Two numbers this fortnight tell the same story. For the first time, the world will spend more on running AI than on training it — $23.3 billion against $19 billion. And in a Gartner poll of 743 audit professionals, 93% of audit leaders use AI but only 38% have a strategy for it. AI has quietly moved into everyday work while the plan stayed in the drawer. The Gulf is moving faster still: the UAE has started shifting half of federal government work onto AI agents, and Saudi data-centre capacity has gone from 68 megawatts to more than 467 in five years. So this issue is about the unglamorous part that decides who actually gets value — naming the number you want to move, picking two workflows, and saying who is accountable. Inside: the news, a one-page AI plan you can write this afternoon, five checks before an agent touches a real process, and our new playbook.
What's moving in AI worldwide
The world will spend more running AI than building it
For the first time, more money goes on running AI every day than on training the models. That is what it looks like when AI stops being a project and becomes part of normal operations.
$23.3B — global spend on running AI models in 2026 — more than the $19B spent training them (Gartner, August 2026)
There is a quiet milestone in Gartner's August forecast. In 2026, global spending on inference — the compute you burn every time a model actually does something — hits $23.3 billion and overtakes the $19 billion spent training models. Fifty-five percent of all AI-optimised cloud spend now goes on running AI rather than building it, and Gartner expects that to reach 59% next year.
The wider number is just as telling. Spending on AI-optimised infrastructure as a service is forecast at $42.3 billion in 2026, up 96% on last year, and $66.1 billion in 2027. Gartner's analyst Hardeep Singh puts it down to two things: fine-tuned, company-specific models moving into customer-facing systems that run continuously, and agentic AI, which is far hungrier than a chatbot because it executes multiple steps on its own.
For a leader, the practical read is about budgets. A pilot has a start and an end. An agent in production has a meter. If you are approving AI work this quarter, ask what it costs to run for a year at full volume, not what it costs to build — and make sure someone owns that number.
Why it matters: Running costs are becoming a permanent line in the budget, not a one-off project cost. Every agent you switch on adds to it, and it keeps ticking long after the pilot is signed off.
Source: Gartner — Forecast: AI-Optimized IaaS, Worldwide
Nearly everyone is using AI. Far fewer have a plan for it.
93% of audit leaders use AI in some form. Only 38% have an AI strategy. That gap is not an audit problem — it shows up in almost every function.
38% — of audit leaders have an AI strategy, though 93% already use AI (Gartner poll of 743 audit professionals, August 2026)
Gartner's August poll of 743 audit professionals produced two numbers that belong side by side. Ninety-three percent of audit leaders report some level of AI use. Thirty-eight percent have an AI strategy. In other words, almost everyone is using it and most are making it up as they go.
Look at where the use sits and the picture sharpens. Sixty percent use AI to draft audit issues, ratings or reports; 41% to review drafts; 35% to prepare stakeholder communications; 37% for general productivity like email and translation. The harder, higher-value work barely registers — 30% for audit testing, 12% for quality assurance reviews. Gartner's James Bourke is blunt about the result: adoption is high, but it is not transforming how the work gets done.
Audit is simply the function that got measured. The same shape shows up in HR, finance, marketing and operations — dozens of people quietly speeding up their own drafting, nobody accountable for whether it adds up to anything. And there is a regional twist worth noting. PwC's 29th Global CEO Survey found 70% of Middle East CEOs say they have a clearly defined AI roadmap, against 51% globally. So in the Gulf the question is not whether there is a plan. It is whether the plan names a number, an owner and a workflow — or just an ambition.
Why it matters: Adoption is not the same as value. Spread thinly across small tasks with nobody accountable, AI gives you slightly faster drafting and nothing that shows up in the numbers.
Source: Gartner — Survey: Audit Teams' AI Use is Common, But Most Lack Strategy
AI in the region: UAE & KSA
The UAE is putting half of federal government work on AI agents
The UAE has started the executive phase of its plan to move 50% of federal government operations, services and tasks onto AI agents within two years.
50% — of UAE federal government operations, services and tasks targeted to run on AI agents within two years (WAM (Emirates News Agency), August 2026)
The UAE's agentic AI project has moved from framework to execution. In August, WAM reported that more than 100 federal officials had attended the workshop launching the project's strategic track, run by the National Committee for the Agentic AI Project. The target it serves — announced in June by Minister of Cabinet Affairs Mohammad Al Gergawi — is to convert 50% of federal government operations, services and tasks to agentic AI models within two years.
What makes this track different is where the agents go. Most government AI programmes start with transactions: renewals, permits, help desks. This one targets policy formulation, future foresight and decision-making efficiency under the UAE Government 4.0 framework, across seven pillars including strategy, governance, government performance and competitiveness. Huda Al Hashimi, Deputy Minister of Cabinet Affairs for Strategic Affairs, described the session as the operational start of deployment. The guiding principle is stated plainly: human leads, AI enables.
The workforce side is running alongside it. The June announcement included a commitment to train 80,000 federal government employees on agentic AI tools. That is the part most private-sector plans skip — and it is the reason government may end up more agent-ready than the companies that supply it.
