and Section 4 (kill triggers, including agency-side performance triggers), and the alternative-cost analysis in Section 6 should include the in-house alternative as the comparison. The thesis is owned by the buyer-side sponsor regardless of who delivers the work.
Key takeaways
- The one-page investment thesis has six sections in fixed order: problem, eval-defined success criteria, budget cap with TCO bracket, kill clause with numeric triggers, ROI staircase positioning, alternative-cost analysis. Each section is two to three sentences.
- The thesis is the artifact the board votes on; the 30-slide deck is the work behind it. Boards that approve from theses approve faster and at higher hit rates than boards that approve from decks.
- Each section has a defensibility test. Problem statement: would an outsider know what the project is for? Success criteria: is the eval bar reproducible? Budget cap: is the bracket honest? Kill clause: are triggers numeric and evaluable? ROI: is one specific stage named with one specific transition? Alternative: is the alternative real and honestly priced?
- The thesis is a forcing function on the sponsor. Sections that need more than three sentences signal the sponsor does not yet have a crisp answer. Boards that demand the thesis before approving budget filter out unready projects, which is what produces higher downstream hit rates.
- The thesis is the board-facing expression of the feature-cost to evaluation-cost shift. Eval-defined success criteria, kill triggers tied to the eval bar, and ROI staircase positioning many flow from treating evaluation cost as the unit of AI economics rather than feature cost.
- The 280-word worked example shows the full thesis at length. Sponsors who can write the thesis at that length have done the work; sponsors who cannot have not yet earned the budget approval the thesis would secure.
The thesis does not replace the slide deck. The slide deck explains the thesis; the thesis is what the board approves. Sponsors who produce both convert; sponsors who produce only the deck defer.
Arthur Wandzel
