AI Trends Report 2026: ten trends from the AI World Series finalists

At a glance

0 in 2

finalists deploy conversational AI

0 in 2

keep a person on every decision that matters

faster to put a new AI agent live on a shared platform

<0 min

to produce meeting minutes that took 1 to 2 days

lower production cost for an AI-generated campaign

<0 in 10

tested AI against a control group

In 2026 the question stopped being whether AI works. The leaders are now asking how to run it safely, at scale, and prove it paid off.

The ten trends

Ten shifts, and the move behind each

Tap a trend to see its move. The bar shows how many finalists already do it.

THE MOVEBuild identity, patterns and audit once, then reuse
THE MOVETurn approvals into pre-approved patterns
THE MOVELet AI interpret, code calculate, people decide
THE MOVEGive the bot transactions and clean handovers
THE MOVECodify brand rules into the pipeline
THE MOVEOffer staff a safe internal option first
THE MOVETreat language as a market, not a setting
THE MOVEPut your proprietary knowledge to work
THE MOVEMeasure capacity redeployed, not jobs cut
THE MOVESet a baseline before you switch it on

Who's in the data

This report draws on every finalist entry across the 2026 AI World Series, spanning all award categories. No organisation is named, and all figures are self-reported by entrants.

7 in 10 finalists came from the Middle East and Africa

Share of finalist organisations, rounded to the nearest 5%

Middle East & Africa70%
Europe15%
Latin America10%
North America5%

All finalist entries, AI World Series 2026 · rounded to the nearest 5%

The practices behind the trends

1 in 2 finalists use conversational AI; under 1 in 10 test it properly

Share of finalist organisations using each practice, rounded to the nearest 5%

45 points between the most common practice and proving it works

Hover or tap a practice to see which trend it drives.

Practices identified in finalist entry forms · an organisation can appear in several rows

Trend by trend

TREND 01

From pilots to platforms

of finalists run genuinely multi-agent systems

The most mature finalists stopped building AI one project at a time. They built a shared platform, then put use case after use case on it.

In a regulated business, every new AI project used to re-argue the same questions: architecture, security review, integration, audit logging and access. The leaders now settle those once.

35 → 20 weeks

to put a new AI agent into production at a European bank. Its platform gives every agent its own identity with minimal access rights, built-in testing and full audit trails. Around 30 use cases now run on it, freeing the equivalent of about 20 full-time roles.

9 agents

in an enterprise intelligence platform, each doing one job, with a dedicated “critic” agent that checks the others’ work before it moves on.

5 families

of agents by department (finance, retail, supply chain, customer, executive) at a Middle East fuel and convenience retailer, all drawing on one governed data layer.

Fewer than 1 in 5 finalists were running genuinely multi-agent systems. Most “agentic” claims in the pool were still early.

THE MOVE

Before scaling use cases, build the platform layer once: agent identity, approved patterns, testing and audit.

TREND 02

Governance becomes the accelerator

of finalists run a formal AI governance framework

Nearly 1 in 3 finalists run a formal AI governance framework. The best use it to go faster, not slower.

The old pattern made compliance a gate every project queued at. The new pattern makes it a co-designer of reusable building blocks.

Four ways governance sped things up

Pre-approved patterns

At the European bank in Trend 1, security, risk and compliance review a pattern once and every later use case inherits the verdict. That is what cut time to production by 40%.

Zero trust for agents

An AI platform vendor gives each agent least-privilege access, checks every action against guardrails, and can revoke an agent’s permissions if its behaviour slips.

Escalation by design

A tax and relocation advisory firm’s AI concierge always says it is an AI and answers only from approved content. It hands any sensitive topic (tax residence, source of funds, sanctions) to a qualified human, with rules mapped to regional data protection law.

Rules before launch

A Gulf bank published its responsible-use framework before switching on an AI assistant for every employee, on the principle that every AI output is a draft until a person checks it.

THE MOVE

Turn recurring approvals into pre-approved patterns, so governance becomes the reason you can scale.

TREND 03

Humans stay in the loop by design

of finalists keep a person on every key decision

1 in 2 finalists explicitly keep a person on every decision that matters. The sharpest split the work cleanly between AI, code and people.

Who does what

AI interprets, code calculates

A European bank uses a language model to read legally complex inheritance documents and extract 35 fields at about 95% accuracy. It deliberately hands the share calculations to deterministic code, because those must be exact and auditable.

AI drafts, people verify

A European life sciences company has AI draft answers for service agents to check. Resolution time halved and self-service resolution rose from 17% to 40%.

