If you're accountable for AI results right now, you're probably funding more than you can defend. Pilots everywhere, a board asking what came back, vendors adding use cases faster than you can evaluate them. Here's what I do about that: I get you behavioural evidence on every bet and a verdict you can stand behind, fund, fix, or kill, so the budget moves to where it earns. I sell no build and no software, and my own company runs on AI agents, so the advice comes from someone who operates this stuff daily and has nothing to gain from your answer being “build more”.
The 2026 landscape
I spend my weeks in executive sessions and portfolio reviews. These are composites from real conversations in 2025 and 2026, anonymised.16 Skip any that don't sound like your building.
Somewhere in your business, a capable manager has already vibe-coded a ServiceNow-sized internal tool your governance can't see. It works, nobody owns it, and other teams now depend on it. You don't want to punish that person. You do want to know what's been built, keep what's good, and retire the rest.
Every renewal on your desk now carries an AI roadmap and a list of use cases they'd like you to fund. A vendor roadmap is priced on adoption, not on your return. Someone at your table has to ask what each use case must prove before it earns budget.
Your pilots demo well, then don't reach the P&L: 95% of enterprise GenAI pilots fail to produce measurable impact. Demand grows faster than budget every year, and most pilots never get a clear decision either way, so the spend continues by default.
Your board's question has changed from ‘what's our AI strategy?’ to ‘what did the AI spend return?’ Answering it takes behavioural evidence: who used it, what changed, what it cost. A business case doesn't answer it.
When agents make building cheap, engineering stops being your constraint and choosing becomes it. Your funding gates, review cadence and operating model were designed for a world where building was the expensive part. The organisations that win the next two years will be the ones that learn fastest per dollar, and most AI portfolios are still optimised for building, not learning.
The fix is an investment discipline, applied bet by bet, held by someone independent. That's the next section.
The front-door offer
2 weeks · fixed fee $25k to $50k8 · work-email gated
Bring your live and proposed AI bets. I map them, rank them on value and evidence, and pressure-test the highest-stakes ones with real behavioural experiments. You walk into the board with a call you can defend on every single one.2
01 Map
Every live and proposed bet in one view. Including the weekend builds.
02 Rank
Value, evidence, adoption, governance risk, time to impact. One page.
03 Test
Fast behavioural tests on the high-stakes bets. Behaviour, not workshop opinion.
04 Decide
A verdict per bet, evidence attached, ready for the board.
Your whole portfolio, mapped: every live and proposed AI bet across the business, on one page
A ranking you can argue from: value, evidence and risk, comparable bet to bet
Real evidence on the big ones: 1 to 3 live behavioural tests on the highest-stakes bets
Something working, where useful: a lightweight demo, concierge workflow, or simulated agent
A funding decision per bet: and a recommended next step for the ones worth backing
Built for the people accountable to the board
Works best in these sectors
The dynamic portfolio view I use to make evidence visible: confidence across the x-axis, experiment depth on the y-axis, value in the bubble size, and the next funding call in the colour.
Exponentially ideas
684 ideas, 136 experiments in account
Stop / explain
Scale / fund
Early / park
Prove next
Confidence
Showing 24 ideas with a confidence value. Effort and value scores are used as metadata, not extra axes.
Shortlist
1. Idea 01
2. Idea 02
3. Idea 03
4. Idea 04
5. Idea 05
Ways to work with me
The door
Half a day with your leadership team. I bring my own agent operation, live, not slides, and an outside read on your portfolio. This is how most engagements start.
You leave with: your portfolio first-pass scored in the room, a working picture of what a business run on agents looks like (costs included), and a shared language for AI funding decisions. Scoped in a 20-minute call.
The retainer
The audit is a snapshot. Bets drift. The retainer keeps the discipline live between budget cycles. A large firm can't sell you this: every recommendation it makes can become its own delivery revenue. Mine can't.
You get: monthly portfolio reviews with every bet re-scored, a verdict on each new AI proposal before it reaches your investment committee, guardrails that catch the weekend builds without punishing the builders, and an independent voice when vendors bring roadmaps or the board wants a straight answer. 6 to 12 month terms, limited seats.
The room
The argument, delivered to your board or your conference: why most AI pilots never reach the P&L, and what the ones that do have in common.
