TIP’s practitioner observations and proposed applications are identified in the relevant pages. Anonymous accounts illustrate a mechanism; they do not establish its prevalence.
Peer-reviewed field experiment · 2026-06-12
The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork
Organization Science / INFORMS. In a P&G innovation experiment involving 791 professionals, AI-assisted individuals matched unaided teams on proposal quality. AI teams were 9.2 percentage points more likely than unaided individuals to produce a top-decile solution.
Scope. One company and early innovation tasks, not realised growth. The AI-team versus unaided-team top-decile difference was not statistically significant. The final analysis excluded non-randomised participants; 776 participants supplied complete post-task data.
Read the original source ↗Peer-reviewed workplace deployment study · 2025-02-04
Generative AI at Work
The Quarterly Journal of Economics / Oxford University Press. Across 5,172 support agents, AI assistance increased issues resolved per hour by 15% on average. Gains were larger for less experienced and lower-skilled agents. Customer requests for supervisors declined. Lower attrition was associated with AI access, especially among newer staff; experienced high performers saw small quality declines.
Scope. One occupation, company and tool. Attrition evidence has weaker causal identification than productivity: selection into tool access could exaggerate the retention effect. This is not evidence for a general headcount reduction percentage.
Read the original source ↗Research explainer and linked randomised-study preprint · 2026-01-29
How AI assistance impacts the formation of coding skills
Anthropic Research. A small experiment recruited 52 mostly junior Python engineers learning an unfamiliar library. The AI group averaged 50% on an immediate comprehension quiz versus 67% without AI: a 17-percentage-point gap. The approximately two-minute completion-time difference was not statistically significant. Explanatory and conceptual uses of AI were associated with stronger understanding than wholesale delegation.
Scope. Short task and immediate quiz, not a longitudinal workplace study. Interaction-pattern comparisons are observational. The source's shorthand '17% lower' should be rendered as 17 percentage points using its reported means.
Read the original source ↗Company annual filing; company-reported attribution · 2026-02-26
Klarna Group plc Annual Report on Form 20-F for 2025
Klarna Group plc. Klarna reports a $7m Q1 2024 reduction in external marketing-supplier spending versus Q1 2023 and attributes 37% of that decrease to AI. Named services include translation, production, CRM and social agencies. The filing also describes workforce reduction alongside AI efficiency, cost optimisation and normal-course employee attrition.
Scope. Management attribution, not an experiment. Supplier reduction is not proof every service moved in-house. The filing does not establish that every leaver went unreplaced, that there were no layoffs, or that all workforce changes were caused by AI.
Read the original source ↗Randomised field experiment; working paper · 2025-11-13
Shifting Work Patterns with Generative AI
NBER; author manuscript on arXiv. The six-month experiment covered 7,137 knowledge workers at 66 firms. In its second half, the 80% of treated workers who used the tool spent about two fewer hours on email weekly and reduced work outside regular hours. Researchers did not detect broader changes in the quantity or mix of tasks from individual tool provision.
Scope. These are specific tool and work-pattern findings, not proof of two days released per week. Some authors were employed by Microsoft, the tool maker. Use the revised estimates rather than the earlier 3.6-hour estimate for frequent users.
Read the original source ↗Central-bank survey of 105 medium and large Australian firms · 2025-11-13
Technology Investment and AI: What Are Firms Telling Us?
Reserve Bank of Australia. Surveyed firms reported using a mix of retraining, redeployment, natural attrition and redundancy when technology displaced work. Most displaced workers were reported as redeployed with minor retraining. Firms expected further adjustments, but few reported material AI-driven productivity gains at the survey date.
Scope. Self-reported liaison sample, not representative of all employers. Technology includes more than AI; expectations are not realised job outcomes. The research does not establish a universal non-backfill policy.
Read the original source ↗Peer-reviewed survey · 2025-04
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers
ACM CHI 2025; Microsoft Research. 319 knowledge workers supplied 936 examples of AI use. Greater confidence in AI was associated with less reported critical-thinking effort; confidence in one's own ability was associated with more. Thinking shifted toward checking information, integrating output and overseeing the task.
Scope. Self-reported correlations. This does not prove AI causes declining intelligence or long-term skill loss.
Read the original source ↗Official decision framework · 2026-09-30
The Sourcing Playbook
UK Government Commercial Agency / Government Commercial Function. The framework compares internal, purchased and mixed delivery using whole-life costs, capability, risks and outcomes.
Scope. Public-sector guidance, not evidence that insourcing saves money or a general legal requirement for private companies. Applying the method here is TIP’s interpretation.
