Everyone Automated. Six Per Cent Got Paid.

Because most deployments put a model next to an existing process instead of rebuilding the process

McKinsey's most recent global survey on the state of AI was published on 25 August 2026 and fielded between 4 May and 8 June 2026, with 1,719 respondents across 97 nations, 36 per cent of them at organisations with over one billion dollars in annual revenue. The adoption figures are unambiguous: nearly 90 per cent use AI regularly in at least one business function, 56 per cent use it in three or more functions, up from 51 per cent, and 44 per cent describe it as scaling across the enterprise, up from 38 per cent.

The return figures are equally unambiguous and they point the other way. Thirty seven per cent attribute at least some EBIT impact to AI, which is essentially where the number sat a year earlier. The high performer group, those attributing at least 5 per cent of EBIT to AI with significant impact, is about 6 per cent of respondents, also flat. Meanwhile 80 per cent of individuals say AI has improved their own productivity. Personal productivity has risen. Enterprise profit has not followed it.

The survey offers a clean explanation. Nearly three quarters of the high performers have fundamentally redesigned workflows around AI, up from 55 per cent a year earlier. Among everyone else, the figure is about a quarter. The organisations getting paid are not using better models. They are running different processes.

This is the distinction that most energy marketing teams have not yet made. Asking a model to draft a case study faster is not workflow redesign; it is the same workflow with a faster first draft and an unchanged review queue, which is where the time was actually going. Workflow redesign means changing who does what, in what order, with what approval, and accepting that some steps disappear. It is an operating model change, and it is uncomfortable, which is why so few have done it.

Projet 54A rope access technician working on a storage tank. The research layer around a purchase is automating quickly. The verification layer, where a specification has to be defensible, is not.

To narrow the field early, and never to settle the decision

Gartner presented survey findings at its 2026 CSO and Sales Leader Conference based on 645 B2B buyers. Forty five per cent used generative AI during a recent purchase, mainly to research vendors and products. Sixty nine per cent prefer to check AI generated insights with a sales representative. Robert Blaisdell, VP Analyst and Chief of Research in the Gartner sales practice, put it plainly: buyers still turn to sales reps to validate AI generated insights.

Two further numbers from the same survey complicate the obvious reading. Sixty seven per cent of buyers say they would prefer a buying experience without a sales representative, and 70 per cent prefer a fully digital self service experience. So the same population wants no seller and also wants a seller to confirm what the machine told them. That is not a contradiction. It describes a sequence. The research phase is being automated by the buyer, and the verification phase is being protected by the buyer.

The credibility numbers explain why. Fifty one per cent of buyers believe they are more likely to encounter misleading information from generative AI, while 49 per cent say the same about sales representatives. Buyers trust neither channel more than the other, so they triangulate, using an average of seven information sources per purchase. In an energy purchase, where a wrong specification means a plant modification rather than a return, that instinct is rational and it is not going away.

The consequence for a supplier's marketing operation is specific. The material a model reads when a buyer asks it about your category is now part of the research phase, and it must be accurate, structured and retrievable. The material a buyer uses to verify, the specification sheet, the delivery commitment, the reference, is part of the verification phase, and it must be defensible by a named human. Automating the first is leverage. Automating the second is a liability with a delay fuse.

Ranked by volume multiplied by tolerance for error, not by vendor enthusiasm

The ordering principle is simple and it is the opposite of how most pilots are chosen. Do not start with the most visible workflow. Start with the one that is high volume, repetitive, and where a mistake is caught internally before it reaches a buyer. That is where redesign is safe enough to be real, and real enough to produce the EBIT line that 63 per cent of organisations currently cannot show.

Account and market research synthesis comes first. An energy salesperson preparing for a meeting with a national oil company or an EPC contractor is reading filings, tenders, local content rules and trade press. The volume is high, the output is internal, and the cost of an error is a corrected briefing note. This is the single highest ratio of hours saved to risk taken in the function.

