GlhipLet's build
WHERE WE WORK

Fourteen sectors. One method.

Commercial AI does not fail on the model. It fails on the integration layer, on the data nobody has cleaned, and on whether the team changes what they do on a Tuesday. That is true in a hotel, a factory, a hospital and a bank. What changes between them is which systems hold the truth and who has to live with the answer.

Hospitality is where our work goes furthest down the stack, and it is why we can say the method travels. Decades of knowledge inside hotel systems, eighty-five integrations, fifty leading groups. Below is where else that method applies, what the published evidence actually shows, and which of our principals has done the work.

Every sector starts the same way, with your people. Training teams on frontier AI has moved from a nice-to-have to a requirement, and it is where most of our relationships begin.

Our hospitality practiceSee all seven offers

Every figure on this page names its source, its basis and its date.

The numbers below come from central banks, regulators, national statistics offices, intergovernmental bodies and peer-reviewed journals. None come from a consulting firm's marketing research, including our own.

Some of them are negative. A randomised trial found experienced developers were slower with AI, not faster. A field experiment found students given unguarded AI scored worse on exams than students with none. Purpose-built legal research tools still hallucinate on up to a third of queries. We publish those alongside the wins because they are the reason our work starts with enablement and evaluation rather than deployment.

Where a figure is an adoption rate rather than a measured outcome, we label it. Where a study is a preprint rather than peer reviewed, we label that too.

The labels a citation card can carry.

Adoption measurePreprintPeer reviewedRandomised trial

The sectors

A hotel revenue manager at a daylit back office desk with a rate calendar open, the restaurant pass behind the glass. Placeholder, generated.

Hospitality and Food & Beverage

What is possible
The commercial questions in a hotel are already written in systems of record: the PMS, the CRS, the RMS, the POS. AI earns its place when it reaches into those systems and changes a decision, not when it produces another dashboard. Rate and inventory decisions, guest messaging across languages, kitchen and procurement waste, and the long tail of back-office work that nobody has ever had budget to fix.
Where our experience is
This is the deepest ground we have. Decades building and operating hospitality technology, a guest-experience platform deployed across tens of thousands of rooms with eighty-five integrations behind it, and direct delivery with more than fifty leading groups. We know which integrations write back, which only read, and which are called supported and are not. One of our principals also served as Managing Director for Asia Pacific at a hotel technology company, so the vendor side of the table is familiar too.
Where we would start
With a cohort. Revenue, marketing and operations leaders spend a day applying frontier AI to the work they already do that week, and keep a portal afterwards that reports what actually stuck. Then one integration, scoped small, against a real system of record.
Evidence3 sources

  • AI waste-tracking cut food waste by 23% to 51% at four of five hotel, restaurant and catering sites, and cut the cost of wasted food per meal by up to 39%.

    Intervention study at five sites in Germany, Switzerland and Greece, measuring baseline waste per meal then re-measuring after installing computer-vision tracking. Honest caveat: the fifth site recorded a 13% increase.

    Sigala et al., Reducing food waste in the HORECA sector using AI-based waste-tracking devices, Waste Management (Elsevier, peer reviewed), May 2025. pubmed.ncbi.nlm.nih.gov

    Peer reviewed

  • Airbnb's customer-support cost per booking fell about 16% year on year, and close to 45% of issues that start with its AI assistant now resolve without a human agent.

    Quarterly shareholder disclosure for the three months to 30 June 2026. Airbnb attributes the cost decline only "in part" to AI, so this is not a clean AI-only effect.

    Airbnb, Inc., Q2 2026 Shareholder Letter, August 2026. s26.q4cdn.com

  • 21.3% of EU accommodation businesses with ten or more staff used at least one AI technology in 2025, against 12.0% across accommodation and food service as a whole.

    Official EU statistics. This is an adoption measure, not an outcome. The gap is the interesting part: hotels are moving roughly twice as fast as food service.

    Eurostat, dataset `isoc_eb_ain2`, Artificial intelligence by NACE Rev. 2 activity, reference year 2025, updated June 2026. ec.europa.eu

    Adoption measure

A retail merchandising manager on the floor of a bright fashion store, the product catalogue on a tablet in hand. Placeholder, generated.

