I took my framework to Harvard and Oxford

I built the AGENTIC Framework, an AI adoption methodology, from practice: two live organisations, hundreds of workflow conversations, no formal AI credentials. This year I took it to Harvard Business School’s AI for Leaders and Oxford’s Generative and Agentic AI Programme to find out what would survive contact with the best academic thinking available. The short version: the spine held, eight things changed, and the biggest gaps I found ran in the other direction. This is the honest account, receipts included.

Why take a working methodology back to school

The framework was already doing its job at a marine conservation non-profit and a minerals venture studio. But the online AI space runs hot, and I’d reached the point where I wanted to know how the organisations with the most at stake actually do this. How Moderna put AI in front of every employee. How P&G scaled past its first wins. What Professor Michael Wooldridge, who earned his PhD in multi-agent systems in 1992, considers real rigour in agentic AI.

So I enrolled in both programmes and made a deal with myself before starting: publish what holds, change what doesn’t, and record every change in the framework’s version history. A methodology that can’t name what it learned from eight weeks with better evidence isn’t being tested. It’s being defended.

What held up

The core of the AGENTIC Framework survived both courses, and in a few places arrived earlier than the material being taught. The convergences were specific enough to be useful rather than flattering.

• Start by listening to the people doing the work. Moderna’s leadership puts it as “listen before you think.” Their highest-value use cases surfaced from the floor, not the boardroom. Management’s list is rarely the best list, and now I have their data to prove it.
• Fix the workflow before you formalise it. Oxford states it verbatim: you cannot automate a broken process. It’s named as one of the four hidden costs of agentic deployment. AGENTIC has it as a stage.
• Design collaboration per step, not per system. Harvard’s automate-or-augment framing works at whole-system level. The Collaboration Spectrum does the same work at finer grain.
• Prove agents in parallel-run before handoff. Oxford’s shadow-mode doctrine: real inputs in, outbound actions blocked, evidence before trust.
• Not everything needs to be an agent. Oxford’s adoption matrix puts agents in one of four solution clusters. “The simpler the better” is a conviction I’ve carried since my systems-design years; hearing it from the Oxford teaching team just gave it a citation.
• Governance is checkpoints, not policy documents. Moderna scripts its compliance surveillance because human oversight doesn’t scale. Principles only work once they’re operational.

One validation stood above the rest. Harvard’s competitive-advantage module argues that foundation models are a general-purpose technology: every competitor buys the same capability from the same vendors, so the models themselves confer no lasting advantage. What does is everything built around them that can’t be bought. The strategy literature calls these complementary assets.

That reframed the AGENTIC Vault for me. I’d been describing it operationally, as the thing that makes the tenth workflow faster than the first. The stronger claim is strategic: two organisations with identical model subscriptions are not equal, because the one with eighteen months of structured specifications, override history, and proven patterns is compounding an asset the other hasn’t started.

What changed after Harvard

The Harvard integration shipped this month as V6.2. The full list is in the version history, but the change that matters most deserves the space here.

The framework used to open with a quiet assumption: that organisations were past “should we use AI?” and stuck on “how do we do this properly?” Harvard has a name for why that assumption fails. The capability-adoption gap: access to AI arrives overnight, but the ability to use it well doesn’t. Most organisations now have the tools. Far fewer have a team that’s fluent in them.

That gap breaks my own entry point. Run Kickoff conversations with a function that has never worked with these tools and you get candidates skewed toward automating today’s pain, described by people who can’t yet imagine the work being different. Worse, their first contact with AI becomes the day it starts changing their job. The adoption stream inherits a trust deficit it didn’t create.

So V6.2 adds a fluency ramp ahead of the first scan: tools in hands on organisation accounts, a one-page guardrails sheet, a champion per function, one shipped quick win each. Weeks, not months, and Moderna supplied the evidence for the sequencing. They rolled ChatGPT out to their entire employee base to play with first and deliberately didn’t chase ROI in that phase, because learning was the return. By the time they redesigned workflows, people already trusted the tools, so they weren’t fighting adoption and implementation at the same time.