Why it matters: If you sell to, are licensed by, or are inspected by UAE government, your counterparty is increasingly going to be software. Procurement, permits and audits will move faster and expect cleaner, structured data from your side.
What to do: Take the two or three processes you run most often with government — licensing, WPS payroll filings, tax, tenders — and check whether your side can keep pace without a person retyping things. Then name one person accountable for AI governance, before a regulator's agent asks who signed off.
Source: WAM — UAE advances integration of agentic AI across federal government
Saudi Arabia has built the place where all that AI will run
Saudi data-centre capacity has gone from 68 megawatts in 2021 to more than 467 megawatts in early 2026 — close to sevenfold in five years.
467 MW — Saudi data-centre capacity in Q1 2026, up from 68 MW in 2021 (Saudi Press Agency, August 2026 (capacity series per Saudi MCIT))
If the global story is that money is moving to running AI, the Gulf story is where that running happens. Saudi Arabia's operational data-centre capacity has gone from 68 megawatts in 2021 to more than 467 megawatts in the first quarter of 2026. The Ministry of Communications and Information Technology puts the intermediate figure at 440MW for 2025, so the curve is still steepening. SPA reported in August that the build-out is supported by investment exceeding SAR 56.2 billion in data centres and digital infrastructure.
The supporting numbers point the same way. More than 60 data centres now operate across the Kingdom, MCIT ranks Saudi Arabia second globally for data-centre market attractiveness, and digital infrastructure reaches close to the whole population. The same August read-out flagged an SAR 11 billion partnership between HUMAIN and Blackstone.
For employers, the practical consequence is about where data is allowed to sit. Plenty of AI projects in banking, healthcare and government supply chains have been parked because the data could not leave the country. That constraint is loosening. What replaces it is a people constraint: megawatts need operators, data engineers, network and security staff — and every organisation in the Kingdom will be hiring for the same roles at the same time.
Why it matters: This is the supply side of the story above. The compute that runs AI is landing inside the Kingdom, which makes local hosting and data residency realistic for regulated industries that had to park projects before.
What to do: If you shelved an AI project because the data couldn't leave the country, price it again — the answer may have changed. And plan the people early: capacity needs operators, data engineers and security staff, and everyone will be hiring them at once.
AI agents & your workforce
How to write your team's AI plan on one page
You do not need a strategy document. You need one page that names a number, two workflows, an owner and a review date. Here is how to write it in an afternoon.
Five steps to a one-page AI plan
- Name the number you want to move. Pick one: days to hire, cost per invoice, time to close, error rate. If your plan doesn't name a number, you can't tell later whether it worked.
- Pick two workflows, not twenty. Choose two high-volume processes that touch that number, and go deep on them. Twenty shallow pilots produce twenty demos and no result.
- Say who is accountable. One named person owns the number and the risk — not a committee, not IT by default. Ambiguous ownership is the most common reason AI work stalls.
- Write the guardrails before go-live. Four lines: what the AI may decide alone, what a person must approve, what data it may see, and how you'll know if it goes wrong.
- Budget for people, not licences. Most of the effort is process change and training, not software. Put your training days and your run cost on the same page as the tool cost.
One page, one number, one owner, two workflows. Set a 90-day review in the calendar before you close the document — that review is what turns a plan into a habit.
Five checks before you let an AI agent touch a real process
Once an agent takes actions instead of making suggestions, the questions change. These five take about an hour to answer and will save you a bad week later.
Five checks to run before an agent goes live
- Can you see what it did?. Every action it takes should land in a log a non-technical manager can read. If you can't reconstruct a decision, you can't defend it.
- How far can it reach?. List exactly which systems, folders and records it can touch. Start narrow. Widening access later is easy; explaining a mass email is not.
- Who signs the risky ones?. Draw the line between what it does alone and what needs a human yes — anything that leaves the company, moves money or affects a person's job.
- What does it do when it's unsure?. The correct answer is stop and ask, not guess. Test this deliberately with a messy, incomplete case before you trust it with a real one.
- Who turns it off?. Name the person and the method, and check it works. An agent nobody can stop at 9pm on a Thursday is not in production — it's at large.
Run all five in an hour. If you can't answer one of them, that gap is your next piece of work — and it is far cheaper to fix now than after go-live.
The One-Page AI Plan: Turning AI Use Into AI Value in the Gulf
Playbook. Almost every team in the UAE and KSA is now using AI. Far fewer can say what it is for. This playbook is the fix: six steps that take a leader from scattered AI use to one page that names the number you want to move, the two workflows you'll change, the person accountable, the guardrails, the money and the people. Written for C-suite and directors in mid-to-large Gulf enterprises who want a plan they can actually hand to their team on Monday.
Spotlight: AI-Assisted Hiring with Hyrra.ai
Hyrra.ai. Hyrra.ai puts AI agents to work on the CV pile — automating first-pass screening and cutting time-to-shortlist, with a recruiter in control of every shortlist and rejection. It's exactly the pattern this edition argues for: one high-volume workflow, done properly, with a person still owning the decisions.