Retail

What is possible
Retail has more clean, high-frequency decision data than almost any sector, which makes it one of the few places where AI effects can be measured honestly rather than asserted. Product content at catalogue scale, pre-sale conversation, search relevance, returns reduction, and dispute handling.
Where our experience is
Commercial strategy is a named layer on every Glhip engagement, built on peer and benchmark modelling, unit economics and KPI design, with a background in key-account strategy and private equity. Behavioural design is a second named layer, which matters more in retail than most sectors: the difference between a recommendation engine that works and one that does not is usually what the store manager does with it.
Where we would start
Pick one workflow with a measurable conversion or returns outcome, build the evaluation before the build, and train the merchandising team to run it. The published evidence is unambiguous that gains vary enormously by workflow, so choosing the workflow is the whole job.
Evidence3 sources

  • Across seven randomised field experiments on a large cross-border e-commerce platform, a generative AI pre-sale chatbot raised sales 16.3% and conversion 21.7%. Other workflows in the same study ranged down to no detectable effect.

    Samples ranged from 44,614 consumers to 13.7 million. Product descriptions cut returns 3.9%. Two workflows, marketing push and ad titles, were not statistically significant. Preprint, not yet peer reviewed, and the platform is not named.

    Columbia Business School, Zhejiang University and Zhejiang University of Finance & Economics, Generative AI and Sales Productivity: Field Experiments in Online Retail, June 2026. arxiv.org

    Randomised trialPreprint

  • 18.6% of EU wholesale and retail businesses with ten or more staff used at least one AI technology in 2025; retail excluding motor vehicles was 15.5%, against a 20.0% all-sector average.

    Adoption measure, not an outcome. Retail is behind the economy-wide rate.

    Eurostat, dataset `isoc_eb_ain2`, reference year 2025, updated June 2026. ec.europa.eu

    Adoption measure

  • Around 14% of US retail businesses reported currently using AI, with about 17% expecting to within six months.

    Biweekly nationally representative survey of US businesses, reference date 3 May 2026. Note the survey question changed in November 2025, so this does not compare to earlier readings.

    US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, May 2026. census.gov

An airline commercial manager at a desk in a daylit airport operations office, a route map on the laptop and the apron through the glass. Placeholder, generated.

Travel and Aviation

What is possible
Aviation is the sector where AI has produced some of the most rigorously measured operational results anywhere, because the industry already instruments everything and regulates what it cannot measure. Flight planning and fuel burn, contrail and emissions avoidance, controller workload, disruption and rebooking, and multilingual service at scale.
Where our experience is
Distribution and commercial depth. One of our principals spent years in global online travel commercial finance and key-account strategy, which is the layer where airline, OTA and hotel economics actually meet. Add the systems knowledge from the hospitality side, since the distribution stack is largely shared.
Where we would start
Almost never with the flight operations layer, which is heavily regulated and already well served. We would start with commercial and service work, and with training the teams who own it, because that is where the unclaimed ground is.
Evidence3 sources

  • A generative AI upgrade to Scoot's customer-service chatbot produced a 19 percentage point improvement in its customer satisfaction scores.

    Reported for the financial year to 31 March 2026. The group reports more than 550 generative AI use cases identified and over 140 implemented. The CSAT sample and instrument are not stated in the report.

    Singapore Airlines Limited, Annual Report FY2025/26, May 2026. singaporeair.com

  • A randomised controlled trial of AI-guided contrail avoidance cut the contrail formation rate 11.6% across 1,232 eligible flights, with no significant difference in fuel burn.

    Dispatcher-led, embedded in normal flight planning, validated against satellite imagery. Among the 112 flights that actually flew the plan as filed the reduction was 62%, but that is 112 flights and the honest headline is 11.6%. Preprint, not yet peer reviewed.

    Google Research, American Airlines and Imperial College London, Efficacy of Scalable Airline-led Contrail Avoidance, March 2026. arxiv.org

    Randomised trialPreprint

  • With speech recognition pre-filling radar labels, air traffic controllers had wrong or missing entries on 4% of commands against 11% without it, and mouse-clicking time fell from 12,800 seconds to 400.

