25 Questions Every Leadership Team Should Be Asking About AI in 2026
The questions that separate AI strategy from AI theatre, reordered into a five-act playbook. Print it, take it to your next leadership offsite, and do not move on until every question has an answer your CFO would defend.
Most AI strategies are answers in search of a question. A vendor shows up with a demo, a pilot gets funded, a steerco gets stood up, and twelve months later nobody can say with a straight face what changed in the P&L.
The fix is upstream. Before the budget, before the vendor selection, before the pilot, the leadership team has to sit in a room and answer the right 25 questions. Not 25 questions about technology. 25 questions about the business: what changes, who pays for it, what gets eroded, what gets amplified, and what you tell the people whose jobs are about to look very different.
The list below is reordered into a five-act sequence: frame the bet, pick the strategy, make the calls, wire it through the organization, and take the people with you. Each act builds on the previous one. Skipping forward is the most common cause of failure.
One housekeeping note. Five of the 25 are math questions. When you get to those, stop debating opinions and run the numbers through a sourced calculator. Most strategy arguments are actually unresolved arithmetic.
Act I — Frame the bet
Most AI strategies fail in the first hour, not the first quarter. The framing is wrong, the menu is fuzzy, and nobody has named what is actually at stake. Before you fund a single pilot, sit your leadership team in a room and answer these four.
- Q1
What does "AI transformation" actually mean for us, and what is the real menu of opportunity?
If the answer is "we will be more efficient," you do not have a strategy, you have a hope. The honest menu has five courses: cost-out, throughput, quality, new products, and new business models. Pick which two you are actually playing for. The other three will pull resources away from the ones that matter.
- Q2
What are the risks of doing this, and the risks of not doing it?
Both sides of the ledger. The risk of moving is wasted spend, change fatigue, security exposure, and a public failure. The risk of standing still is margin compression as competitors run with 30 to 45% lower unit cost. Name both, then weigh them. A board that has only heard the upside has not been briefed.
- Q3
Which businesses have already transformed successfully with AI, which have failed, and what can we steal from each?
The winners (Klarna in support, Moderna in research ops, Bayer in regulatory) share three traits: a measured baseline, executive sponsorship, and a culture that ships. The cautionary tales (the big consulting-led "AI factories" that quietly disbanded in 2025) share one trait: no measured baseline. Steal from the structure, not the headcount number.
- Q4
If a competitor launched tomorrow, AI-native from day one, what would they do that we would not, and why are we not doing it?
This is the most useful exercise in the room and the one most teams skip. Spend 90 minutes designing the AI-native competitor that would beat you. Then ask why your business cannot ship that. The answer is usually a sacred cow, not a technical constraint.
- Q5
Which of our current competitive advantages does AI erode, and which does it amplify?
Scale economies in repetitive work get eroded. Proprietary data, distribution, brand, and trusted relationships get amplified. Map your moats into one column or the other. The eroding column is where AI is a defensive necessity. The amplifying column is where it is offensive leverage.
Act II — Pick the strategy
Five questions about structure, sequencing, and ambition. This is where "we should do something about AI" becomes a programme with a budget, an owner, and a written plan.
- Q1
How do we go from a bunch of disconnected experiments to a coherent strategy with real rigor?
A coherent strategy has three artefacts: a written outcome thesis (what changes in the P&L), a portfolio map (which workflows, in what order), and a measurement cadence (monthly value tracking, not annual). Without all three, experiments stay experiments and the budget gets cut in the next cycle.
- Q2
Should AI transformation live with a central steerco or be embedded in every function?
Both, sequenced. Year one is central: standards, security, vendor selection, the first three flagship use cases. Year two pushes ownership to the functions, with the centre shifting to enablement and governance. Pure federation in year one creates twenty pilots and zero platforms. Pure centralisation in year three creates a bottleneck nobody can route around.
- Q3
Where do we actually start?
Where the workflow is high-volume, the data is structured, and the value per unit is measurable. That is almost always finance (AP, AR, close), service operations (tier-1 deflection), or sales operations (lead enrichment, proposal drafting). Start there, ship in 90 days, then use the win to fund the harder bets.
- Q4
How do we avoid bolting AI onto existing processes and reimagine them from first principles?
Run the process map twice. Once with the current steps, once as if the process were being designed in 2026 with no legacy. Compare them. The delta is the real opportunity. Bolting AI onto the old flow gets you 15%. Redesigning gets you 60%. The hard part is having the political authority to redesign.
- Q5
If a task that used to cost $100 now costs $1, what is suddenly worth doing that was not worth it before?
This is the single highest-leverage question in the strategy session. The answers (personalised onboarding for every customer, deep research before every sales call, daily competitive intelligence, per-account quarterly business reviews) are usually services your customers always wanted but you could never afford to deliver. Find five and you have a roadmap.
Act III — Make the calls
Strategy fails in execution because the decision frameworks are missing. These five questions give you the rails: how to measure value honestly, where to take risk, how to stay flexible as the vendor market churns.
- Q1
How do we tell the difference between AI activity and AI productivity?
Activity is measured in tokens, users, and pilots launched. Productivity is measured in invoices touched per FTE, days to close, DSO, and cost per ticket. If your AI dashboard does not show one of the second list, it is a vanity dashboard. The fix: every use case ships with a baseline measurement and a monthly delta, or it does not ship.
- Q2
What is our risk framework for go/no-go decisions on AI tooling and systems?
Four axes, scored low/medium/high: data sensitivity, decision reversibility, customer visibility, and vendor maturity. Anything scoring high on three or more axes needs board approval. Anything scoring low on three or more axes is a function-level decision and should not need a committee. Write the matrix once, apply it consistently.
- Q3
How do we build a data strategy that gets us progressively AI-ready without blocking the work we could start today?
