McKinsey finds that AI’s largest gains come from faster, better decisions. Its new measure, decision throughput, tracks the share of your decisions that AI informs, speeds up or makes.
It’s Thursday evening. You have 3 quotes to price, a customer who is 20 days past terms and a decision to make about replacing the van. Your September figures will reach you at the end of October.
Each of those decisions is waiting for you. Every day it waits has a cost. None of that cost appears in your accounts.
In August, McKinsey published a paper by Lari Hämäläinen and Steffen Fuchs called The decision dividend: How AI creates economic value. They argue that AI’s largest gains come from faster decisions, better use of existing assets and opportunities a business would otherwise miss. They propose a measure to track those gains. They call it decision throughput: the share of decisions that AI informs, speeds up or makes.
The paper speaks to chief executives of large companies. Its argument applies directly to an owner-managed business, where most decisions pass through one person.
In a small business, the return from AI depends on how far you redesign the way decisions get made.
1. The cost of deciding
McKinsey starts with a gap in what companies measure. Most know exactly what they spend on labour, materials and equipment. Few can put a figure on what they spend making decisions.
The cost is hard to see because it sits inside ordinary work. Managers, analysts and frontline staff spend hours gathering information, comparing options, checking with colleagues and waiting for approval. As a company grows, meetings and review cycles multiply. McKinsey calls decision-making one of the largest operating costs a company carries, and one of the least visible.
In a small business the structure is simpler. You are the approval process. Most decisions about prices, people, purchases and cash pass through you, because you built the business and for years made every decision yourself.
So each decision carries 2 costs. The first is your time. The second is the wait between the moment the information exists and the moment someone acts on it.
The wait is easy to underestimate. A 5% price rise in a business turning over £1m adds £50,000 a year, if volumes hold. Delay it by 6 months and £25,000 is gone. No line in the profit and loss account records the loss.
McKinsey makes the same point at corporate scale. A $50 million initiative postponed by 4 quarters can become a $75 million to $100 million problem. The business loses a year of learning, improvement and financial return.
2. The falling price of a decision
McKinsey cites research estimating that AI can cut the cost of many decisions by more than 90%. A task that needed about $40 of human labour in 2023 is now estimated to cost a fraction of a cent. Decisions that took weeks can take seconds.
The paper’s example is Klarna, the Swedish payments company. Using AI, Klarna cut the time to resolve a customer service query from 11 minutes to under 2. It estimated the saving at $40 million for 2024.
Those ratios hold for simple, single-step decisions. Some agent-based decisions cost 100 to 1,000 times more, because the AI checks and reworks its own output. McKinsey says they still cost less than doing the work by hand.
The speed of the fall is new. McKinsey compares how long major technologies took to match the cost of what they replaced. The steam engine took about 55 years, the electric dynamo about 30 and the internet about 7. On McKinsey’s estimate, AI decision-making took about 2, though few companies have yet made it work at scale.
Part of the reason is that users rent the infrastructure. The 4 largest cloud and data centre companies are expected to spend more than $700 billion on capital equipment in 2026. Everyone else pays by use. A business with 8 employees can use the same AI models as a FTSE 100 company, for the price of a few monthly subscriptions.
3. Two ways to adopt AI
Most companies still report little or no measurable effect on earnings. McKinsey traces the gap to how they put AI to work.
The typical company adds copilots, chatbots and dashboards to the processes it already runs. McKinsey calls these companies bolt-on adopters. Their returns are modest: under $1 for each $1 invested, with no clear payback period.
A smaller group redesigns whole workflows around AI. A workflow is the sequence of steps from information to action. Redesigning it changes who decides, when, and on what information.
These companies report about $3 of profit for each $1 invested and payback within 1 to 2 years. They often see EBITDA rise by roughly 20%. (EBITDA is profit before interest, tax, depreciation and amortisation.)
Treat those figures with care. They are self-reported by the firms McKinsey classes as AI high performers, about 6% of the 1,993 respondents to its November 2025 global survey. Plan on less.
