AI adoption methodologies: what Anthropic and OpenAI propose, and how they differ

An AI adoption methodology is the set of phases, instruments and decision criteria by which an organisation moves from experimenting with artificial intelligence to capturing measurable value. The two most relevant vendors of the moment — Anthropic and OpenAI — publish official methodologies with opposing philosophies: Anthropic’s is top-down, because it begins with strategy and governance; OpenAI’s is bottom-up, because it begins with employees discovering use cases.

What follows: what each methodology examines, why one is top-down and the other bottom-up, the table that summarises the difference, the impact each vendor documents, and which approach suits which organisational profile.

I spent the last few weeks reading both guides closely, including their extended and sector-specific versions, and in this analysis I share what each one examines, why they differ, and what that difference implies for anyone who has to choose a path today.

Why vendor methodologies interest me

The problem both guides set out to solve is the same one, and it is well documented: 92% of companies plan to invest in AI over the next three years, yet barely 1% consider their investments to have reached full maturity. The gap is not technological — the models are already capable — but organisational: most initiatives stall between the promising pilot and sustained adoption. The vendor methodologies are, in essence, two different diagnoses of where that chain breaks and therefore two different prescriptions for repairing it.

There is a further reason I recommend reading them side by side: each reveals its author’s mental model. No framework is neutral. Both are the codification of how each vendor has watched its own customers succeed — and fail — and that imprint shows in every design decision.

What Anthropic’s methodology analyses

Anthropic’s Enterprise AI Transformation Guide, built on implementations with organisations such as Cox Automotive, Thomson Reuters and NBIM, structures adoption into three sequential steps: laying the foundations, launching a pilot and scaling the impact.

What struck me most is not the phases — any consultancy proposes three or four — but what information each one gathers. The foundation phase examines the organisation before the technology: clarity of the AI strategy and its link to business objectives, executive alignment and the formation of a steering committee, definition of the governance framework — ethical guidelines, model evaluation processes and incident response — and a readiness assessment that determines the appropriate transformation route. Only with that foundation is the pilot selected, under a demanding criterion I welcome: it must demonstrate value within 30 to 60 days and be measured on four fronts — adoption, efficiency, quality and satisfaction.

The extended version of the guide, produced with AWS, widens the model to four stages — develop the strategy, create business value, build for production and deploy — and devotes a whole chapter to a point that, in my experience accompanying organisations, almost all of them discover late: volume of data is not enough. Data has to be high quality, accessible and compliant with regulation, which demands mature data governance from day one, not as a later corrective project.

Why Anthropic’s approach is top-down

I find three factors that explain the top-down orientation, and all three are structural rather than stylistic.

The first is brand identity: Anthropic has positioned itself on the safety and reliability of AI — it was one of the first companies in the sector to certify against ISO 42001 for responsible AI — and a methodology that starts from governance is the natural extension of that thesis. The second is its customer base: its reference cases come from regulated or reputationally exposed industries (finance, legal, pharmaceutical, sovereign funds), where no use case reaches production without prior sign-off from risk and compliance. The third is its theory of failure: in Anthropic’s reading, AI projects do not die for lack of ideas but for lack of foundation — without executive alignment, governance and reliable data, the best pilot in the world never scales. Hence its information gathering is an organisational diagnosis that precedes any experiment.

What OpenAI’s methodology analyses

OpenAI splits its methodology across two complementary documents. AI in the Enterprise: Lessons from Seven Frontier Companies condenses seven field lessons: start with evaluations (evals), embed AI in the products, start now and invest early, customise the models, put AI in the hands of business experts, unblock the developers, and set bold automation goals.

The information gathering proper lives in the second guide, Identifying and Scaling AI Use Cases, built from more than 300 implementation cases and the experience of more than two million enterprise users. Its mechanics are distributed: instead of a central committee identifying opportunities, the whole organisation is taught to recognise them. To that end it proposes six use “primitives” — content creation, automation, research, coding, data analysis, and ideation and strategy — and three opportunity zones where AI performs consistently: repetitive low-value tasks, skill bottlenecks and navigating ambiguity. Discovery is triggered with workshops, hackathons and a field technique I find as simple as it is brilliant: asking each team for its anti to-do list, the inventory of what they find tedious and would rather never do again. I have found that the first use cases hide there in almost any organisation.

