The whole company sits between saved time and retained profit
From saved time to retained profit, the whole company sits in between
An employee uses generative AI to reduce a two-hour task to one hour. Changes of this kind can already be measured in writing, customer service, and software-development experiments. The harder question for the company starts here: did it pay one hour less in wages, remove a software subscription, reduce an outsourcing bill, or complete an extra piece of billable work? In most cases, not yet.
The speed of an employee's task can change after one use. A company's costs and revenue are set by budgets, headcount plans, customer demand, approval rights, legacy systems, and contracts. Saved time is often scattered across small parts of many people's days. Salaries are still paid, software seats are renewed, and downstream review and compliance work remain. The team may feel markedly faster while the financial statements barely move.
As of September 2026, enterprise AI is producing an uneven set of operating changes. Some customer-service, route-optimisation, and high-volume transactional workflows have carried efficiency into lower unit costs or real resource reductions. Much knowledge work remains at the stage of employee time savings, local output growth, better customer experience, or capability building. In other projects, review, rework, and error costs consume the gain.
The figures come from the same 2026 McKinsey survey of 1,719 managers in 97 countries. They measure, respectively, individual experience, management attribution, and the survey's high-performer definition; they are not interchangeable. Source
What is happening now
What companies have actually gained so far
AI can already produce substantial changes in speed and quality when tasks have clear boundaries and outcomes are easy to judge. In a randomised writing experiment involving 453 professionals, completion time fell by 40% and blind-rated quality rose by 18%. A field study of 5,172 customer-service workers and roughly three million conversations found a 15% increase in successfully resolved issues per hour, with gains near 30% for less experienced workers. Across three randomised developer experiments involving 4,867 people, completed tasks increased by 26.08%.
These results show that a tool can change an individual task or work unit. They do not automatically tell us what happened to company payroll, staffing, software, outsourcing, or capital expenditure. Task completion is the front end of the operating chain. The company must still absorb the additional output, preserve quality, handle exceptions, and find either orders or costs that can actually be removed.
Swipe horizontally on mobile to view the full table
| Observed work | Measured change | What must still change at company level |
|---|---|---|
| Professional writing | Time fell 40% and blind-rated quality rose 18% in a 453-person experiment | The output must enter real approval, client-delivery, and billing processes |
| Customer service | 5,172 agents resolved 15% more issues successfully per hour | Waiting, escalation, scheduling, and unit service cost must also change |
| Software development | Completed tasks rose 26.08% across three experiments | Code still requires review, testing, deployment, maintenance, and a commercial use |
| Mature open-source projects | 16 experienced developers were 19% slower across 246 real issues | Perceived speed cannot substitute for measured completion time at equal scope and quality |
| Knowledge work | Email time fell in an experiment across 66 companies | Meeting time, task volume, and the mix of tasks did not change significantly |
Experimental sources: professional writing, customer service, developer experiments, METR, and the Microsoft 365 multi-company experiment.
Company-level data therefore moves more slowly. The U.S. Census Bureau's Business Trends and Outlook Survey found formal AI use at about 17.9% of firms. Among adopters, about 64% had made no institutional adjustment because of AI, and 95.7% had not changed total employment over the following six months. In another survey of nearly 6,000 executives in the United States, United Kingdom, Germany, and Australia, 89% said AI had not changed their company's labour productivity over the previous three years. Firms are using the tools, but use has not yet changed how most organisations operate.
Sources: U.S. Census Bureau and Firm Data on AI.
From working time to the accounts
Why one hour saved is not one hour of cost removed
Companies pay salaries by the month, not by the email or first draft. Saving twenty or forty minutes in an employee's day does not usually change payroll. The saving starts to enter the accounts only when the company hires one fewer person, leaves a vacancy unfilled, reduces overtime or outsourcing, or directs the time to work that increases revenue.
“Equivalent to a certain number of full-time employees” is not cash either. It divides saved hours by a standard work year to produce a capacity estimate. Actual staff reductions, genuinely avoided hiring, and moving employees to higher-value work are three different outcomes. If the company also adds model, cloud, data, integration, training, and review costs, apparent labour savings can coexist with higher total expenditure.