The economist Paul David found the same pattern when American factories electrified, and McKinsey cites his work. Early adopters fitted electric motors to the belts and shafts the steam engine had turned. In 1899, electric motors supplied less than 5% of the mechanical drive in US factories.
The productivity surge came in the 1920s, 4 decades after the first central power station opened. It arrived when owners rebuilt their factories around a separate motor on each machine, in single-storey buildings laid out to follow the work. Electric power paid off once the factory itself changed.
The return on a new technology arrives when the work is redesigned around it.
4. Where the money comes from
A labour saving has a ceiling. It can never be larger than the cost of the work it removes.
Take a business turning over £1m that spends £60,000 a year on administrative wages. If AI removed half that work and you cut the hours to match, you would save £30,000, or 3% of turnover. The price rise delayed by 6 months cost £25,000 on its own.
McKinsey reaches the same conclusion for large companies. In most of them, administrative costs are too small a share of revenue to explain the gains the leaders report. At some industrial companies, it adds, the largest gains came from other sources while headcount stayed level.
The strongest evidence comes from small firms. In an INSEAD working paper, Hyunjin Kim, Rembrand Koning and Dahyeon Kim report a field experiment with 515 high-growth start-ups. Firms in the treatment group were shown how other businesses had reorganised production around AI.
Those firms found 44% more uses for AI, mostly in product development and strategy. They were 18% more likely to win paying customers and generated 1.9 times the revenue of the control group. They needed 39.5% less outside investment. Their demand for staff did not change.
The gains were largest among the strongest firms.
McKinsey’s Klarna figures carry a footnote: the numbers may have changed since 2024. In May 2025, Klarna’s chief executive, Sebastian Siemiatkowski, told Bloomberg the company was hiring human customer service staff again. He said cost had carried too much weight in how the AI service was judged, and quality had fallen. Klarna’s answer was to invest again in human support.
When an AI service is judged on labour cost alone, the design favours the cheapest way to close each query.
5. Three sources of value, at your size
McKinsey names 3 sources of value: faster action, better use of existing assets, and opportunities that would otherwise be missed. They tend to arrive in that order. Each makes the next possible, and the gains compound.
On faster action, McKinsey adds a condition. Companies must also shorten their planning cycles and workflows, which tend to be long.
In an owner-managed business, the planning cycle is usually the month-end or the year-end. If your management accounts reach you 4 weeks after the month closes, every decision they inform starts at least 4 weeks late. If you see figures only in the annual accounts, some of the events they describe are more than a year old.
The monthly rhythm was set by the cost of producing figures by hand. AI has cut that cost. A weekly view of cash, sales and margin, taken from Xero and checked against limits you have agreed, can reach you every Monday morning.
The first pillar of Beyond the Return is management reporting that supports decisions while they still matter. Decision throughput puts a number on that idea. The redesign is a change of rhythm. You decide each week, on figures that are days old, and you keep the monthly meeting for the choices that need more thought.
The second source is the assets you already own. McKinsey finds that production schedules, maintenance plans and operating settings are often based on outdated assumptions. Equipment delivers less than it could, and companies buy capacity they may not need.
In a small business the assets include cash tied up in debtors and stock, plus vans and machines. Debtors are the clearest case. On annual sales of £1.2m, each 10 days of debtor days ties up about £33,000 of cash, more if you charge VAT.
Added to the current process, AI drafts the chasing emails faster. In a redesigned process, AI reads the aged debtors report each morning and sends reminders on a schedule you have agreed. It flags any customer paying 10 days slower than their usual pattern. At 45 days it stops and puts the account in front of you with a recommendation: call, hold further work or agree a payment plan.
The decision about who to chase moves from monthly to daily. You make only the decisions that need you.
The third source is opportunity. McKinsey lists the signs of slow decisions: launches that slip, stock in the wrong place, capacity given to the wrong work. When decisions are cheap, a company can test more options and act before conditions change.
Reckitt, the UK-based consumer health and hygiene group, worked with McKinsey to apply AI to its pricing, promotion and portfolio decisions. Leaders could model scenarios and test trade-offs before changing anything in the market. Reckitt also redesigned its planning so staff used the analysis in everyday decisions.
