Value
Does AI pay off? Here is how it shows.
The value shows up either in the bottom line or in room for people to do other work. Here are the examples by step, and what others have measured, good and bad.
Move the needle
How the value shows up
The same project delivers different value depending on when you measure. First time is saved, then quality improves, and finally it shows up in operations. Most projects are judged on the first measure - the value is in the last.
Direct impact on results
Shows up in the books. Revenue that isn’t lost and costs that aren’t incurred.
Room for other work
Time people get for other things. When there’s room, it’s not just output that grows - the work that has sat in the backlog finally gets done. When small things pile up they create bottlenecks elsewhere, and they usually land on the people who should be working on other value-creating work.
Room
Time is freed
The to-do list shrinks
Work that has waited gets done
Bottlenecks open up
What piled up stops holding other work back
Better service
Customers get answers and follow-up
Customer loyalty
Customers are less likely to go to a competitor, and ask for more
Winning a new customer can cost 5X to 25X more than keeping an existing one. The research is not new, but the rule still holds today. That’s why customer loyalty is often the most valuable outcome - but also the hardest to tie directly to a single project. Source: Harvard Business Review, 2014
18 examples
The examples start conservatively, at the second-lowest value. Move the slider on a card to match your size.
- Time trackingStep 04Implementation: Medium
Time logging in the ticketing system
What this solves
- FirstLogging takes minutes
- ThenLogged as you go instead of at the end of the week
- In the long runHours aren’t forgotten, and overruns and gaps in work show up straight away
Measured: Time spent logging → logging frequency → uninvoiced hours
Open the example →Return
ISK 10.1ma year
ISK 842k a month
Direct impact on results
0ISK 25m Model cost: ISK 190 a day per person × 10 → ISK 437k a year
The cost of the environment, implementation and operation comes on top.
Example: 10 consultants, 1 hour a week each, ISK 22,000
How it’s calculated
Forgotten hours a week × staff × billed hourly rate × 46 (working weeks)
- Sales and marketingStep 03Implementation: Large
Answering questions before a purchase
What this solves
- FirstAn answer in minutes, out of hours too
- ThenFewer give up before they buy
- In the long runMore enquiries turn into sales
Measured: Response time → share who buy → sales
Open the example →Return
ISK 3.6ma year
ISK 300k a month
Direct impact on results
0ISK 25m Estimated model cost: ISK 25 per enquiry × 1,200 a year → ISK 30k a year
The cost of the environment, implementation and operation comes on top.
Example: 100 enquiries a month, 2 percentage points more buy, ISK 150,000 margin per sale
How it’s calculated
Enquiries a month × more who buy × margin per sale × 12
- FinanceStep 04Implementation: Large
Invoicing in line with contracts
What this solves
- FirstInvoices are checked against contracts before they go out
- ThenMissed items and outdated prices are caught straight away
- In the long runRevenue doesn’t leak and price increases come through
Measured: Invoice discrepancies → corrections → revenue recovered
Open the example →Return
ISK 2.7ma year
ISK 225k a month
Direct impact on results
0ISK 25m Estimated model cost: ISK 20 per invoice × 900 a year → ISK 18k a year
The cost of the environment, implementation and operation comes on top.
Example: 75 invoices a month, 2% with an under-billed item, ISK 150,000 on average
How it’s calculated
Invoices a month × share with an under-billed item × average amount × 12
- Software teamsStep 03Implementation: Small
Tests written alongside code
What this solves
- FirstTests are written with the code
- ThenFewer bugs ship with the code
- In the long runMore releases, less risk
Measured: Tests → production bugs → release frequency
Open the example →Return
ISK 2.0ma year
ISK 167k a month
Direct impact on results
0ISK 25m Estimated model cost: ISK 1,000 a day per person × 3 → ISK 690k a year
The cost of the environment, implementation and operation comes on top.
