After Adoption: Six Analysts on the GenAI Operating Model

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

Eighteen months ago the question every board was asking was whether to adopt generative AI. The slide in the deck said pilot, prove, scale. The budget line said training and licences. By late 2026 that question has shed its interest: adoption is happening, and in most mid-market and enterprise organisations the only remaining debate is whether to pay for it on seat licences or on consumption. The real question has shifted underneath, and most leadership teams have not noticed.

What a board should be asking in a 2027 planning cycle is not whether the pilots worked. It is whether anything got built around the pilots that would still be running under a new chief financial officer. An operating model: who owns the agents, which process they touch, what metric each decision gets held against, where the data lives, how a new use case gets through legal and into the stack without six months of standing-committee review. These are the parts nobody budgeted. They are not interesting to show a board and they are tedious to build, and that is where the next two years of value is sitting.

I went and read what the six analysts you would hire to tell you this have published in 2026, and the consensus is tighter than I expected. They disagree on who should wear the title and on which single number proves it worked, but they agree on the diagnosis. The pilots are not the problem anymore. The silence around them is.

The Six Voices on This Question

• Sylvain Duranton, Managing Director and Senior Partner, Global Leader of BCG X, Boston Consulting Group.
• Vinciane Beauchene, Managing Director and Partner, Boston Consulting Group.
• Aditya Challapally, lead author of the MIT NANDA “State of AI in Business” report.
• Rita Sallam, Distinguished VP Analyst, Gartner.
• Nitin Mittal, Global AI Leader, Deloitte.
• Jim Rowan, US Head of AI, Deloitte Consulting LLP.

Strategy, Not Tools, Is What Sorts Winners from Losers

Boston Consulting Group spent the first half of 2026 arguing this point in print. The firm’s annual AI at Work study, released on 3 June 2026, covering twelve thousand frontline workers across fourteen markets, landed on a finding that cuts through the usual adoption chart. Sixty-six percent of those workers report receiving limited or no guidance on how to use the time that AI saves them. The adoption number is no longer the problem. The absence of a reason the time is being saved is.

Sylvain Duranton returned to the finding a month later in a Forbes column co-signed with BCG colleagues on 6 July 2026. His framing is worth printing straight. More than ever, he writes, employees expect their leaders to focus the workforce’s AI efforts on a few high-impact topics, and to rethink the organization and ways of working. The sentence does three useful things at once. It names the audience, which is employees, not shareholders. It admits that the leadership has not focused the effort. And it asks for organisational redesign, not for another training programme.

Employees expect their leaders to focus the workforce’s AI efforts on a few high-impact topics, and to rethink the organization and ways of working.

Sylvain Duranton, BCG, 6 July 2026

Vinciane Beauchene, co-author of the June report, is more pointed about the shape of the gap. The pattern across the twelve thousand workers is that enthusiasm fades into disengagement where no strategy follows the rollout; a workforce that was given the tools but not the steer concludes within a quarter that the exercise was never serious. Any agency that has lived through a “we bought the seats, now what” conversation with a client will recognise that pattern on sight.

Adoption Is a Bridge Metric, Not an Outcome

MIT NANDA’s “GenAI Divide” report, led by Aditya Challapally and published mid-2025, carries the figure that every 2026 analyst piece has been built against. Across the sample of leadership interviews, employee surveys and public deployments, ninety-five percent of surveyed organisations are getting no measurable return from their generative AI initiatives. The report’s second number is the one worth keeping: of custom enterprise AI tools built in the sample, only five percent have reached production.

Challapally’s own framing of why that five percent works, reported in Fortune’s coverage of the research, is that the surviving programmes pick one pain point, execute well, and partner smartly with the companies who use their tools. Nothing in that sentence is about the model. It is about scope, operational discipline, and organisational readiness to use the thing. The implication for a 2027 operating model is direct: a programme that cannot name one pain point, one partner, and one owner should not be in the next budget.

Leadership Presence Is Where Every 2026 Report Lands

Deloitte’s 2026 State of AI in the Enterprise, surveying 3,235 leaders between August and September 2025 across 24 countries, is the broadest of the six voices. Its headline finding, in the report’s own words, is that enterprises where senior leadership actively shapes AI governance achieve significantly greater business value than those where governance sits inside a committee. Nitin Mittal, Global AI Leader at Deloitte, and Jim Rowan, US Head of AI at Deloitte Consulting LLP, are both named on the report’s contact line.

The report avoids a single attributed quote and does the arithmetic instead. Thirty-four percent of surveyed organisations are deeply transforming operations around AI, thirty percent are redesigning key processes, and thirty-seven percent remain at what the authors call a surface level of use. The line that separates the first two groups from the third is the presence of a named executive who signs the operating model.

Read against the BCG work, this is the clearest point of agreement in the six voices. A programme without a named executive sponsor who turns up to the quarterly review is a programme that stays in the thirty-seven percent.