    Simulation with twelve Vienna approach controllers over 28 hours. Simulation rather than live operations, and published 2023.

    SESAR 3 Joint Undertaking (EU institutional body), Reduced flight time and fuel burn on the cards with controller voice recognition tool, March 2023. sesarju.eu

An event director with a floor plan on a tablet in an exhibition hall during set up. Placeholder, generated.

Events and MICE

What is possible
Exhibition and event businesses run on three things AI is genuinely good at: matching attendees to content and to each other, producing enormous volumes of multilingual material against immovable deadlines, and turning post-event data into something a sponsor will pay for next year. The honest position is that this sector publishes adoption enthusiastically and outcomes almost never.
Where our experience is
We run events as a business line, not as a theory. Glhip Learning delivers hands-on cohorts monthly across the region, which means the operational problems of an events business are ones we solve for ourselves every month: room logistics, multilingual material, participant tracking, and proving afterwards that something actually landed.
Where we would start
With the content and matching workload, since that is where the volume is, and with training the event team to run it themselves. An events business that has to call an agency every time it needs an asset has bought a dependency, not a capability.
Evidence3 sources

  • 91% of exhibition-industry companies worldwide report using AI, up four points in six months, and 62% are testing or implementing AI specifically for company and process efficiency, up ten points. Thailand ranks second worldwide, at 67%, for companies testing or implementing AI to improve customer experience.

    Survey of 466 companies across 59 countries, concluded June 2026. Adoption measure, not an outcome.

    UFI, The Global Association of the Exhibition Industry, 37th UFI Global Exhibition Barometer, July 2026. ufi.org

    Adoption measure

  • 93% of incentive travel professionals report their team using AI tools for incentive travel tasks, with content creation at 61% and destination research at 51% the leading applications.

    2,708 respondents across 85 countries, fielded May to July 2025. Adoption measure.

    Incentive Research Foundation with the SITE Foundation and Oxford Economics, The IRF 2026 Trends Report, January 2026. theirf.org

    Adoption measure

  • 91% of business-events professionals reported using AI in some way, with 15% classed as leaders and 20% lagging; the top concern was data security and privacy at 59%.

    Survey of 92 professionals, December 2024. Small sample, and the survey partner is an event-technology vendor, though the published figures are adoption rates rather than product claims.

    PCMA, AI Pulse Check: How Event Planners Are Using Gen AI, March 2025. pcma.org

    Adoption measure

All three figures above are adoption rates. We searched government, intergovernmental and peer reviewed sources for a measured outcome in this sector and found none, so we say that rather than dress an adoption rate up as a result.

A senior software engineer at a standing desk in a daylit studio, code and a results dashboard on the monitor. Placeholder, generated.

Technology and Software

What is possible
This is the sector with the most evidence and the least agreement. Rigorous studies point in opposite directions depending on who is using the tool, how experienced they are, and whether anyone measured properly. That is not a reason to avoid the sector. It is the reason the work has to be instrumented.
Where our experience is
Platform and integration is a named layer, run by a principal who co-founded a Bangkok technology consultancy specialising in AI, cloud and digital transformation, and who holds a Google Cloud partner practice. Agentic systems engineering is a second named layer. Both flagship Build offers, MCP and Agent Platforms and Glhip Model Factory, come out of this ground.
Where we would start
With an evaluation harness before a rollout. The findings below show that self-reported productivity and measured productivity diverge sharply, and that senior engineers gain far less than junior ones. Any programme that skips measurement is buying a perception.
Evidence3 sources

  • In a randomised controlled trial, experienced open-source developers took 19% longer to complete real tasks when allowed to use early-2025 AI tools, while believing AI had made them 20% faster.

    Sixteen experienced developers, 246 real issues on repositories they had contributed to for around five years. Confidence interval +2% to +39%. The researchers have since said their follow-up experiment gives an unreliable signal, so treat this as a 2025 snapshot rather than a standing truth.