Example: 3 developers, 1.7 fewer bugs a year each, ISK 400,000 per bug (analysis, fix, disruption)
How it’s calculated
Developers × bugs avoided per developer a year × cost per production bug
- Software teamsStep 04Implementation: Large
Agents on maintenance
What this solves
- FirstBugs fixed faster
- ThenThe team can focus on new features
- In the long runLess technical debt, more frequent releases, and bottlenecks move, for example to code review
Measured: Time per bug → share of new development → technical debt
Open the example →Return
383 hoursa year
32 hours a month
Room for other work
Extra capacity equal to 0.2 full-time roles
Worth ISK 4.6m if the time is used
01,500 h Estimated model cost: ISK 1,500 a day per person × 3 → ISK 1.0m a year
The cost of the environment, implementation and operation comes on top.
Example: 3 developers, 25% on maintenance, agents take 30% of it, ISK 12,000
How it’s calculated
Developers × 1,700 hours × share of maintenance × share agents take × hourly cost
- FinanceStep 03Implementation: Medium
Asking the financial data directly
What this solves
- FirstA summary that took hours takes minutes
- ThenThe position is reviewed weekly, with room to think outside the box
- In the long runDeviations are caught sooner
Measured: Time spent summarising → review frequency → variance from budget
Open the example →Return
384 hoursa year
32 hours a month
Room for other work
Extra capacity equal to 0.2 full-time roles
Worth ISK 3.9m if the time is used
01,500 h Estimated model cost: ISK 190 a day per person × 2 → ISK 87k a year
The cost of the environment, implementation and operation comes on top.
Example: 2 staff, 16 hours a month each. Overruns stopped aren’t counted
How it’s calculated
Hours summarising a month × hourly cost × 12
- Customer serviceStep 04Implementation: Large
A chatbot that answers enquiries
What this solves
- FirstSimple enquiries drop
- ThenStaff get time for harder cases
- In the long runBetter service, and customer loyalty grows
Measured: Resolved without a person → first-contact resolution → churn
Open the example →Return
360 hoursa year
30 hours a month
Room for other work
Extra capacity equal to 0.2 full-time roles
Worth ISK 3.6m if the time is used
01,500 h Becomes a direct impact on results if a new hire is avoided.
Estimated model cost: ISK 25 per conversation × 9,000 a year → ISK 225k a year
The cost of the environment, implementation and operation comes on top.
Example: 750 cases a month, 40% resolved, 6 minutes each
How it’s calculated
Cases a month × share resolved × minutes per case ÷ 60 × hourly cost × 12
- Sales and marketingStep 03Implementation: Medium
Preparing for customer meetings
What this solves
- FirstA customer summary ready before the meeting
- ThenAccount managers can look after more customers
- In the long runCustomers ask for more and stay longer
Measured: Preparation time → meetings a month → growth and churn
Open the example →Return
276 hoursa year
23 hours a month
Room for other work
Extra capacity equal to 0.2 full-time roles
Worth ISK 3.3m if the time is used
01,500 h Becomes a direct impact on results when it turns into more business.
Estimated model cost: ISK 190 a day per person × 2 → ISK 87k a year
The cost of the environment, implementation and operation comes on top.
Example: 2 account managers, 3 hours a week each, ISK 12,000. Extra business isn’t counted
How it’s calculated
Account managers × hours saved a week × 46 (working weeks) × hourly cost
- FinanceStep 04Implementation: Medium
Invoice coding and reconciliation
What this solves
- FirstFewer manual entries
- ThenLess time goes on month-end closing
- In the long runThe team works on other things rather than data entry, and finance stays lean with little overhead
Measured: Manual entries → days to close → share of analysis work
Open the example →Return
240 hoursa year
20 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 2.4m if the time is used
01,500 h Estimated model cost: ISK 10 per invoice × 3,600 a year → ISK 36k a year
The cost of the environment, implementation and operation comes on top.
Example: 300 invoices a month, 4 minutes saved on each
How it’s calculated
Invoices a month × minutes saved ÷ 60 × hourly cost × 12
- Legal and complianceStep 03Implementation: Medium
Contract review
What this solves
- FirstDeviations from standard terms flagged automatically
- ThenThe lawyer reads what matters
- In the long runContracts close sooner
Measured: Time per contract → deviations found → incidents after signing
Open the example →Return
113 hoursa year
9 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 2.0m if the time is used
01,500 h Estimated model cost: ISK 300 per contract × 75 a year → ISK 23k a year
The cost of the environment, implementation and operation comes on top.