- 01The task worksOutput can be accepted with real data, permissions, and error costs
- 02The individual is fasterNet improvement remains after learning, waiting, checking, and rework
- 03The workflow is fasterApprovals, hand-offs, meetings, and exception queues do not consume the gain
- 04The company does moreAccepted output, volume, release speed, or customer results increase
- 05Resources changeHiring, outsourcing, overtime, legacy software, or low-value work genuinely decline
- 06Operations improveRevenue, unit cost, loss rates, or service capacity change
- 07Cash remainsA balance remains after model, cloud, integration, review, risk, and transition spending
Swipe horizontally on mobile to view the full table
| Common claim | What it actually represents | When it enters profit |
|---|---|---|
| Hours saved | Gross time saved by an individual or task | The time is consolidated and reassigned in a way that changes billable output or real expenditure |
| FTE equivalent | Theoretical capacity based on standard working hours | Hiring, replacement hiring, or outsourcing falls, or added output produces retained revenue |
| Avoided hiring | Future spending that did not occur relative to a growth plan | The original plan was credible, volume materialised, and another cost did not replace it |
| Lower cost | Reduced spending on payroll, outsourcing, software, fuel, or another account | The reduction did not come from lower activity or poorer quality and exceeds new AI costs |
| Higher profit | Revenue or cost changes that reach the income statement | Full investment and transition costs are deducted in the same operating period |
| Cash recovery | More cash received or less cash paid | Receivables, capital expenditure, restructuring payments, and contract payments are included |
Where the gain goes
Where saved time and extra output go
Efficiency that does not enter profit has not necessarily vanished. Rework can consume it. It can stop at the next approval, become more service or better quality, move to employees or customers, or accumulate as data and capability that may matter later. The distinction is who benefits and when the company can retain part of that benefit in its revenue or cost base.
Consumed by review and rework
More output creates more checking, exception handling, and repair. Deloitte's report for the Australian government required renewed quality assurance after errors and a refund of the AUD 97,587.11 final payment. The errors generated rework and a refund rather than revenue.
Held up in the next process
Email becomes faster, but meetings, approvals, and the mix of work remain. The individual gains time while the team's delivery pace is still determined by the same coordination work.
Reinvested in more output
Grindr reported roughly 2.5 times the engineering output while its technology team still grew by about 15%. The added capacity funded more work rather than an immediate reduction in staffing.
Transferred first to employees, customers, and suppliers
Employees spend less time on after-hours email, customers receive faster service or lower prices, and model, cloud, and software suppliers receive new spending. The deploying company's profit may not move at first.
Retained as a future option
P&G's 791-person experiment found that an individual working with AI could match the idea quality of human teams. Ideas still had to pass selection, development, launch, sales, and market tests before becoming profit.
Related sources: Australian parliamentary hearing record, Grindr disclosure, and the P&G randomised experiment.
Following the cases to economic results
Which operations are closest to turning AI into money
Economic results are usually easier to see when a workflow is close to a transaction, a unit cost, or a resource that can be removed. Customer service can track cost per contact, automated resolution, escalation, and team size. Route optimisation can track miles, fuel, and vehicle use. High-volume back-office work can track outsourcing, replacement hiring, and throughput. These operations are not necessarily the most sophisticated, but their revenue and cost outcomes are clearer.
Flywire disclosed that 45% of customer contacts were automatically resolved, cost per contact and handling time fell by 30%, volume rose by 19%, and team size fell by 2%. UPS's ORION route system reduced miles and fuel, with early disclosures reporting more than $400 million in annual savings and cost avoidance. In both cases, the algorithm's output connects to operating resources that can be counted.
Klarna reported that AI handled about 80% of customer-service conversations and linked $39 million of 2024 savings to capacity equivalent to more than 700 full-time employees; total company headcount also declined. Real resource reduction is visible, although the separate model, processing, and service cost of AI was not disclosed. At Block, engineering output, incident rates, staffing, and adjusted operating income all changed materially. Adjusted operating income rose 56% to $728 million in the first quarter of 2026, while GAAP operating loss was $172 million, including $852 million of restructuring and other charges. Productivity, staffing changes, reliability, and business performance are mixed into the same company result.