    METR (independent non-profit), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025. metr.org

    Randomised trial

  • Across three randomised controlled trials covering 4,867 developers, an AI coding assistant produced a 26.1% increase in completed tasks, with gains concentrated in junior developers at 27% to 40% against 7% to 16% for seniors.

    Peer reviewed. Trials ran 2022 to 2023 on autocomplete-era tooling, and two of six authors were employed by the tool vendor.

    Cui, Demirer, Jaffe, Musolff, Peng and Salz, The Effects of Generative AI on High-Skilled Work, Management Science, 2025. economics.mit.edu

    Randomised trialPeer reviewed

  • UK government developers in a cross-government trial reported saving 56 minutes per working day, while the objectively measured code suggestion acceptance rate was 15.8%.

    1,273 licences across 50-plus organisations over three months. The 56 minutes is self-reported by 424 respondents and the report itself warns of optimism bias. The 15.8% is the defensible hard number.

    Government Digital Service, UK Department for Science, Innovation and Technology, AI coding assistant trial: UK public sector findings report, September 2025. gov.uk

A grid operations analyst in a daylit control room, forecast screens behind him and a rooftop solar array through the window. Placeholder, generated.

Sustainability

What is possible
Forecasting is where AI has delivered its least disputed wins, and most sustainability problems are forecasting problems wearing a different hat. When you can predict generation, demand, weather or flow more accurately, you buy less headroom, spill less, and dispatch less standby. The gains are physical and they are audited by people who have no commercial interest in the answer.
Where our experience is
One of our principals spent years as Chief Operating Officer of a company producing sustainable oils and proteins for the cosmetics, pet food, alternative protein and nutrition markets, having joined as Factory Manager, with responsibility for production, health and safety and finance across a site of around 250 people. Sustainability claims that survive audit come from people who have had to make them about their own operation. Cloud cost and energy optimisation sits with the platform layer.
Where we would start
With measurement integrity, then forecasting, then training the operations team to act on a forecast they did not produce. The third step is where most of these programmes quietly die.
Evidence3 sources

  • An AI flood forecasting model matched or beat the incumbent global system's same-day accuracy at up to five days of lead time, in watersheds with no gauges at all.

    Peer reviewed, trained and tested on 5,680 gauges representing 152,259 gauge-years, benchmarked against the EU Copernicus system. Now operational and free in over 80 countries. The Southwest Pacific was the highest-scoring continent in the evaluation.

    Nearing et al., Global prediction of extreme floods in ungauged watersheds, Nature, March 2024. nature.com

    Peer reviewed

  • An intergovernmental weather centre's operational machine-learning forecast delivers medium-range error reductions typically of 5% to 15% against its physics-based system, using roughly 1,000 times less energy per forecast.

    Full-year 2025 operational verification. Tropical cyclone position errors reduced substantially, though intensity errors were larger, which is a real limitation for the region.

    European Centre for Medium-Range Weather Forecasts, Forecast performance 2025, ECMWF Newsletter No. 187, May 2026. ecmwf.int

  • A deep learning solar nowcasting model cut a national grid operator's solar forecast error from 650 MW to 233 MW mean absolute error, 2.8 times better, for forecasts up to two hours ahead.

    Delivered into the control room of a statutory electricity system operator. The source describes carbon and cost benefits qualitatively and publishes no figure for either, so we do not imply one.

    National Energy System Operator (UK statutory public corporation), Solar NowCasting innovation project improves solar forecasting, July 2024. neso.energy

A data scientist beside a whiteboard of evaluation diagrams, a model comparison dashboard on the laptop. Placeholder, generated.

AI and Data Science

What is possible
The sector that sells AI is also the one most exposed to the gap between what AI is believed to do and what it measurably does. The work here is evaluation, data quality, model routing and the unglamorous plumbing that decides whether anything else works.
Where our experience is
All three practices meet here. Glhip Model Factory routes each task to the cheapest model that clears a quality bar built from the client's own workload, which is only possible because the evaluation comes first. MCP and Agent Platforms makes existing software operable by agents. Glhip Learning exists because none of it lands without people who can use it.
Where we would start
By building the eval set before anything else, from real tasks in the client's own workload. It is the least exciting deliverable we produce and the one that decides whether the rest of the programme can be defended in six months.
Evidence3 sources

  • Access to a generative AI assistant raised customer-support productivity 15%, measured as issues resolved per hour, with gains concentrated in less-experienced agents and small quality declines among the most experienced.