Example: 75 contracts, 1.5 hours saved on each, ISK 18,000. Reduced risk isn’t counted
How it’s calculated
Contracts a year × hours saved × hourly cost
- Customer serviceStep 03Implementation: Medium
Case history summaries
What this solves
- FirstThe agent sees the case history in seconds
- ThenThe customer doesn’t have to repeat themselves
- In the long runFewer complaints and more customer loyalty
Measured: Time per case → repeat enquiries → churn
Open the example →Return
180 hoursa year
15 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.8m if the time is used
01,500 h Estimated model cost: ISK 10 per case × 3,600 a year → ISK 36k a year
The cost of the environment, implementation and operation comes on top.
Example: 300 cases a month, 3 minutes saved on each
How it’s calculated
Cases a month × minutes saved ÷ 60 × hourly cost × 12
- ITStep 02Implementation: Small
Documenting legacy systems
What this solves
- FirstThe first description of an old system gets written
- ThenNew people don’t have to ask the one person who knows
- In the long runLess key-person risk
Measured: Documented systems → questions to key people → time to productivity
Open the example →Return
138 hoursa year
12 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.7m if the time is used
01,500 h Estimated model cost: ISK 1,000 a day per person × 1 → ISK 230k a year
The cost of the environment, implementation and operation comes on top.
Example: 3 hours a week, ISK 12,000. The cost if they leave isn’t counted
How it’s calculated
Key person’s hours answering a week × 46 (working weeks) × hourly cost
- Software teamsStep 01Implementation: Small
An assistant for software development
What this solves
- FirstEach person faster on simple tasks
- ThenThe team starts working the same way
- In the long runMore predictable and visibly faster delivery of code
Measured: Time per task → variance in ways of working → plan reliability
Open the example →Return
Up to138 hoursa year
12 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.7m if the time is used
01,500 h Estimated model cost: ISK 250 a day per person × 3 → ISK 173k a year
The cost of the environment, implementation and operation comes on top.
Example: 3 developers, 1 hour a week. Can be zero without a shared way of working
How it’s calculated
Developers × hours saved a week × 46 (working weeks) × hourly cost - licence cost
- ITStep 04Implementation: Medium
Internal service desk
What this solves
- FirstCommon requests are resolved automatically
- ThenIT staff get time for harder cases
- In the long runFewer disruptions to operations
Measured: Automatic resolution → waiting time → incidents
Open the example →Return
158 hoursa year
13 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.6m if the time is used
01,500 h Becomes a direct impact on results if a new hire is avoided.
Estimated model cost: ISK 25 per request × 1,800 a year → ISK 45k a year
The cost of the environment, implementation and operation comes on top.
Example: 150 requests a month, 35% resolved, 15 minutes each
How it’s calculated
Requests a month × share resolved × minutes ÷ 60 × hourly cost × 12
- HRStep 03Implementation: Small
Onboarding new starters
What this solves
- FirstNew starters get answers straight away, without interrupting colleagues
- ThenProductivity comes sooner
- In the long runA better first experience and less chance they look elsewhere
Measured: Time to full productivity → onboarding experience → first-year attrition
Only 12% of employees think their workplace does a great job of onboarding new starters. Gallup
In a survey, just over half of new starters said they were undertrained after onboarding - and four in five of them planned to look elsewhere soon. Paychex, 2022
Open the example →Return
140 hoursa year
12 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.4m if the time is used
01,500 h Becomes a direct impact on results when attrition and rehiring are avoided.
Estimated model cost: ISK 190 a day per person × 4 for 65 days → ISK 49k a year
The cost of the environment, implementation and operation comes on top.