Four more Harvard changes shipped alongside it. The Prioritisation Matrix now flags strategic value separately from operational value, because a portfolio scored only on efficiency automates a hundred small processes and never changes what the organisation is. The adoption stream gained a segmentation step (capability and buy-in place every affected person, and each quadrant needs a different intervention) plus three design laws drawn from Pernod Ricard’s deployment data: design for fast feedback, spend on the interface, and teach the limitations openly. Their sales tool delivered weekly recommendations and hit 85 percent acceptance. Their marketing model asked people to interpret response curves over months, and stalled. Same company, same technology, opposite adoption outcomes, and the difference was design.

The governance stream now covers who may build, not just what gets built: three access tiers, from consumers through controlled democratisation to full producers. And the scaling guidance finally answers a question I’d been dodging: where the AI capability should sit organisationally. Past ten or so workflows, hub-and-spoke wins. P&G reached the same conclusion scaling across a business too diverse for either a pure central team or pure local autonomy.

What’s coming from Oxford

The Oxford integration lands next as V7, with its own write-up. The headline changes: Greenlight gains a one-page business case. My own line was always that scoring is analytical and commitment is political, and Oxford’s PROFIT structure, with its explicit do-nothing option, is the missing artefact between a green score and a resourced decision. Kickoff candidates get routed through four solution clusters before anything enters the pipeline, so a better prompt or a retrieval setup counts as a first-class outcome. And governance gains an adversarial layer, because the framework’s own success creates its sharpest new risk: once specifications drive everything, poisoning the Vault becomes the highest-value attack. More on all of that when it ships.

What neither of them teaches

Here’s what surprised me most. Two of the best AI programmes available, and the two ends of the adoption lifecycle are missing from both.

Both courses assume use cases arrive somehow. Harvard’s cases begin once leadership has chosen a direction. Oxford’s business-case module starts from “identify the opportunity” without a mechanism for identification. Neither has an operational answer to the question every organisation actually asks first: where do we even start? That’s exactly the job of Kickoff’s two-pass model, and I checked carefully. It has no equivalent in either curriculum.

The exit is missing too. Oxford gestures at lifecycle management. Harvard doesn’t touch it. Neither treats switching an agent off as a governed event with the same rigour as switching one on, which is the entire argument of the framework’s decommission path.

Five more gaps, briefly. Neither course has the two-layer specification (a clean SOP for humans, an exhaustive spec for agents, intentionally divergent). Neither treats human overrides as structured learning data with typing and severity, the way Nurture does. Neither watches for tool substitution, the adoption signal where people quietly do the work in ChatGPT because the built system isn’t meeting their needs. Neither scores key-person risk when prioritising, though staff dependency is often the argument that wins the sponsor. And neither makes stage gates machine-verifiable through required artefacts.

None of this is a criticism of the courses. They teach leaders to reason about AI, and they do it with evidence I could never generate alone: polls across more than a thousand executives, deployment data from global rollouts, faculty who wrote the field’s foundations. But reasoning and Monday morning are different jobs. The gap between them is the practitioner layer, and it’s exactly why I built one.

What I’d tell you if you’re doing the same

Take the courses after you’ve built something, not before. Every case study landed harder because I had two implementations to test it against, and every framework they taught got audited against workflows I actually run. The credential was the least valuable thing I left with.

Be prepared to push back as well as absorb. I still think maturity assessments produce reports rather than next steps. I still think the act-early-versus-wait question is economics rather than courage, and the falling cost curve answers it more honestly than enthusiasm does. Disagreement, written down, is part of the pressure test.

And keep the receipt. Everything this process changed is dated and reasoned in the framework’s version history, next to the things it didn’t change. V6.2 is live now. V7 arrives with the Oxford piece. If your AI adoption methodology can’t name what it learned from eight weeks with better evidence, it isn’t being tested. It’s being defended.

Version log

Published 3 July 2026 · Post-Harvard update. This article documents the V6.2 Harvard integration and previews the Oxford V7 integration. The individual V6.2 changes are also noted on the relevant framework pages and recorded in the public version history.