First draft production for non binding content is second. Blog explainers, event abstracts, newsletter copy, social posts and internal enablement material all carry no contractual weight. The redesign that matters here is not the drafting, it is removing a review stage: if the content cannot create an obligation, it does not need legal review, and the queue that was the real bottleneck disappears.

Translation and localisation is third, and it is chronically underrated in energy because the buying centres are genuinely multilingual. Published English material translated into the working languages of Gulf, North African and Francophone West African procurement teams expands reach at near zero marginal cost, provided the source material is accurate. The risk profile is inherited: a correct source translates to a correct target.

Enquiry enrichment and routing is fourth. Matching an inbound enquiry to a company, a sector, a probable project and an owner is pattern work with a human fallback. The failure mode, a misrouted enquiry, is recoverable within hours.

The monthly reporting pack is fifth, and it is the one most teams skip because it feels like overhead rather than growth. It is also the workflow whose automation produces the evidence needed to justify the other four. Pull the numbers, assemble the narrative, flag the anomalies, and have a human decide what the anomalies mean.

Buy the workflow change. Be sceptical of the category

Gartner's position on the agent market is worth reading before any procurement conversation. It predicts that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It estimates that only around 130 of the thousands of vendors claiming agentic capability are real, and describes widespread agent washing: existing chatbots and robotic process automation rebranded without substantive change.

Anushree Verma, Senior Director Analyst at Gartner, framed the maturity accurately: most agentic AI projects are early stage experiments or proofs of concept, most agentic AI propositions lack significant value or return on investment, and many use cases positioned as agentic today do not require agentic implementations. That last point is the one to carry into a vendor meeting. A workflow that runs on a schedule, follows a fixed sequence and has a deterministic output does not need an agent. It needs automation, which is cheaper, more predictable and easier to audit.

The same Gartner material does project real change: at least 15 per cent of day to day work decisions made autonomously through agentic AI by 2028, from zero in 2024, and 33 per cent of enterprise software applications including agentic AI by 2028, from under 1 per cent in 2024. The direction is not in doubt. The timing of value capture is. In a January 2025 poll of 3,412 attendees, 19 per cent reported significant agentic AI investment, 42 per cent conservative investment, 8 per cent none, and 31 per cent were waiting or unsure. Conservative is the modal position and it has aged well.

For an energy supplier the pragmatic stance is this. Spend the budget on redesigning the five workflows above using tools you already own or can rent cheaply, measure the hours and the cycle time before and after, and defer agent procurement until a specific workflow has a demonstrated bottleneck that only autonomy removes. McKinsey's figure that around 20 per cent of organisations now say operating costs including tokens constrain their AI use is a reminder that running cost is a real line item, not a rounding error.

Measure the queue, not the output

The most common measurement mistake is to count artefacts produced. More drafts, more posts, more briefings. That number always rises and it never correlates with anything commercial, which is how a function arrives at the 37 per cent position: visible activity, invisible margin.

Measure three things instead. First, cycle time from request to published or delivered, because that is what redesign actually compresses. Second, the number of approval stages a given artefact passes through, because removing a stage is the redesign and keeping it is the reason nothing changed. Third, reallocated hours, specifically whether the hours saved moved to work that a buyer sees, or simply evaporated into more internal review.

Then connect it to the buyer evidence. If buyers consult an average of seven sources and 45 per cent of them start with a generative tool, the commercial test of a redesigned content operation is whether your material is present and accurate in that research phase, and whether a named human is available fast when the buyer moves to verification. Those two conditions are measurable, and together they are a better proxy for pipeline effect than any volume metric.

A final caution on headcount. Thirty nine per cent of McKinsey's respondents expect AI to reduce their organisation's headcount in the coming year, up from 32 per cent. Only 14 per cent reported an actual AI related workforce decline over the past year, against the 32 per cent who had expected one the year before. Expectation has been running roughly twice ahead of outcome. Plan the workflow change on its own merits, and treat any business case that depends on a headcount reduction arriving on schedule as unfunded.