Example: 4 new starters, 2 weeks sooner, 50% productivity gap, ISK 350,000 a week. Attrition avoided isn’t counted
How it’s calculated
New starters a year × weeks sooner at full productivity × productivity gap × weekly cost
- Legal and complianceStep 03Implementation: Small
Questions on internal rules
What this solves
- FirstAnswers in minutes instead of searching and making sense of large handbooks
- ThenFewer questions to compliance
- In the long runBetter compliance and understanding of the rules in force
Measured: Response time → enquiries → audit findings
Open the example →Return
120 hoursa year
10 hours a month
Room for other work
Extra capacity equal to 0.1 full-time roles
Worth ISK 1.2m if the time is used
01,500 h Estimated model cost: ISK 25 per query × 360 a year → ISK 9k a year
The cost of the environment, implementation and operation comes on top.
Example: 30 enquiries a month, 20 minutes each
How it’s calculated
Enquiries a month × minutes ÷ 60 × hourly cost × 12
- HRStep 03Implementation: Small
Answers from the employee handbook
What this solves
- FirstFewer repeat questions to HR
- ThenEveryone gets the same answer in seconds
- In the long runHR develops people rather than answering questions
Measured: Enquiries → response time → share of development work
Open the example →Return
60 hoursa year
5 hours a month
Room for other work
Extra capacity equal to 0.0 full-time roles
Worth ISK 0.6m if the time is used
01,500 h Estimated model cost: ISK 15 per query × 360 a year → ISK 5k a year
The cost of the environment, implementation and operation comes on top.
Example: 30 enquiries a month, 10 minutes each
How it’s calculated
Enquiries a month × minutes ÷ 60 × hourly cost × 12
- For everyoneStep 02Implementation: Small
Minutes and follow-up
What this solves
- FirstThe minutes are ready straight away
- ThenActions from meetings aren’t forgotten
- In the long runDecisions turn into action faster
Measured: Time spent writing up → open actions → time from decision to action
Open the example →Return
60 hoursa year
5 hours a month
Room for other work
Extra capacity equal to 0.0 full-time roles
Worth ISK 0.6m if the time is used
01,500 h Estimated model cost: ISK 50 per meeting × 120 a year → ISK 6k a year
The cost of the environment, implementation and operation comes on top.
Example: 10 meetings a month, 30 minutes each
How it’s calculated
Meetings a month × minutes writing up ÷ 60 × hourly cost × 12
Typical value development with worked examples - not measured results from specific projects. Assumptions: internal cost ISK 10,000 an hour (about ISK 18m a year per full-time role), specialists ISK 12,000-18,000, billed rate ISK 22,000, 46 working weeks a year. Direct impact on results shows up in the books - revenue that isn’t lost or costs that aren’t incurred. Room for other work is time people get for other things - it only turns into money if it’s used. Model cost is the use of the AI models and the processing. It was measured at ISK 190 a day per person for time logging; elsewhere it is estimated from the volume, and heavier work such as agents in software development costs more. The cost of the environment itself, implementation and operation is not included and comes on top.
First 12 months:
Per year at full effect:
Your assumptionsWhat others have measured
Published results, good and bad. They show where the value lies - and where it fails to show up.
For companies and public bodies
95% with no measurable return
No strategy · MIT
In a 2025 MIT study, about 95 percent of AI pilots had no measurable impact on the bottom line. What set the other 5 percent apart was not the technology, but whether the tools learned from use and fitted the processes. Buying and partnering with specialists succeeded about twice as often as building in-house.
Sources: MIT NANDA (PDF) · Fortune
3% of working hours
Step 01 · Denmark
A study of about 25,000 workers at 7,000 Danish workplaces: chatbot users reported saving 2.8 percent of their working hours on average. No measurable effect was found on earnings or hours worked. Everyday use pays off, but on its own it does not change how the business runs.
26 minutes a day
Step 01 · UK civil service
Twenty thousand civil servants got Copilot for three months and said they saved 26 minutes a day on average - nearly two weeks a year. The UK tax authority measured something more modest, two to three percent of the working week, and 17 percent said they saved nothing. Licences pay off, but modestly. The big step comes when the tool connects to data and processes.