Read the original source ↗Research study · 2026-02-18
Adoption and efficacy of LLMs in US consumer financial complaints
Nature Human Behaviour. An analysis of 1,134,512 US financial complaints estimated a 6.9 percentage-point increase in favourable-relief probability associated with LLM adoption (95% confidence interval: 4.9–8.9 points).
Scope. An instrumental-variable estimate, not a UK insurance trial. Favourable relief includes outcomes beyond cash payments; do not turn this into a universal P&L percentage.
Read the original source ↗Government estimate · 2026-04-02
Government response on the new subscription contracts regime
UK Government. The government estimates UK consumers spend £1.6bn a year on unwanted subscriptions.
Scope. A market exposure estimate, not a measured effect of AI or a company-level revenue forecast. This resource is used for the estimate, not as legal guidance on implementation dates.
Read the original source ↗Regulator research · 2026-02-26; updated 2026-06-30
New research reveals how to cut phone and internet bills
Ofcom. Ofcom reports 28% of broadband customers out of contract, paying on average £7–£9 more monthly than those in contract.
Scope. Evidence of the cost of inaction in broadband. It does not measure AI-driven switching or establish the same exposure in another market.
Read the original source ↗Owner-reported case and white paper · Account updated 2026-10-22
The report doesn’t fit. The insurer says no. What next?
Building Trust. A structural-warranty dispute involving a 16-home building describes AI-supported organisation of evidence alongside professional and collective work. The account records an accepted net settlement of £576,022.75.
Scope. One case, based on the owner’s account and retained documents. Accepted settlement is not a statement of cash received or the project budget. The result cannot be attributed entirely to AI or treated as typical.
Read the original source ↗Consultation · 2026-07-30
Consultation on the Code of Practice on disciplinary and grievance procedures
Acas. Acas’s consultation discusses growing AI use in disciplinary and grievance processes and concerns about longer, more complex grievances.
Scope. The consultation closed on 23 September 2026; proposals are not current law. It does not establish an increase in successful employee claims.
Read the original source ↗Vendor-reported detections · Reporting period March 2025–May 2026
The Autonomous CFO: agentic AI expense audit
AppZen. AppZen reports 1,548 AI-generated receipts flagged for human review during its reported period.
Scope. Vendor data with no general prevalence denominator. A flag is not a confirmed fraudulent claim, and detection counts are not a rate of employee fraud.
Read the original source ↗Observational study · 2025-07-22
Google users are less likely to click on links when an AI summary appears
Pew Research Center. In March 2025 browsing data from 900 US adults, users clicked a traditional result on 8% of visits with an AI summary versus 15% without one.
Scope. Observed Google behaviour in a particular sample and period; not a causal estimate, a ChatGPT study or a universal current traffic-loss rate.
Read the original source ↗Product announcement · 2025-04-29
Mastercard unveils Agent Pay
Mastercard. Mastercard announced infrastructure for authorised AI-agent payments, including registered agents and consumer controls.
Scope. Evidence of infrastructure development, not the share of purchases made by bots. Authorised customer agents are different from malicious or unwanted traffic.
Read the original source ↗Threat assessment · 2025 assessment to 2027
Impact of AI on cyber threat from now to 2027
UK National Cyber Security Centre. NCSC assesses how AI may increase the efficiency and frequency of cyber activity and affect vulnerability exploitation and connected systems.
Scope. A threat assessment and forecast, not a measured loss probability for your business. Test your own workflows and exposure.
Read the original source ↗Published judgment · 2025-06-06
Ayinde and Al-Haroun: verification of legal authorities
Judiciary of England and Wales. The judgment addresses false legal citations and the responsibility of legal professionals to verify research and authorities.
Scope. Specific proceedings, not a general error rate for AI legal tools. The practical application here is a review workflow, not advice on a particular dispute.
Read the original source ↗Company-reported case · 2024-11-25
Klarna January–September 2024 results, page 3
Klarna. Klarna's own results presentation states that AI helped it reduce marketing-agency spending by 25%.
Scope. A historical claim from one purchaser. It is not an estimate of the revenue at risk for all agencies or proof that all external work moved in-house. The separate press-release page currently serves a bot check; the linked results PDF was accessible and the claim is on page 3.
Read the original source ↗Live product documentation · Checked 2026-10-22
Shopify: product discovery for agentic storefronts
Shopify. Shopify documents product feeds to AI channels with structured titles, descriptions, options, images, prices, availability and other attributes. It also documents product-data mapping and eligibility requirements.
Scope. Evidence that the mechanism exists on one commerce platform. It does not measure how often AI buyers reject products, prove incremental sales, or establish that the same integration is suitable for every B2B product. Availability varies by channel and eligibility.