When the cases multiply — the guide itself warns that the challenge moves quickly from discovery to prioritisation — the central instrument enters: the Impact/Effort matrix, a quadrant scoring each case against value to the company and implementation effort. The result is four categories with different treatment: quick wins to build momentum, self-service for individual use, transformational projects that merit planning, and deferred initiatives worth parking until the technology makes them cheaper.

Why OpenAI’s approach is bottom-up

The bottom-up orientation also answers to the vendor’s DNA. OpenAI arrived in the enterprise from mass consumer use: by the time IT evaluates a corporate implementation, much of the workforce already uses the tool personally. Its methodology does not need to convince employees to experiment; it needs to channel experimentation that is already happening. Added to that is its culture of iterating with evals — measure before trusting, the first of its seven lessons — and an explicit bet on volume: in its model, a thousand employees exploring find more value than a committee of twelve planning. Leadership’s role is not to originate the use cases but to sponsor, prioritise and scale them.

The essential difference, in one table

If I had to summarise the difference in one sentence, it would be this: Anthropic answers how to govern adoption; OpenAI answers how to discover where to apply it. They are answers to different questions, which is why they complement each other more than they compete.

Documented benefits and impact

Both guides back their approach with customer results. Anthropic’s guide with AWS documents productivity improvements of between 20% and 50% in support, engineering, content and back office functions, and notes that the most advanced performers attribute more than 10% of their results to generative AI implementations. OpenAI’s guide, for its part, highlights cases such as an assistant handling two thirds of support chats and cutting resolution times from 11 minutes to 2, in an organisation where 90% of employees integrate AI into their daily work.

A clarification I consider obligatory: these figures come from the vendors and their cited customers, and should be read as illustrative rather than as promises transferable to any context. That said, the differential impact of each approach is clear. The top-down approach produces scalability with control: when the pilot works, the governance foundation already exists and scaling does not stall in legal or security. The bottom-up approach produces speed with legitimacy: the cases are born from those who live the process, which raises real adoption — precisely the point where I have seen most initiatives stall.

Which approach suits which organisation

In my view the choice is contextual rather than ideological, and three variables decide it.

The first is the regulatory profile: an organisation in finance, health or the public sector can hardly permit distributed discovery without a prior governance framework; a services or retail company with low-risk data can begin with the quick wins. The second is the existing digital culture: where employees already experiment with AI on their own account, the bottom-up approach channels energy that otherwise becomes shadow AI; where that base does not exist, imposing it produces empty workshops. The third is the available structure: the top-down model presupposes a steering committee, a centre of excellence and data governance teams; the bottom-up one presupposes a critical mass of experimenters and internal capacity to evaluate dozens of cases at once.

And here I reach the conclusion that matters most to me as a professional working in Latin America: neither methodology was designed for an organisation of 80 to 500 people. The structures the top-down approach presupposes, and the critical mass the bottom-up one presupposes, are the exception in our region’s mid-market. My reading is not that the frameworks should be discarded — their rigour is real and I have learned a great deal from both — but that they should be compressed: take the governance discipline of the first and the discovery mechanics of the second, at a scale of effort proportional to the size of the organisation and to its cost of discovery.

Read also: the benchmark of AI adoption methodologies from the large consultancies and the technology vendors, and what is missing to apply them in the mid-sized company.

Frequently asked questions

What is an AI adoption methodology?

It is the set of phases, information-gathering instruments and decision criteria by which an organisation moves from experimenting with AI to capturing sustained value. The methodologies published by Anthropic and OpenAI are today the most cited references in the market.

What is the difference between Anthropic’s approach and OpenAI’s?