    Peer reviewed. Phased rollout across 5,172 agents at a Fortune 500 software firm, the majority of them working in the Philippines.

    Brynjolfsson, Li and Raymond, Generative AI at Work, The Quarterly Journal of Economics, April 2025. digitaleconomy.stanford.edu

    Peer reviewed

  • Between 17% and 20% of US businesses reported using AI in any business function, rising to 37% among firms with 250 or more employees and falling below 20% among the smallest.

    Biweekly survey of a sample drawn from around 1.2 million US employer businesses, December 2025 to May 2026. Adoption measure. Scale is the strongest predictor of adoption.

    US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, May 2026. census.gov

    Adoption measure

  • In a randomised controlled trial, developers forecast a 24% speed-up, experienced a 19% slowdown, and afterwards still reported a 20% speed-up.

    The single most useful finding in this document. Perceived productivity is not evidence of productivity, and a programme measured by survey is measuring the wrong thing.

    METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025. metr.org

    Randomised trial

A building operations manager with a tablet in a bright workspace lobby of glass, stone and plants. Placeholder, generated.

Real Estate and Workspaces

What is possible
Buildings are the most heavily instrumented assets most organisations own and the least interrogated. Plant optimisation, demand and occupancy forecasting, maintenance scheduling and lease and document work. The measured savings are real, in single to low double digits, and the literature is unusually honest that most published results were badly measured.
Where our experience is
One of our principals co-founded a Bangkok property services business covering search, deal management, legal coordination, ownership transfer, renovation and ongoing property management, so the operational reality of running property here is first-hand rather than theoretical. Market entry, due diligence and cross-border structuring sit with the governance layer.
Where we would start
With the building management system and what it actually records, which is almost always less than the specification claims. Then a single plant or a single portfolio process, measured against a baseline agreed before the work begins.
Evidence3 sources

  • Across 104 field demonstrations in real occupied buildings, AI-based HVAC control produced duration-weighted average energy cost savings of 13% in commercial buildings and 16% in residential.

    Peer-reviewed systematic review. The authors found 71% of the studies used protocols liable to produce unreliable estimates; those figures come from the 29% judged reliable. Only 13 of 104 papers reported deployment cost, so this is a savings figure, not a net return.

    Purdue University, University of Colorado Boulder and NREL, Lessons learned from field demonstrations of model predictive control and reinforcement learning for residential and commercial HVAC, Applied Energy, December 2025. arxiv.org

    Peer reviewed

  • Hong Kong's government engineering department reported 5% to 10% chiller plant energy savings at AI-optimised pilot sites, with a public hospital plant improving average coefficient of performance from 3.52 to 3.74.

    Neural networks predicting cooling demand plus particle-swarm optimisation, retrofitted onto existing plant from January 2022. The plant cannot run both modes at once, so savings were measured against a benchmark model built from historical data rather than a direct comparison.

    Electrical and Mechanical Services Department, Government of the Hong Kong SAR, Chiller Energy Optimization using Artificial Intelligence: Experience in Hong Kong, January 2024. emsd.gov.hk

  • A US federal property agency found machine-learning optimisation of air handling units saved between 5% and 11% of whole-building energy across four office buildings, while the directly measured result was demand forecasting accurate to within 5%.

    Independent field validation by a national laboratory. Pandemic ventilation guidance blocked the optimised start-up sequences, so savings were modelled rather than directly measured, and neither site met the agency's five-year payback target. Published 2022.

    US General Services Administration, evaluated by the National Renewable Energy Laboratory, Energy Management Information System with Automated System Optimization, September 2022. gsa.gov

A hospital administrator at a clinic administration counter with a scheduling grid on a tablet. Placeholder, generated.