Measurement Is the Honest Blocker, and Everyone on This List Admits It

Gartner’s position is the one every executive in the room will quietly admit to their own team. Rita Sallam, Distinguished VP Analyst at Gartner, presented the firm’s top data and analytics predictions for 2026 at the Orlando summit on 11 March 2026. The thread running through those predictions is that measurement, not model capability, decides whether an enterprise AI programme survives its first finance review. Sallam’s 2024 call that at least thirty percent of generative AI projects would be abandoned after proof of concept by the end of 2025 has dated well. Her 2026 message is that the survivors are the ones who wrote down a measurable business outcome before the tool was switched on.

Deloitte’s own 2026 numbers reinforce this. Twenty percent of surveyed organisations say AI has already increased revenue. Seventy-four percent say they hope it will. Hope is not an operating-model metric. A programme that cannot name the number it has moved, with a baseline and a date, cannot be extended through a change at the finance function. That is not a style preference. It is how an operating model survives.

Skills Are a Harder Line Item Than Licences

Deloitte’s report calls insufficient worker skills the biggest barrier to integrating AI into existing workflows. Fifty-three percent of surveyed enterprises are educating the broader workforce on AI and forty-eight percent are running upskilling or reskilling programmes. The implication here is that the tool budget was the easy part of the 2025 cycle. Training is paid in time, not in money, and time is harder to approve.

BCG’s sixty-six percent no-guidance figure is the same gap seen from the worker side. Beauchene and Duranton are consistent across their two 2026 pieces on this: strategic clarity, not training hours, is the lever that moves adoption from use into value. An operating model that funds licences but not the hour a week it takes to tell a team which decision the tool is supposed to help them make is an operating model that funds half the work.

Where the Six Disagree

The agreement is on the diagnosis. The disagreement is on ownership. BCG puts responsibility on senior leaders with a direct, named organisational role. Deloitte puts it on a cross-functional governance layer anchored by board visibility. MIT NANDA, by implication, puts it on whoever picks the first pain point and the first partner. Gartner puts it on whoever can prove the number moved. There is no settled answer on who should wear the title of chief AI officer, and the honest reading is that it matters less which function holds the title than that someone does, and that the operating model is written down before the next budget cycle closes. A thing almost nobody has and almost nobody is asking for is a thing that stays unbuilt.

What a WordPress Build Section Looks Like for This Operating Model

This is where we would depart from Econsultancy and from most of the analyst houses. The operating model does not live only on a slide; it lands as work inside the systems your content, commerce and support functions already use. For an enterprise running WordPress multisite or a WooCommerce catalogue, here is what the build actually looks like, written in sprints rather than in strategy bullets.

• Governance baked into the template, not into the editor’s memory, two to three sprints. Instead of training editors to add AI-generated disclosures and machine-readable provenance tags by hand, bake both into the Gutenberg block pattern. A new editorial piece inherits provenance metadata, source attribution and a classification field from the template rather than from editorial discipline. This is Deloitte’s “governance actively shaped” line, implemented on a WordPress stack.
• Centralised AI provider key management, one to two sprints. Route every call to OpenAI, Anthropic and open-source models through a single server-side endpoint that applies per-team budget caps, logs usage, and lets security rotate keys without a plugin audit. Without it, licence sprawl hides the real cost by month three, and nobody can report honest spend to a CFO.
• Measurement baseline before the first use case ships, four to six weeks of instrumentation plus a quarter of observation. For a WooCommerce catalogue that means GA4 custom dimensions on every AI-touched surface (product descriptions, personalised recommendations, support replies), plus server-side logging of which agent ran which task. Sallam’s point in one line: if the dashboard does not exist on day one, the programme will not be extendable at day 120.
• The named owner, named in the mu-plugin, one day. Trivial technical work, hard political work. The content-side equivalent of a CODEOWNERS file: every section of the site has an editorial owner whose name shows up in the admin, and every AI-augmented workflow has a business owner whose approval is required to change the prompt template. This is Challapally’s “one pain point, one partner, one owner” principle, rendered as configuration.
• Quarterly test, fix, retest cadence, baked into the sprint calendar, ongoing. Not an annual audit. The right beat is ninety days, which is what WordPress VIP itself runs on its own properties, and which matches our own view after running that cadence on thewpclan.com.

For a mid-market multi-brand estate, the first three items are eight to ten weeks of concentrated build, plus one quarter of measurement before the operating model has anything honest to show a board. The quarter is not optional, and it is the part that gets cut. That is the main reason most operating models fail at the second budget review.

What a Board Should Ask in the 2027 Planning Cycle

If I were briefing a board for a 2027 cycle on this, three questions would get asked before any new budget line was approved. Who owns this programme inside this company, by name, at executive level, with a seat on the quarterly review. Which single business metric will move if this operating model works, and what is its current baseline and its target. What is the governance decision that survives a change at the chief financial officer. The answers can be short. What cannot be short is the honest answer to each. If any of the three come back as a function or a committee rather than a person and a number, the operating model does not exist yet, and funding more pilots will not bring it into being.

We build and run WordPress and WooCommerce estates for organisations that have done the first eighteen months of AI adoption and are now sitting with the second question. If that is a conversation worth having before your 2027 planning closes, that is what the agency is here for. A second pair of eyes before a budget signs is usually cheaper than a second eighteen months of pilots.