Anthropic proposes a top-down sequence: strategy, governance and steering committee first, then pilots and scaling. OpenAI proposes bottom-up mechanics: employees discover use cases through six primitives and workshops, and leadership prioritises them with an impact and effort matrix.

Can the two methodologies be combined?

Yes, and in my experience that is the advisable route: the governance discipline of the top-down approach prevents uncontrolled proliferation, and the discovery mechanics of the bottom-up one prevent paralysis by planning. The combination requires adapting the scale of effort to the size of the organisation.

Which approach suits a mid-sized company in Latin America?

Neither in its pure state: both presuppose structures or critical mass that the regional mid-market rarely has. The advisable route is a proportional synthesis — a brief readiness diagnosis, discovery with the process owners, and prioritisation by impact and effort — ideally accompanied by a third party with experience in the segment.

Conclusion: methodology first, AI second

Having studied both guides closely, my first conclusion is one of recognition: Anthropic and OpenAI have made an enormous contribution by publishing, openly and free of charge, the distilled knowledge of thousands of real implementations. Only a few years ago this level of methodological detail — prioritisation instruments, measurement criteria, governance sequences tested in the field — existed only inside large consulting engagements, or was learned the hard way, project after project. Today any organisation, of any size, can reach it before investing its first peso in AI. That is a gift worth taking seriously.

The methodologies are not interchangeable recipes but two complementary diagnoses of the same problem: why investment in AI does not turn into maturity. One organises control; the other organises discovery. And together they teach something no technology can teach on its own: that AI adoption is, before anything else, an exercise in organisational knowledge. Knowing which processes you have, which data feeds them, who governs them, where they hurt and what relieving them would be worth — that is what both guides, each in its own way, force you to gather.

Which is why my central recommendation is one of order, not of tool: implement the methodology before moving any process to AI. The temptation to do it the other way round is understandable — the technology is available, the demos convince, and competitive pressure pushes — but moving a process to AI without having diagnosed it first carries a silent cost that is paid later. A poorly understood process is not corrected by being automated: it is accelerated with its defects included. A use case chosen without prioritisation criteria consumes the budget and the political capital that the real quick wins deserved. And an adoption without a baseline cannot demonstrate its value, because nobody measured the starting point — so when the inevitable question of what we gained from this arrives, there will be no defensible answer.

Applying the methodology first inverts that equation. The prior diagnosis reveals which processes are genuinely ready and which need order before algorithms. Early governance — the great merit of Anthropic’s approach — turns legal, risk and security into enablers of the project instead of obstacles discovered halfway through. Structured discovery — the great merit of OpenAI’s approach — ensures the use cases are born from those who live the process, which is the only reliable source of real adoption. And measuring from day zero turns every pilot into evidence: the organisation learns what works in its own context, not in somebody else’s success story.

My conviction, after testing both frameworks against the reality of companies in our region, is that the order of the factors does alter the product. The organisation that spends a few weeks gathering information with method — its strategy, its processes, its data, its people — before touching the technology gets further, faster and with fewer shocks than the one that starts with the tool. The Anthropic and OpenAI methodologies are published, they are rigorous, and they complement each other. The best moment to read them is before the first pilot. The second best moment is today.

References

1. Anthropic. The Enterprise AI Transformation Guide. resources.anthropic.com/enterprise-ai-transformation-guide
2. Anthropic. Building Trusted AI in the Enterprise (guide with AWS, PDF). www-cdn.anthropic.com
3. Anthropic. The Enterprise AI Transformation Guide for Retail (PDF). resources.anthropic.com
4. Anthropic. The Enterprise AI Transformation Guide for Healthcare & Life Sciences. resources.anthropic.com/hcls-transformation-guide
5. Anthropic Academy. Driving Enterprise Adoption of Claude (course). anthropic.skilljar.com
6. OpenAI. AI in the Enterprise: Lessons from Seven Frontier Companies (PDF). cdn.openai.com
7. OpenAI. Identifying and Scaling AI Use Cases (PDF). cdn.openai.com
8. OpenAI. Identifying and Scaling AI Use Cases (web version). openai.com