Healthcare and Medical Services

What is possible
Healthcare holds the highest-quality AI evidence of any sector, because the field will not accept a claim without a trial. Screening and diagnostic support, risk stratification that changes clinical decisions, and the documentation burden that drives clinician attrition. The trials show real effects, and they also show that partial adoption produces partial results.
Where our experience is
This is a sector we approach through governance first. Multi-jurisdiction privacy including PDPA and GDPR, AI governance charters, data protection and procurement all sit with named layers. One of our principals has run production in a regulated environment with responsibility for health and safety across a 250-person site. We would be candid in the room that we are not a clinical informatics firm, and that anything touching a clinical decision needs a clinical partner alongside us.
Where we would start
In administration, not diagnosis. Documentation, scheduling, coding, correspondence and multilingual patient communication carry real burden, sit outside the regulated clinical decision path, and are where enablement lands fastest.
Evidence3 sources

  • In a randomised trial of 105,934 women in a national screening programme, AI-supported reading detected 6.4 cancers per 1,000 screened against 5.0 per 1,000 with standard double reading, while cutting screen-reading workload 44.2%, with no significant increase in false positives.

    Randomised, controlled, parallel-group, single-blinded. The comparator is routine double reading by two radiologists, not a single reader.

    Lund University and Skåne University Hospital, Screening performance and characteristics of breast cancer detected in the MASAI trial, The Lancet Digital Health, March 2025. doi.org

    Randomised trial

  • In a pragmatic randomised trial of 15,965 patients in Taiwan, an AI electrocardiogram mortality-risk alert sent to treating physicians reduced 90-day all-cause mortality from 4.3% to 3.6%.

    Hazard ratio 0.83. Thirty-nine physicians randomised to receive or not receive the alerts, across emergency, outpatient and inpatient settings. Single hospital system, so generalisability is untested.

    Tri-Service General Hospital and National Defense Medical Center, Taipei, AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial, Nature Medicine, April 2024. nature.com

    Randomised trial

  • In a three-arm randomised trial across roughly 48,000 outpatient visits, one ambient AI scribe cut physician time in notes by 9.5% while the other showed no significant change.

    238 physicians across 14 specialties, measured objectively from electronic health record audit logs rather than self-report. Adoption was partial, at 33.5% and 29.5% of visits. Burnout scores improved for both tools. Not vendor funded.

    UCLA David Geffen School of Medicine and RAND Corporation, Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI, December 2025. doi.org

    Randomised trial

A plant quality engineer on a clean daylit manufacturing floor, inspection images on a tablet. Placeholder, generated.

Manufacturing and Industrial

What is possible
Visual quality inspection, predictive maintenance, yield optimisation, scheduling and the supply chain around them. The gap between what a well-instrumented plant achieves and what the average firm reports is larger here than in any other sector, and understanding why is most of the consulting value.
Where our experience is
The most direct operating credential we hold. One of our principals joined a manufacturer as Factory Manager and rose to Chief Operating Officer, with responsibility for production, human resources, health and safety and finance across a site of around 250 people, and was previously Logistics and Enterprise Resource Planning Manager at a Thai manufacturer. ERP and integration work sits with the platform layer.
Where we would start
On the shop floor, with the people who will have to trust the model. The evidence below is blunt about this: instrumented plants report transformational numbers while the economy-wide realised gain is close to zero. The difference is not the model.
Evidence3 sources

  • A national industry ministry reported that 230 flagship smart factories recorded defect rates 50.2% lower, production efficiency 22.3% higher, development cycles 28.4% shorter and carbon emissions 20.4% lower than pre-transformation baselines.

    Covering plants across 31 provincial regions and over 80% of manufacturing sectors, deploying nearly 2,000 scenarios including AI-powered quality inspection. Important caveat: the comparison period and method are not stated, and these figures cover smart-factory transformation as a bundle, with AI inspection as one named component.

    Ministry of Industry and Information Technology, People's Republic of China, reported February 2025. govt.chinadaily.com.cn

  • Across roughly 6,000 senior executives in the US, UK, Germany and Australia, AI is reported to have raised productivity by around 0.29% over three years, with nine in ten reporting no own-firm impact at all. The same executives expect 1.4% over the next three years.

    Four national firm panels. 69% of firms actively use AI. Self-reported by executives rather than independently measured.