An hour a day - for some
Step 01 · Australian Government
Over 7,600 Australian public servants had Copilot for six months in 2024. Users felt they saved about an hour a day on summaries, first drafts and search, and 86 percent wanted to keep it. But only about a third used it daily, and up to 7 percent said it made tasks longer because checking the answers ate up the savings.
Sources: ACS Information Age · DTA
25% faster - or 19 points worse
Step 01 · BCG and Harvard
758 BCG consultants were given GPT-4. On tasks within its reach they completed 12 percent more tasks, 25 percent faster and at over 30 percent higher quality. On a task beyond it, they were 19 percentage points less likely to get it right. Knowing where the limits are is part of the ground rules.
Source: Harvard Business School
812 Canadian dollars
Step 01 · Air Canada
Air Canada’s chatbot told a passenger he could claim a bereavement fare after the fact, which was wrong. A British Columbia tribunal held the airline liable in 2024: the company is responsible for what its chatbot says. The amount was small, the lesson large. Ground rules and testing come before AI talks to customers.
15% more issues an hour
Step 03 · Customer support
Over 5,000 support agents got an assistant built on the company’s past conversations. They resolved 15 percent more issues an hour on average, and less experienced agents about 30 percent more. The most experienced gained little. What the best people knew became available to everyone.
98% of advisor teams
Step 03 · Morgan Stanley
Morgan Stanley connected an assistant to about 100,000 internal reports and documents for its financial advisors, and 98 percent of advisor teams have adopted it. The answers draw on the firm’s own data, not general information from the web.
Sources: Morgan Stanley · CNBC
From 11 minutes to two
Step 04 · Klarna
An agent took over two thirds of Klarna’s customer service chats in its first month, and resolution time fell from 11 minutes to under two. Later the company hired people again for the harder cases. The value was real - but it took a person where judgement matters.
Source: HubSpot
For software teams
55% faster - or 19% slower
Step 01 · Two studies
In a controlled experiment, developers with Copilot finished a simple task from scratch 55.8 percent faster. Two years later METR measured experienced developers on their own codebases: they were 19 percent slower with AI, even though they believed they were faster. The assistant alone guarantees nothing. The way of working decides.
Source: TechCrunch (METR)
-7.2% in stability
Step 02 · DORA
In Google’s 2024 DORA report, every 25 percent increase in AI adoption came with a 1.5 percent drop in delivery throughput and a 7.2 percent drop in stability, even as documentation, code quality and reviews improved. Individuals get faster, the team does not automatically. The way of working has to keep up.
Sources: Google Cloud · DORA (PDF)
24% more changes
Steps 02-03 · Microsoft
Tens of thousands of Microsoft developers adopted Claude Code and Copilot CLI in early 2026. Those who adopted the tools delivered about 24 percent more merged changes, and the difference held for four months. The authors note themselves that a merged change is not the same as value.
75% of new code
Step 03 · Google
In April 2026 Google said 75 percent of all new code was written by AI and approved by engineers, up from just over a quarter in autumn 2024. The figures are Google’s own. The key word is approved: every change goes through review.
Sources: Google (2026) · Google (2024)
4,500 developer-years
Step 04 · Amazon
Amazon used AI to upgrade its Java systems. The average upgrade went from about 50 developer-days to a few hours, over half of its production systems were upgraded in under six months, and the gain was put at about 260 million dollars a year. 79 percent of the auto-generated changes passed review without further edits.
Source: Andy Jassy (LinkedIn)
1,300 a week
Step 04 · Stripe
Stripe’s agents deliver over 1,300 merged changes a week with no human-written code. Every one goes through review by an engineer. It worked because tests, isolated environments and access to internal tools were already in place.
Figures are as published by each organisation. Some are self-reported by participants.
Where is the value for you?
A half-hour conversation is often enough to see which step applies and where the value is greatest.
Talk to us
Get in touch directly with the people behind this. We answer ourselves.
Andri Örvar Baldvinsson
andriorvar@apro.is
Páll Jónsson
pall@apro.is
Hlöðver Þór Árnason
hlodver@apro.is
Or email us at apro@apro.is