AI advises, people decide

A Middle East ports group added plain explanations of every speed recommendation for its vessels after captains said they would not trust a black box.

Clear clinical boundaries

A fertility app’s AI companion uses the couple’s own records to tailor answers, but is built never to diagnose and to say clearly when to see a doctor.

THE MOVE

For every AI workflow, decide in advance which parts AI interprets, which parts code calculates and which parts a person signs off.

TREND 04

Conversational AI becomes a service channel

of finalists deploy conversational AI

1 in 2 finalists deploy conversational AI. The leaders have moved it from answering questions to completing transactions and handing over cleanly to people.

27 million

conversations and 800,000+ transactions through a Middle East transport authority’s virtual assistant since launch, worth AED 80 million. Call-centre demand fell 70%.

28% → 85%

of non-Arabic, non-English speakers resolving their own travel queries at stations, after the same authority piloted a 15-language AI agent.

1 in 3

branch interactions handled by life-size AI avatars in a European bank’s branch of the future, giving bankers 20% more time for advice. The branch’s public rating rose from 3.8 to 4.6.

+4.5 pts

in first-contact resolution and +3.3 points in retention conversion at a North American energy retailer, after it gave agents real-time AI assist.

87% vs 81%

of AI-booked meetings went ahead, against those booked by bankers, at a European bank using an AI voicebot to book advisory appointments.

THE MOVE

Give your assistant real transactions and a structured handover summary, so a customer never has to start again with a person.

TREND 05

Generative content factories

of finalists produce generative content at scale

Fewer than 1 in 5 finalists use generative AI to produce content at scale, but those that do report the largest cost reductions in the pool.

The model that works is not a creative shortcut. It is a production line with brand rules built in and a human approving every output.

80%

lower production cost and 12 weeks down to 7 for a Gulf food brand’s fully AI-generated campaign: films, voiceovers and music. It beat its planned reach by 17%, and short videos held 98% completion on one social platform against a 90% benchmark.

75%

lower cost and 5x faster for a government business-education platform: 140 bilingual lessons presented by AI avatars trained on local presenters, plus an AI tutor. A second phase of 360 more videos followed.

500k+

engagements in six months for an insurer’s one-minute explainer videos that use AI presenters to make insurance terms understandable.

AI + CGI

at the same food brand: AI for creative scenes and traditional CGI for its packaging, so the product was always shown exactly right.

THE MOVE

Write your brand rules into the production pipeline and keep a person on final approval.

TREND 06

Sovereign and private AI

of finalists run private or on-premises AI

Fewer than 1 in 5 finalists run their AI on private or on-premises infrastructure. They also report some of the broadest staff adoption.

The logic is simple: give staff a safe, capable internal option and they stop pasting sensitive data into public tools.

<2 min

down from 24–48 hours to produce official meeting minutes at a Middle East transport authority. Its meeting assistant runs entirely on its own servers and uses an Arabic-capable open model, with 99% transcription accuracy across English and Arabic dialects and 12,500 staff hours saved in its first quarter.

9 countries

where a Gulf bank gave its whole workforce a privately hosted AI assistant from day one, with no data shared with outside AI providers. Staff report spending more than 10% less time on drafting, reports and analysis.

On device

behavioural signals processed on the customer’s own phone by an AI platform vendor, improving privacy and how quickly it picks up what customers want.

THE MOVE

Provide a governed internal AI option before shadow AI spreads, and decide early which data must never leave your walls.

TREND 07

Arabic-first and multilingual AI

of finalists build multilingual AI as a core feature

More than 1 in 3 finalists build multilingual AI as a core feature, and nearly 1 in 3 build specifically for Arabic.

For finalists in the Middle East, language is the market, not a setting.

+25%

renewals at a regional academic publisher whose AI assistant summarises and translates peer-reviewed Arabic scholarship for researchers worldwide. The assistant influenced nearly 40% of new institutional sales.

15 languages

including low-resource languages and colloquial speech, in a transport authority’s station assistant, built for the migrant workers and visitors the network serves.

3 languages

English, Arabic and Persian in a fertility app, so couples can describe intimate symptoms in the language they think in.

Minority languages

served in a European branch, where AI avatars speak to customers in languages bankers do not.

THE MOVE

Test your AI in the dialects your customers actually use, and treat language access as a growth market.

TREND 08

Your documents are the moat

of finalists point AI at their own documents and data

Nearly 1 in 3 finalists point AI at their own documents, transcripts and feedback, the knowledge no competitor can copy.

700,000

uses of a generative AI assistant that explains complex public tender documents to suppliers on a Latin American procurement marketplace, with 83% positive ratings. Complaints fell around 40% over two years.