Read the original source ↗Primary research paper by Salesforce AI Research · 2025-05-24
CRMArena-Pro: holistic assessment of LLM agents
Salesforce AI Research. In the paper's tests, leading agents achieved about 58% success in single-turn tasks and about 35% in multi-turn settings. The benchmark covered 19 expert-validated tasks across sales, service and configure-price-quote scenarios.
Scope. A May 2025 benchmark of the tested systems using synthetic business data. These figures are historical experimental results, not present-day model capability, live customer failure rates or a forecast for your operation. The practical lesson is to test changing instructions and complete workflows.
Read the original source ↗Regulator survey · 2024-11-21
Bank of England and FCA: AI in UK financial services 2024
Bank of England and FCA. Among 118 responding financial-services firms, the top three providers accounted for 73% of named cloud providers and 44% of named model providers. Forty-six per cent of respondents reported only partial understanding of the AI technologies they used.
Scope. A survey of UK-regulated financial-services firms. Provider shares refer to providers named in the survey, not global market share, spend or outage probability. Partial understanding is self-reported and does not itself demonstrate a control failure.
Read the original source ↗Peer-reviewed randomised experiment · 2023-07-13
Experimental evidence on the productivity effects of generative artificial intelligence
Science. 453 professionals completed occupation-specific writing tasks. The published abstract reports 40% less task time and 18% higher graded quality with ChatGPT access.
Scope. Short professional writing tasks in an online experiment, not complete jobs or organisations. A direct example of a 40% time reduction on defined work. It supports exploring substantial opportunities, not a universal company saving.
Read the original source ↗Peer-reviewed field experiment; publication of earlier working paper · 2026-03-11
Navigating the Jagged Technological Frontier
Organization Science. The study involved 758 consultants. On the within-frontier task set, AI users completed 12.2% more tasks and used about 25.1% less time; assessed quality improved. On the selected outside-frontier task, AI users were 19 percentage points less likely to reach the correct answer.
Scope. Specific consulting tasks and a historical GPT-4 configuration, not whole consulting engagements or enterprise capacity. Demonstrates why the work and its quality must be tested. The widely repeated 2023 'over 40%' figure referred to quality, not time released. The published article summarises quality/performance improvement as over 30%; avoid mixing versions.
Read the original source ↗Consultancy-reported deployments and illustrative activity accounting · 2026-09-16
The agentic transformation office: Redefining the economics of change
McKinsey. The article describes 35–40% time reductions on many core tasks and a recent deployment freeing more than 25 days. The existing TIP exhibit transcribes ten recurring activities totalling 65 to 39.5 effort-days monthly: 25.5/65 = 39.2% released, or five effort-days becoming 3.04.
Scope. Recurring transformation-office work, not all transformation activity or the enterprise. A relevant roughly-40% example grounded in reported implementation and a defined work mix. Describe its evidence type; do not present it as a controlled study or audited enterprise result.
Read the original source ↗Allocation model and separate anonymised company cases · 2026-07-24
How AI agents can help FP&A better steer the business
McKinsey. The existing model releases 50 points from activities, adds 20 to business steering and leaves 30 for allocation. Separately, the primary article reports more than 40% of FP&A capacity reallocated away from gathering data and manual reporting at a telecom company.
Scope. FP&A, not every finance activity. The allocation diagram and reported telecom result are distinct evidence. Shows why gross release, reinvestment and remaining capacity must be separated, and how a reported function result differs from a model.
Read the original source ↗Proposed 2030 operating model · 2026-08-04
Escaping the pilot trap: Building HR for the agentic era
McKinsey. The proposed allocation moves operational delivery from 40% to 5% and programmes from 30% to 10% of human time. The 55 points shift into 35 additional points of strategy and 20 of agent-capability management; the total stays 100%.
Scope. Future allocation of HR human time, not an observed 55% reduction in staffing or working hours. Supports the distinction between changing the contribution and taking people out. Agent ownership itself requires capacity.
Read the original source ↗Consulting-experience opportunity and separate company examples · 2026-06-22
From campaigns to continuous growth: AI capabilities shaping marketing
McKinsey. The article describes execution falling from 60–70% to as little as 10–15% of marketers' time, a potential 45–60 percentage-point shift. A separate company case reports 35–50% time savings in campaign activation.
Scope. Experience-based opportunity and selected workflow cases; no controlled whole-function study or common denominator across companies. Illustrates the scale of possible work redesign. Keep the allocation opportunity separate from observed campaign-activation claims and from cost reduction.
Read the original source ↗