Swipe horizontally on mobile to view the full table
| Company or project | What happened | Where the value currently sits |
|---|---|---|
| Klarna | AI handled about 80% of service conversations; management cited $39 million of savings; total headcount declined | Resource reduction occurred, but stand-alone AI cost and long-term service outcomes were not separated |
| Flywire | 45% of contacts were automatically resolved; cost per contact and handling time fell 30%; volume rose 19% and team size fell 2% | Close to a unit-cost result; company-level cash contribution was not reported separately |
| UPS ORION | Miles and fuel fell; reported annual savings and cost avoidance exceeded $400 million | Real cash-bearing resources changed, although savings and avoided future spending were combined |
| Block | Code changes per engineer rose about 2.5 times, incidents fell, staffing declined, and profit measures diverged | Operating improvement occurred alongside restructuring and other business changes |
| Grindr | Engineering output rose roughly 2.5 times; management estimated $60 million of avoided-hiring value | Primarily a hiring counterfactual, not cash that had already left the cost base |
| DBS | More than 430 AI/ML use cases; management attributed about SGD 1 billion of economic value | Revenue, cost, and risk value were combined rather than reported as workflow-level profit |
| P&G | In a 791-person experiment, an individual with AI could match the idea quality of human teams | Value remained at idea generation and selection, not launch, sales, or profit |
| 66 companies / Microsoft 365 | Email work and after-hours activity declined | Meetings, task volume, and task mix did not change materially |
| Amazon Blue Jay | Development accelerated and reached production testing before operational use stopped | The project ended without an observed continuing operating return |
| Deloitte / DEWR | The report underwent renewed quality assurance after errors and the AUD 97,587.11 final payment was refunded | Value turned into rework, reputational cost, and a refund |
| Publicis | Growth, margin, and free cash flow improved in FY2025 | Company results improved, but AI's contribution was not separated |
| WPP | About £300 million was invested in AI, data, and platforms while revenue and profit declined | Cash investment is visible; AI-attributable cash inflow has not been reported separately |
Case sources: Klarna, Flywire, UPS, Block, DBS, Amazon, Publicis, and WPP.
One company can occupy several states at once
Nine operating realities
“Does AI pay?” is not a simple yes-or-no question. Within one company, customer service may already be lowering unit cost, R&D may be reinvesting time in more experiments, legal may be adding review, and finance may still be paying for both old and new systems. A specific workflow inside a specific company is the more useful unit of observation.
It feels faster but is not
Employees believe they save time, while complete measurement shows no gain or even slower work because of checking and switching.
Employees receive the time
Email and after-hours work decline, but the company does not convert that time into more output or lower cost.
A local efficiency island forms
Writing, coding, or ideation becomes faster while delivery remains constrained by meetings, approvals, data, and downstream capacity.
The constraint moves downstream
Once generation accelerates, selection, review, compliance, exception handling, and responsibility become the new queues.
Capacity funds more work
Staffing does not fall; the same or a larger team serves more customers, builds more features, or improves quality.
Customers and suppliers benefit first
Customers receive faster service or lower prices, while cloud, model, and software providers receive new spending and the deploying firm's profit is unchanged.
Old and new costs coexist
AI tools are added while legacy software, workflows, approvals, and staffing remain, causing total spending to rise first.
Cost genuinely declines
Replacement hiring, outsourcing, overtime, fuel, error losses, or legacy software spending leaves the cost base and exceeds the new investment.
The project stops or creates a loss
The effect is too weak, the risk too high, or the workflow cannot absorb it; the project ends or produces rework, refunds, or incidents.
From speed to retained value
Which companies can turn faster work into profit
Operations that approach a complete economic result tend to share several features. The task occurs frequently and its outcome can be judged quickly. AI addresses the current constraint on revenue, cost, or losses rather than an easy-to-demonstrate but unimportant step. One process owner can change staffing, budgets, and exception rules across the workflow. Downstream demand or a backlog can absorb the added capacity.
Resources must also be changeable. When savings are distributed across a few minutes in hundreds of employees' days, fixed payroll does not fall automatically. If manual data entry declines but the company keeps the same outsourcing contract and legacy software, costs simply stack. Even when revenue rises, competition may pass the gain to customers through lower prices, employees may receive some through pay and working time, and suppliers may receive some through model, cloud, and platform charges. The company retains only what remains after this distribution.
Success criteria are clear, errors can be found, and checking and rework do not consume the time saved.
It changes orders, unit costs, waiting, losses, or current backlog rather than an incidental step.
The owner can change hand-offs, approvals, staffing, budgets, and exception handling, not merely purchase the tool.
Orders, customer requests, service commitments, or clearly removable costs are available.
Hiring, outsourcing, overtime, software seats, and low-value work can genuinely decline.
Old systems and duplicate workflows are retired on schedule rather than maintained indefinitely.
Lower prices, higher pay, and supplier charges do not transfer the entire benefit elsewhere.
Quality, risk, and maintenance costs do not reverse the early gain in later periods.
Watch operating changes, not usage heat
What management should keep watching
Seat counts, calls, prompts, and generated content show that employees are using the tools. They do not show an operating result. More useful measures follow the work from output and hand-offs through capacity and into the accounts: did accepted output increase, did the full workflow shorten, did demand absorb the additional capacity, which spending actually left, and did cash remain?