In a small business the equivalent is the quote. In many trade and service firms, the owner prices every job personally, often in the evening, from memory. Quotes go out days after the enquiry. Margins vary with what the owner remembers of past jobs.
Added to that process, AI tidies the quote and writes the covering email. The quote still waits for the owner.
In a redesigned process, the last 2 years of jobs sit in one place: estimated cost, actual cost, price, margin, won or lost. AI drafts each price from that history and your written pricing rules. A member of the team sends any quote under £5,000 that meets the target margin. You see the exceptions: large jobs, new types of work and anything below target.
Quotes can go out the same day. The business can also see its win rate and margin by type of job, month by month.
Every redesign shortens the wait between information and action.
6. Counting decisions
Without clear financial measures, many companies count AI activity: licences bought, models deployed, pilots launched. McKinsey says those numbers measure ambition and say little about returns. A company can own many AI tools and use few of them well.
Manufacturers learned long ago to judge a factory by its output. McKinsey applies the same test to AI. That is the purpose of decision throughput, the percentage of decisions that AI informs, speeds up or makes.
McKinsey’s example is a manufacturer that resets its production schedule once a week. If it checks demand, stock and equipment every hour, it can reset the schedule several times a day. McKinsey treats rising decision throughput as an early sign of financial results to come.
In a small business the measure fits on 2 sides of A4. List the decisions that recur each week, month and quarter. Pricing a job, chasing a debt, reordering stock, approving overtime and deciding when to hire all belong on it.
Mark each decision that AI informs, speeds up or makes within limits you have set. Divide the marked decisions by the total. If 5 of 30 are marked, your decision throughput is 17%. Track it each quarter.
The count alone can mislead. A business could automate 20 decisions about sorting email, leave pricing untouched and post a high score with no change in profit. McKinsey’s version of the warning is that each decision must create more value than it costs to make.
Start with the decisions that move profit and cash. Track the results in money and days: debtor days, quote turnaround, margin by type of job and the hours you spend deciding each week.
A companion McKinsey article, Is that AI agent worth it? Agentic economics and the modern operating model, adds a warning on cost. The same AI task can cost 30 times more depending on how it is run. The article recommends judging AI work by the business result it completes.
Every decision on the list should be tested against 3 questions. How often does it come up? What does a late or wrong call cost? Can you write down the rule you use?
Decision throughput moves before the profit does.
7. Write the rule before you choose the tool
Is there a better way to make this decision? That question comes before any choice of software. McKinsey expects the companies that move first to redesign their workflows before they select tools.
The INSEAD researchers name the difficulty the mapping problem: finding where and how AI creates value in the way a business works. They conclude that this search is a key constraint on the gains from AI.
In a small business the redesign starts on paper. You can hand a decision to AI, or to a member of your team, only once someone has described it. In many owner-managed businesses the rules for pricing, credit and purchasing exist only in the owner’s head. They were never written down because, for years, the owner made every one of those decisions.
Writing the rule down is the redesign. It sets out who decides, on what information, by when and within what limits. The same written rule lets a member of your team make the decision, with or without AI.
The companion article sets a condition for every autonomous AI system: a defined mandate, a budget and a stopping rule. For a small business that means 3 written answers for each decision you hand over.
What may it decide? Up to what value? When does it stop and ask you? The credit control process above has all 3.
McKinsey adds a fourth: who is accountable for the result. In a business of 10 people, that is usually you.
The companion article also argues that data and context now separate one company’s results from another’s, because the AI models themselves are converging. The same AI agent gives very different results depending on the customer information, operating data and decision history it receives.
Your context is your job history, your Xero records and your notes on why you priced a job the way you did. Any competitor can buy the same AI subscription. None of them has your last 3 years of job costings.
AI working from poor Xero records produces wrong answers faster.
The article’s advice is to start capturing that context now. In a small business, a short decision log will do: the price of each quote, the reason for it and the outcome. In 12 months it gives an AI tool the history it needs to price your work.
The same article notes that process improvements which once took years can now be matched in months or weeks. Records and judgement take years to build. I take that argument further in The Knowledge Your Competitors Can’t Buy.