You may have seen me at: Stanford, where I guest lecture in the course where the method is taught, and internal keynotes for PEXA, NAB and an ASX-listed enterprise group.
How I decide
Unchanged for years, because they keep being right. Built testing customer demand; they apply unchanged to AI.
Every verdict is anchored to what real people did when offered the thing, not what they said in a survey or a steering meeting.
Most of the cost of a failed product is paid before anyone tests whether it should exist. So the test goes before the build, while being wrong is still cheap.
RACQ killed a multimillion dollar idea in 3 days. The verdict landed while the budget decision was still open, which is the only time it's useful to you.
ANU stopped 24 weak ideas in 12 weeks. Every cheap kill hands its budget to a bet that earned it.
AI made building faster, which makes choosing harder. When every function can ship a pilot, the scarce skill is deciding which of the 24 things you could fund deserve capital at all. The difference between the pilots that stall and the ones that pay isn't model quality, it's evidence discipline before the next cheque.
Why trust the read
Before I advised anyone, I ran the pre-AI version of your world from the inside: enterprise innovation lead, repeat founder. A decade of enterprise experiments taught me what evidence looks like before capital moves. The full record is in section 6 and the timeline in section 7.
Then I rebuilt my own company on the new stack. Agents write most of my code, run the CRM and prep my meetings; the figures at the top of this page come from their commit history. I pay the token bills, review the failures, and I've killed my own AI projects when the evidence said to. So when a vendor tells you what agents will do for you, I can tell you what they actually did for me, and what it cost.
AI does the portfolio administration. I make the portfolio call. That's why the audit takes 2 weeks, not a quarter, and why the verdict is signed by the person who did the work: 20% of my operations run on agents, 172 hours of AI development went into 45 days of client work, hundreds of AI-run customer interviews get recruited, conducted and synthesised in days, and one production MCP server shipped inside Rapidly software this year.
The record · appendix A
Every dollar is a bet that paid off, or one we never made.
Tracked per experiment, per company, since 2017. Both sides count: money made by ideas that passed, and money not spent on ideas that failed. Every figure below sits on a public case study.
The record since 2017 also includes Racing Victoria, Reece, Treasury Wine Estates, Bupa and NSW Health.
“Within the first 15 minutes, I could see there was something I'd been missing that I didn't even know I needed.”
Patrick Smith · Head of Innovation, Tabcorp
“You are smarter when you are making decisions based on behavioural data, not your gut. The data is real.”
Richard Guy · Experimentation Manager, AGL
One page from a completed audit, client identifiers redacted: the bet, the evidence gathered in 2 weeks, and the call the board accepted. If the format doesn't convince you, the audit won't either. Better to know now.
Who you're hiring · appendix B
I ran innovation inside large companies before I advised any: Head of Innovation at ThoughtWorks Australia, then Sportsbet. Before that I built four companies and sold two: Exertrack, acquired by Techstars-backed Gyminee, and GetViable, acquired by Bigcolors after 1,500 founders across 35 countries used it to test their ideas. Since 2017 I've run Exponentially from Melbourne, profitable and customer funded, doing one thing: testing whether ideas deserve capital before the capital is spent. I sit on the PHORIA board and trained with the Australian Institute of Company Directors.
The method behind the audit is pretotyping. It was created at Google by Alberto Savoia and is taught at Stanford. I trained directly with Alberto, guest-lecture in the Stanford course where he teaches it, and am one of a small number of practitioners listed on his site. That's the only place the method's name appears in this memo. What you're buying is the verdicts it produces, now applied to AI bets: agents, copilots, models.
“Leslie Barry is one of the world's leading pretotyping practitioners. He has trained thousands of innovators and helped enterprises save millions by testing ideas before building them.”Alberto Savoia · creator of pretotyping, first Engineering Director at Google
Between engagements: I write Fund, Fix, Kill, on the AI bets I back and stop in my own business, and keep a practical YouTube channel of experiments that worked and failed.
Next step
Bring your AI spend. Two weeks later you walk into the board with a verdict on every bet, and the evidence attached.
Not ready for the audit? Email one bet you want more confidence in to leslie@exponentially.com. In 20 minutes on a call you'll get my honest read and a yes or no on whether this is a useful next step. No pitch.
Leslie
Leslie Barry · Melbourne · LinkedIn
Footnotes