Swipe horizontally on mobile to view the full table
| Where to look | Operating signals to record | The question they answer |
|---|---|---|
| Usable output | Accepted output per all-in hour, first-pass acceptance, review minutes, rework, and defects | Whether employees are truly faster or merely generating more |
| Full workflow | AI-step time versus end-to-end cycle, hand-offs, approvals, work in progress, and exception escalation | Whether the faster step changes the downstream process |
| Organisational capacity | Paid or accepted output per employee, backlog, service commitments, and capacity use | Whether local time savings become deliverable company output |
| Staffing and outsourcing | Hiring, replacement hiring, attrition, contractors, overtime, and reassigned hours | Whether a theoretical FTE estimate becomes a real resource change |
| Revenue and customers | Volume, price, conversion, renewal, response time, and retention | Whether demand absorbs the capacity and the company retains the gain |
| Full cost | Model, cloud, data, integration, review, incident, and total cost per accepted result | Whether back-end spending consumes the front-end saving |
| Legacy-cost removal | Retirement of old software seats, duplicate processes, outsourcing contracts, and parallel systems | Whether the company is still paying for two production systems |
| Cash and capital | Payback, working capital, capital expenditure, restructuring payments, cash conversion, and capital turnover | Whether an income-statement improvement finally reaches cash |
These signals need to sit within the same workflow and operating period. Higher company profit alongside higher cost for one AI project does not establish that the project created profit. Lower employee hours alongside lower business volume does not automatically establish productivity either. A useful record is continuous: accepted output for comparable work rises, the end-to-end cycle shortens, demand absorbs the capacity, staffing or other expenditure changes as planned, and an operating gain remains after the new technology costs are deducted.
Conditional paths, not one forecast
How the next 12–36 months may diverge
Companies will not follow one route. Across the next two or three budget cycles, the important question is not whether tool use keeps rising. It is whether companies begin removing legacy resources, changing end-to-end workflows, and retaining part of the resulting revenue or cost benefit.
AI becomes an additional cost layer
Individual tools spread quickly while staffing, legacy software, meetings, and approvals remain, and review and governance keep growing. Usage and supplier revenue rise, but unit cost, operating profit, and cash flow do not improve.
Change signal: legacy seats, outsourcing, approvals, and real spending decline across consecutive budget periods.Operating leverage follows transition spending
The company rewrites the full workflow, names an owner and exception route, retires legacy systems, and lets hiring or outsourcing respond to capacity. Early costs are high; unit cost, expense ratios, and cash conversion improve later.
Contrary signal: full unit cost still has not declined after two operating periods of deep integration.Output expands while customers receive most of the gain
The company handles more business, but competition forces lower prices, faster service, or higher quality. Output per employee rises while margins may remain flat because customers receive most of the efficiency benefit.
Change signal: gross margin expands while price and customer retention hold.Profit concentrates after supply expands
Peers simultaneously increase the supply of content, code, service, or products while demand cannot absorb all of it. Prices come under pressure and companies with data, distribution, brands, and lighter legacy burdens gain share.
Contrary signal: unit cost and profit improve at the median company without wider dispersion between firms.Value appears mainly through fewer losses and greater option value
High-loss industries retain human responsibility while AI first reduces omissions, expands checking, or accelerates experimentation. Near-term profit changes little; incident losses, quality, and future product capacity become the main returns.
Validation signal: loss frequency, rework, launch success, or risk-capital requirements improve persistently.“Returns need more time” cannot be extended indefinitely. If a high-volume workflow has used AI deeply for two budget or operating cycles without changing end-to-end unit cost, resource reductions, legacy-system retirement, or cash results, the more useful explanation shifts from “the return has not arrived” to “the current design does not produce an economic result the company can retain.”
The distribution of profit continues beyond the deploying company
What this means for enterprise software, labour, management, and competition
Enterprise software: addition first, compression later
In the near term, companies pay for legacy seats and AI add-ons at the same time, raising customer cash spending and supplier revenue. If agentic tools gradually complete whole stretches of work, low-use seats, duplicate features, and per-person pricing will come under pressure. Pricing may move towards usage, accepted work units, or outcomes.
Labour: slower hiring usually precedes layoffs
Fragmented time savings first move into other tasks or employee welfare. Companies can more easily reduce replacement hiring, contractors, and overtime before large-scale layoffs. Over time, junior and repetitive roles may shrink, while review, domain judgement, and responsibility become scarcer. If the junior pipeline contracts too quickly, experienced review capacity may become a constraint later.
Management: tools do not remove coordination automatically
Personal assistants reduce information-processing work but do not cancel meetings, approvals, or conflict between teams. Early deployment often adds policy, review, and exception management. Coordination work and management spans change only when decision rights, process ownership, and exception authority change as well.