Not every decision needs AI. The article warns against assuming every problem needs the most capable model. Smaller models, fixed rules and ordinary software often do the job. A bank rule in Xero settles a recurring coding decision at no extra cost.
A decision you cannot describe is a decision you cannot delegate.
8. Where to start
McKinsey’s advice to chief executives fits a small business with few changes.
Start where decisions matter most, and prove the gain in 1 or 2 areas before expanding. Keep experiments cheap so you learn quickly. Define who is accountable, track economic results and treat delay as a cost.
Proving the gain in one area first also lowers the cost of the next. At Freeport-McMoRan, the mining company, about 60% of the code written for the first site was reused at later sites. Each deployment ran faster than the last.
The small business version needs a pad of paper and 90 days.
List the decisions that recur each week, month and quarter.
Note how often each one comes up, who makes it and how long it waits. Estimate what a late or wrong call costs.
Choose the 1 or 2 decisions with the highest cost of delay or error.
Write the redesign: who decides, on what information, by when, up to what value, and when you step in.
Choose the tool last. It may be an AI assistant, a Xero rule or a spreadsheet.
Run the new process for 90 days. Measure the result in money and days.
Keep what works, drop what doesn’t, then move to the next decision.
For most owner-managed businesses I would start with pricing or credit control. Both move profit or cash within weeks.
The first redesign is the slowest. The written rules, the cleaner records and the habit of measuring carry over to the next one.
A small business can run this cycle faster than a large company. It has one approver, a handful of systems and no committee to persuade. The randomised trial McKinsey cites was run with small firms.
9. What this means for your business
AI has made a decision cheap to produce. The return depends on how many of your important decisions it reaches, and how fast the business acts on them.
In an owner-managed business, that starts with you. Write down how the business decides. Let the system and the team decide within those limits. Measure the result in money and days.
That is workflow redesign at the scale of an owner-managed business. It produces decisions made on time, on current figures, by the right person or system.
The return on AI starts with the decisions waiting on your desk.
Sources and Further Reading
Lari Hämäläinen and Steffen Fuchs, “The decision dividend: How AI creates economic value”, McKinsey & Company, August 2026, mckinsey.com/industries/industrials/our-insights/the-decision-dividend-how-ai-creates-economic-value.
Lari Hämäläinen, Mark Patel, Sven Blumberg, Tanguy Catlin and Wasim Lala, “Is that AI agent worth it? Agentic economics and the modern operating model”, McKinsey Quarterly, 13 July 2026, mckinsey.com/capabilities/quantumblack/our-insights/is-that-ai-agent-worth-it-agentic-economics-and-the-modern-operating-model.
Hyunjin Kim, Rembrand Koning and Dahyeon Kim, “Mapping AI into Production: A Field Experiment on Firm Performance”, INSEAD Working Paper 2026/20/STR, March 2026, insead.edu/faculty-research/publications/working-papers/mapping-ai-production-a-field-experiment-firm.
McKinsey Global Survey on the state of AI, November 2025 (1,993 respondents in 105 countries), as cited in “The decision dividend”.
Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, American Economic Review, vol. 80, no. 2 (Papers and Proceedings), May 1990, pp. 355–361.
Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month”, press release, 27 February 2024, as cited in “The decision dividend”.
Irina Ivanova, “As Klarna flips from AI-first to hiring people again, a new landmark survey reveals most AI projects fail to deliver”, Fortune, 9 May 2025, reporting Sebastian Siemiatkowski’s interview with Bloomberg, fortune.com/2025/05/09/klarna-ai-humans-return-on-investment.
About the Author
Noel Guilford FCA is a chartered accountant and business adviser who works with a small number of deeply engaged business owners through a structured monthly advisory programme, The Business Alignment Programme. He is the founder of Guilford Accounting and writes on practice design, advisory strategy, and the intersection of AI with professional judgement.
To see which of the five stages your business should work on first, take the 7-minute Business Alignment Diagnostic.
To discuss how this approach might work for your business, book a discovery call at calendly.com/noelguilford.