    National Bureau of Economic Research, Firm Data on AI, Working Paper 34836, February 2026, revised March 2026. nber.org

  • Manufacturing showed the strongest year-on-year growth of any US sector in work-related generative AI adoption, up about 58%, or 14.5 percentage points.

    Central bank analysis triangulating three separate surveys to November 2025. Adoption measure. The note explicitly does not measure output or productivity effects.

    Board of Governors of the Federal Reserve System, Monitoring AI Adoption in the U.S. Economy, April 2026. federalreserve.gov

    Adoption measure

A 50.2% defect reduction in flagship plants and a 0.29% realised economy wide gain are not contradictory. They are the whole argument. Measured gains in tightly scoped, well instrumented settings are large. Diffuse gains from unfocused adoption are close to nothing.

A training director at the front of a daylit seminar room, participants at laptops behind him. Placeholder, generated.

Education and Trade Associations

What is possible
The evidence here is the most directly relevant to what Glhip sells, and the most double-edged. Well-designed AI tutoring produces large, replicated learning gains. Unguarded AI access produces worse outcomes than no AI at all. The design of the intervention is not a detail, it is the entire result.
Where our experience is
Enablement is the practice most of our relationships start with. Glhip Learning runs hands-on cohorts monthly across the region for non-technical leaders, built around the work people already do that week, with a private portal afterwards that reports skills completed and work shipped rather than seats filled. Every lesson ships as an installable skill. On the association side, one of our principals is President of a bilateral chamber of commerce in Bangkok and a faculty member at Chulalongkorn University's Faculty of Law, so member-organisation governance is lived experience rather than a target market.
Where we would start
By designing the guardrails into the programme rather than adding them afterwards. The third study below is the reason. An organisation that hands its people an unguarded assistant and calls it enablement has run the control arm of an experiment that has already been done.
Evidence3 sources

  • A meta-analysis of 49 controlled experiments found AI-assisted learning improved outcomes with a pooled effect size of 0.449, after correcting for publication bias.

    Randomised and quasi-experimental designs comparing AI-assisted learning against traditional instruction, covering achievement, attitude and motivation. Effects varied by educational level.

    Liu, Meng, Zhu and Rong, Does AI-assisted Learning Improve Students' Learning Outcomes? Evidence From A Meta-analysis, The Asia-Pacific Education Researcher (Springer), March 2026. link.springer.com

  • In a randomised crossover trial, students learned more from an AI tutor than from in-class active learning, by 0.63 standard deviations, in less time.

    194 students, each experiencing both conditions in consecutive weeks. Median 49 minutes against a 60-minute class. The tutor was purpose-built by the research team, in a single course at one elite university, so generalisability is limited.

    Kestin, Miller, Klales, Milbourne and Ponti, AI tutoring outperforms in-class active learning, Scientific Reports (Nature Portfolio), June 2025. pmc.ncbi.nlm.nih.gov

    Randomised trial

  • Students given an unguarded AI tutor scored 48% better on practice problems and 17% worse than the no-AI control on exams once access was removed. A safeguarded tutor that gave hints rather than answers produced a 127% practice gain and eliminated the exam penalty.

    Field experiment with nearly 1,000 students, four 90-minute sessions. Working paper, not yet peer reviewed. This is the finding that should shape any enablement programme.

    Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman, The Wharton School, Generative AI Can Harm Learning, 2024. static1.squarespace.com

All three studies are education. We found no published evidence specific to trade associations or chambers of commerce from a source we would cite, so the association credential stands on its own.

A financial crime operations lead at a desk in a bright bank office, a transaction network graph on the monitor. Placeholder, generated.

Financial Institutions

What is possible
Financial services has the deepest regulatory scrutiny of AI anywhere, which means the sector's own regulators have done the measurement. Financial crime detection, payments integrity, supervisory reporting and the governance apparatus around all of it. The published results are strong on detection and blunt about the false-positive cost that comes with it.
Where our experience is
Commercial strategy on this engagement is held by a principal whose background includes foundation models for finance, global online travel commercial finance and key-account strategy, and private equity. Governance and risk, including multi-jurisdiction privacy and AI governance charters, is a separate named layer. We would be direct that we are not a regulated financial services adviser.
Where we would start
With the false-positive economics, because that is where these programmes are won or lost. A detection lift that triples the review queue has not saved anyone anything, and the source below publishes exactly that trade-off rather than hiding it.
Evidence3 sources

  • Sharing payment-system-level network analytics with banks found 12% more illicit accounts on average than bank-only detection, rising to 26% for financial crime patterns not previously seen.