Cloud, model, and consulting suppliers receive revenue first
Companies pay for models, computing, data, integration, and organisational change before their own profit improves. Suppliers therefore receive revenue first, although their own net profit still depends on computing costs, competition, and pricing. Longer-term profit is more likely to remain around proprietary data, distribution, workflow control, and assurance.
Competition: higher average efficiency does not mean everyone earns more
Large companies have data, customers, and scale but carry more legacy systems. New companies have lighter workflows but less trust, data, and distribution. Industry productivity can rise through expansion by high-productivity firms, contraction or exit by low-productivity firms, and shifts in market share rather than equal improvement across incumbents.
Customers: faster and cheaper become the new baseline
When more firms can respond faster and expand supply, competition turns what was once a chargeable efficiency into a service standard. Without a brand, proprietary workflow, or customer relationship that retains part of the value, efficiency becomes a cost of competing rather than extra profit.
Technology installation and economic results rarely arrive on the same day
History has seen similar transitions
The spread of steam power was not simply a matter of machine performance. Firms with heavy sunk investment in water power found it harder to replace their production systems, while new entrants could adopt the new power source more readily. Productivity gains therefore emerged through entry, expansion, and exit rather than every incumbent upgrading in the same year. Related research
Electrification also required factories to rearrange machinery, material flows, and work organisation, but historical evidence does not support reducing the story to “decades without returns”. Industries with the complementary conditions saw labour-productivity gains relatively quickly. The distinction matters: a historical transition can explain why returns differ, but it cannot justify waiting indefinitely for any low-return project. Related research
Enterprise IT, ERP, and barcodes offer closer parallels. After hardware installation, companies still had to address master data, standard processes, employee skills, decision rights, and legacy-system retirement. Migration, training, and data-cleaning costs could depress performance at first. The early labour-productivity gain from barcodes alone was modest; their larger value came when point of sale, inventory, replenishment, pricing, and logistics shared the same network of data. Enterprise IT, ERP, and barcodes
Productivity gains in internet retail did not occur evenly across every firm either. Entry and expansion by high-productivity retailers, together with contraction and exit by low-productivity firms, lifted the industry average. AI may spread through a similar process: widening differences between companies and workflows before hiring, market share, and exit change average industry performance. Related research
Final assessment
Final assessment
Companies have not yet become more profitable across the board because of AI, but that does not mean all the tools are ineffective. Repeated evidence already supports faster individual tasks and better quality in some work units. The missing part is the second half of the operating chain: approvals, hand-offs, demand, staffing, software, and budgets have not changed at the same speed. Saved time therefore remains inside workflows, moves to employees and customers, funds more output, or is absorbed by model, cloud, review, and transition costs.
Profit appears earlier under a recognisable set of conditions. The task is frequent and easy to accept, AI addresses the current constraint on cost or revenue, downstream demand can absorb the added capacity, an owner can change the full workflow, and hiring, outsourcing, fuel, error losses, or legacy software can genuinely leave the resource base. Automated customer resolution, route optimisation, and some transactional workflows already show this path.
Much knowledge work remains in the middle. Companies receive faster drafts, code, analysis, and ideas while continuing to pay fixed salaries and retain legacy systems and existing coordination structures. Until volume, price, or resources change, local productivity will not enter profit in the same proportion. If new investment continues to stack, the intermediate state itself becomes a continuing cost.
The signals that would change this assessment are concrete: accepted output keeps rising, end-to-end cycles shorten, replacement hiring, outsourcing, overtime, or legacy-software spending begins to leave, unit service cost falls after full AI spending is deducted, revenue grows without an equivalent price reduction, and cash conversion improves. If deep use passes two operating cycles without these changes, the company is no longer dealing only with a timing problem. Its current workflow and business model cannot retain the efficiency AI creates.
Main sources
Main sources
- McKinsey, The State of AI 2026
- U.S. Census Bureau, Business Trends and Outlook Survey AI working paper
- Firm Data on AI
- Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence
- Generative AI at Work
- The Effects of Generative AI on High-Skilled Work
- METR, Early-2025 AI Experienced Open-Source Developer Study
- Shifting Work Patterns with Generative AI
- The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise
- Still Waters, Rapid Currents: Wage and Employment Effects of AI
- EIBIS–ORBIS evidence on AI adoption, productivity and employment
- Mapping AI to Production: A Field Experiment at the Firm Level
- Klarna Group plc, 2025 annual filing
- Block, Q1 2026 shareholder letter
- Flywire, Q2 2026 investor materials
- DBS Annual Report 2025, CIO statement
- Grindr, Q2 2026 shareholder letter
- Parliament of Australia, committee hearing record
- WPP Annual Report 2025
- Publicis Groupe, FY 2025 results