1 in 10

vehicle sales now use a financing option an automotive retailer only discovered by analysing 400,000 customer feedback records with AI.

20% → 100%

of comments read by a payments client of a feedback platform, which then hit its highest-ever NPS. The same AI found customers of a global carmaker giving top NPS scores while still complaining in their comments.

3 in 5

product questions now resolved without an agent by a life sciences company’s AI product finder, built on its own manuals and service data.

THE MOVE

Inventory the documents, transcripts and feedback only you hold, and make them searchable and usable by AI before buying another generic tool.

TREND 09

Growth without matching headcount

of finalists aim AI at their own people first

More than 4 in 10 finalists aimed AI at their own people first. The story they tell is about capacity redeployed, not jobs cut.

4x clients, 2x staff

at a Latin American software company over three years. It took a new AI product from idea to paying customers in about six months, and is careful to say AI was one lever, not the only one.

1,500 hours

of staff time a month freed at stations by a transport authority’s AI agent and moved to safety, accessibility and complex help.

87%

internal AI adoption at a life sciences company that trained 40 regional change agents to bring hundreds of colleagues along.

Savings > cost

at a European bank that redesigned its back office for AI from scratch: staff savings consistently outweigh the full running cost of the AI, including model usage.

THE MOVE

Report capacity freed and where it went, fund change agents alongside the technology, and track AI running costs against the savings.

TREND 10

The evidence gap

of finalists ran a controlled test of AI impact

Fewer than 1 in 10 finalists tested AI against a control group, and nearly 1 in 3 leaned on projections, industry benchmarks or pre-launch estimates rather than measured results.

AI ambition is running ahead of AI measurement. Where finalists did measure properly, the numbers were compelling.

9.1% vs 2.1%

take-up of savings and investment products from a European bank’s AI-led campaign, more than four times the rate of an untouched control group.

$41.2m

in incremental revenue in 90 days for a North American telecoms client of a behavioural AI platform, in early 2026. Better timing of outreach alone accounted for nearly half of it.

3 measures

at once from a European life sciences company: service NPS up 7.9 points, online sales up 12% and $1.6 million saved, for the same programme.

The weakest evidence presented industry benchmarks as if they were the entrant’s own results, or described a platform that had not yet launched.

THE MOVE

Record a baseline before you switch AI on, keep a control group where you can, and report your own numbers, not the industry’s.

The results

Finalists reported reductions of 40% to 99%

Reduction after deploying AI, selected finalist results

Hover or tap a result to see the trend behind it.

The deepest cuts came where AI replaced slow, manual production or paperwork outright, rather than speeding up one step of it. Self-reported by finalists; measurement periods vary.

Regional lens

Each region brought its own flavour

Middle East & Africa70% of finalists

SIGNATURE THEMES

Government as an AI leader
Sovereign and on-premises AI
Arabic-first models
Generative content at scale
Proptech

IN ONE EXAMPLE

Meeting minutes in under 2 minutes, on the organisation’s own servers

What set the strongest entries apart

Five habits the strongest entries shared

1
They were in production, not in pilot.
The standouts reported months or years of live use at real scale, not a launch plan.
2
They measured their own baseline.
Before-and-after figures from their own operation, rather than benchmarks borrowed from elsewhere.
3
They engineered trust.
Explanations for recommendations, escalation rules, agent identities and audit trails were part of the build, not an afterthought.
4
They started from a business problem.
The best began by measuring the pain (call reasons, hours lost, documents delayed) and only then chose the AI.
5
They were honest about limits.
A security AI still partly fed by hand, a health app refusing to claim outcomes it had not yet proven, a growth story that credited AI as one lever among several. Candour made the claims credible.

What held entries back. Weaker entries described platforms with little actual AI, wrote results in the future tense, or submitted the same unchanged text to many categories.

The 2027 playbook

Ten moves for 2027. Tick the ones already on your plan.

0/10

on your plan

Based on all finalist entries to the AI World Series 2026, spanning every award category. Figures are self-reported by entrants and unaudited. No organisation is named.

THE 2027 EDITION IS OPEN

Doing any of these ten? Put it in front of the judges.

Next year's AI Trends Report 2027 will be written from the AI World Series 2027 finalists. Entries that show a baseline, a control and a commercial result score highest. Entry registration closes 28 May 2027.

How to enter Browse the categories

THE COMPANION REPORT

CX Trends Report 2026

The best CX programmes of 2026 no longer compete on how well they answer customers. They compete on how rarely customers need to ask.