    Machine learning on a fully synthetic dataset of 1.8 million accounts and 308 million transactions. The trade-off is published honestly: flagging the top 0.2% riskiest accounts gave a 5% false-positive rate, while pushing recall to 51% carried a 73% false-positive rate. No real customer data was used.

    Bank for International Settlements Innovation Hub with the Bank of England, Project Hertha: identifying financial crime patterns in real-time retail payment systems, June 2025. bis.org

  • A national treasury recovered $1 billion of check fraud in one fiscal year using machine learning, within more than $4 billion in total fraud and improper payments prevented and recovered, up from $652.7 million the year before.

    Only the $1 billion is explicitly attributed to machine learning, so the full $4 billion should not be. Methodology for calculating prevented amounts is not stated.

    US Department of the Treasury, October 2024. home.treasury.gov

  • Australia's financial regulator identified 624 AI use cases in use or development across 23 licensees, and found nearly half lacked policies on consumer fairness or bias.

    Targeted regulator review during 2024, not a representative sample. The regulator's own headline was the governance gap, not the adoption number, and the page should present it that way.

    Australian Securities and Investments Commission, REP 798 Beware the gap: Governance arrangements in the face of AI innovation, October 2024. asic.gov.au

A content producer at an edit desk in a daylit production studio, a video timeline on the monitor and a cinema camera behind. Placeholder, generated.

Media and Content Production

What is possible
Production volume at a cost that changes what is worth making, localisation into markets that were previously uneconomic, and archive and metadata work that has sat undone for decades. Alongside that, an accuracy and rights problem that is measured, serious and not going away.
Where our experience is
AI Media and Marketing Production is a standing offer: on-brand images, short-form video, avatars and ad creative generated with frontier models and packaged for distribution. We produce our own material this way, including the placeholder imagery in our own design work, which we label as generated rather than pass off as photography. Rights, licensing and IP transfer sit with the governance layer, run by a principal with three decades of licensing and contracts experience.
Where we would start
With provenance and rights before volume. The second and third findings below are why. A production pipeline that cannot say where an asset came from is a pipeline that will eventually have to stop.
Evidence3 sources

  • Weekly use of AI chatbots for news rose from 7% to 10% globally in a year, reaching 17% among 18 to 24 year olds.

    Survey of 97,520 respondents across 48 markets, around 2,000 per market, including ten Asia Pacific markets. Online panels, so offline populations are under-represented.

    Reuters Institute for the Study of Journalism, University of Oxford, Digital News Report 2026, June 2026. reutersinstitute.politics.ox.ac.uk

  • 45% of AI assistant answers about news contained at least one significant issue, with 31% showing serious sourcing problems and 20% containing major accuracy issues.

    Structured evaluation by professional journalists at 22 public service media organisations across 18 countries and 14 languages, scoring more than 3,000 responses. The worst-performing assistant had significant issues in 76% of responses.

    European Broadcasting Union with the BBC, News Integrity in AI Assistants, October 2025. ebu.ch

  • A national copyright office has registered more than 7,000 claims that include AI-generated material where that material was disclaimed under its registration guidance.

    Stated in sworn Senate testimony by the Register of Copyrights. Counts registrations granted since March 2023, not works created, with no breakdown by medium.

    United States Copyright Office, Testimony of Shira Perlmutter before the Senate Judiciary IP Subcommittee, May 2026. copyright.gov

Your sector is not the hard part. Your systems are.

If your industry is not on this list, the method still applies. We start where the data actually lives, we build the evaluation before the build, and we train the people who will own it afterwards. What we will not do is claim depth we do not have. If your problem needs a specialist we are not, we will say so in the first conversation.

Let's buildHow we work