Stop trying to fix everything before starting anything. The right move is parallel: ship the first three use cases on the data you have, while a small dedicated team cleans the next three domains in priority order. A perfect data layer in 18 months loses to a working pilot in 90 days.
- Q4
How do we build systems with model-agnostic harnesses so we are insulated from a volatile vendor market?
Three rules. One: every model call goes through an internal abstraction, never a vendor SDK in business logic. Two: prompts, evaluations, and routing live in your repo, not the vendor console. Three: every workflow has a documented eval set so you can swap models in a week, not a quarter. Vendor lock-in is a choice, not a fate.
- Q5
If we lean on external AI vendors and LLM costs spike, what is our exposure?
Model the scenario. Map your top five use cases to their tokens-per-transaction and current per-token cost. Now run the math at 3x and 10x. If 3x kills the business case, the use case needs a smaller model, a cached layer, or an on-prem fallback before it ships. Cost has dropped 90% in two years. That trend will not continue forever.
Act IV — Wire it through the organization
Strategy and frameworks do nothing without the right people in the right loops. Five questions about how use cases flow up, how leadership models the work, and how the functions traditionally cast as blockers become the heroes of the story.
- Q1
How do we bubble use cases up from employees and prioritize/productionize them from the top?
A two-lane process. Lane one is bottom-up: a low-friction submission form, a 30-day review cadence, named sponsors per function. Lane two is top-down: a portfolio committee that picks the five worth productionising each quarter, with budget and a delivery owner. Lane one without lane two creates a graveyard. Lane two without lane one creates a strategy disconnected from the work.
- Q2
How does leadership get its hands dirty and actually build, so it can lead by example?
Every member of the executive team builds and ships one AI workflow per quarter, presented to the board. Not delegated, not "directed." Built. If the CEO cannot describe their own use case in 90 seconds, the rest of the organisation will read the strategy as performative. This is the single highest-leverage cultural intervention available, and it is free.
- Q3
How do we find the AI A-players already in our org and make what they do the gold standard?
They are already there, usually in roles you do not expect: a finance analyst who quietly automated the close, a support agent who built a knowledge bot, a product manager who replaced their research with a research agent. Find them with a one-question survey ("what have you built or automated for yourself this year?"), pay them, promote them, and put them on the steerco.
- Q4
How do we test for AI curiosity, literacy, and interest during hiring?
Add a 15-minute working session to the loop. Hand the candidate a real workflow problem and access to a chat model. Watch how they use it. The signal is not whether they get the right answer. It is whether they iterate, criticise their own prompts, and notice when the model is wrong. This separates curious from credentialed faster than any reference check.
- Q5
How do we make IT, legal, and compliance partners and heroes in the story, not bottlenecks?
Bring them into the use-case selection, not the use-case review. The order matters. If they are consulted at design time, they shape the architecture. If they are consulted at deploy time, they block it. Give them a budget, a public win to claim, and a seat on the steerco. The framing shifts from "the bad guys" to "the team that made it possible to ship."
Act V — Take the people with you
The hardest five. Strategy and tooling decisions are intellectually demanding. People decisions are emotionally demanding. Most AI programmes that stall do so on this act, not the previous four.
- Q1
How honest is the "we will redeploy people to higher-value work" narrative, really?
Partially. The honest version is: roles change, some shrink, some grow, and a minority disappear. Redeployment is real for the people whose roles are 30 to 60% routine. It is fiction for the roles that are 90% routine. Be specific by function. A blanket "no one will lose their job" pledge corrodes trust the moment the first restructuring lands.
- Q2
Where does the line fall between our responsibility for upskilling and our employees'?
A workable split: the company funds the curriculum, the time, and the access to tools. The employee owns the choice to engage and the application to their own work. Mandating learning without time is theatre. Offering time without expectation is wasted. The contract has to be explicit and the same for everyone.
- Q3
How do we transform a culture when most employees are apathetic, fearful, or quietly waiting it out?
Not with town halls. With visible wins from people who look like them. The most effective intervention is a monthly "show and tell" where individual contributors demo what they built and what it changed in their week. Apathy converts to curiosity faster watching a peer than listening to an executive. Fear converts to participation when the peer is still employed and visibly better off.
- Q4
How do we message our AI strategy honestly and with empathy for the people doing the work?
Three rules. Name what changes for whom, by function, with timelines. Acknowledge what people will lose, not just gain. Give them agency in the design. The message that builds trust is "here is what is changing, here is what we are doing for you, here is what we need from you." The message that destroys trust is "this will only make your job better." Nobody believes it.
- Q5
How do we create a culture of experimentation without taking on security risk we cannot stomach?
Two surfaces. A sandbox tier with synthetic or fully anonymised data, where experimentation is encouraged and approval is automatic. A production tier with real data, where the risk matrix applies and approval is explicit. Most companies have only the production tier and wonder why nobody experiments. Build the sandbox first and most of the cultural battle is already won.
Stop arguing. Start sourcing.
Three of the five acts hinge on numbers, not opinions: the cost of the bet, the payback period, and the working capital unlocked. The AI Automation ROI Calculator shows every constant, cites every source, and emits a one-page PDF you can paste straight into a CFO memo.
If you can only answer five
Most leadership teams will not get through all 25 in a single session, and that is fine. If you have one offsite, prioritise these five: Q1 from Act I (what does transformation actually mean), Q1 from Act II (how do we move from experiments to strategy), Q1 from Act III (activity vs productivity), Q2 from Act IV (leadership building, not just sponsoring), and Q4 from Act V (honest, empathetic messaging).
Those five are the load-bearing walls. The other 20 are the rooms inside. Get the walls right and the rest can be scheduled across the next quarter without anything falling over.