69 points of view from our industry teams — filter by industry or topic, or search. Figures in each article are sourced or labelled.
Industry
Topic
69 insights
Generative AI assistants made individual bankers faster. Agents that execute whole workflows under human supervision are what move the cost base — and they need a different operating model to do it safely.
2026 is the year programmable money moved from pilots to live value: real-value cross-border trials among central banks, bank deposit tokens on public chains and a US stablecoin law about to bite. Banks need a position on all three forms of money, not a bet on one.
The open-source record shows where AI in financial services is really being built: agent frameworks and finance reference agents, data connectors, and a thin but important layer of governance. Stars are loud; the signals that matter for banks are quieter.
Trading, discom operations, generation, finance and regulatory teams each rebuild the same forecasts and market data. A data product factory and an internal marketplace — on AWS, Azure or on-premise — turn that duplication into governed products with owners, contracts and subscriptions.
Short-term power trading is growing, the Real-Time Market is its fastest-growing segment, and market coupling, derivatives and deviation-settlement reform are changing the rules. Desks that win will be the ones whose forecasts, positions and settlements are trusted by gate closure.
Open models, agent frameworks and MCP connectors are now part of almost every bank’s AI stack. The licence, supply-chain and regulatory questions they raise are answerable — but only if intake is designed for AI, not just for software.
The easing cycle is over almost before it finished, and rates are rising again. The lesson of the past three years is that margin is now earned customer by customer, and that the primary relationship is the only deposit franchise worth defending.
FMCG makers know what they ship to distributors. What sells on to millions of kirana stores arrives late, in a dozen formats, and is argued about at month-end. Joining distributor, outlet and claim data into one governed record changes what a sales team can do on Monday morning.
Grievance portals made it easy to complain. Reading, routing and answering every grievance fairly is still mostly manual. Triage agents can do the reading and routing — while grievance officers keep every reply.
RERA asks developers to keep 70% of what buyers pay for a project in a separate account and to withdraw only against certified progress. Most of the effort goes into matching receipts, bookings and bank lines by hand. One governed model of buyers, projects and bank accounts turns it into a daily reconciliation with people deciding the exceptions.
Parents no longer judge their India centres on headcount and rate cards. They want outcomes they can see — products owned, controls that hold, AI that pays back. That shift is as much a data problem as a talent one.
Voluntary alliances have dissolved and US supervisors have stepped back, yet European and UK prudential expectations have hardened. For banks, climate is becoming a risk-management and data discipline rather than a public commitment, and financed-emissions data remain the weakest link.
By the time warranty cost shows up in the P&L, the failure has been in the field for months. Joining claims, service reports, build records and supplier data turns warranty from an accounting line into an early-warning system — provided people keep the quality decisions.
In a hospital, a fluent but wrong answer is worse than no answer. The assistants worth deploying answer only from the record, show where every number came from, and leave the decision with a clinician or a named owner.
Smart prepaid metering gives Indian discoms the data to see where energy goes missing, from the feeder to the consumer. The losses fall when that data is joined to billing, the network and the vigilance team's list — and when officers, not algorithms, decide who is inspected.
One wrong field on a shipping bill or a bill of entry can hold a container for days. Extraction, classification suggestions and pre-filing checks can catch most mismatches before submission — as long as a licensed broker still decides and files.
Embedded finance keeps growing, but the US sponsor-bank enforcement wave showed that when a partnership fails, the bank's ledger and licence are what customers fall back on. The winners will be banks that can prove, every day, whose money is where.
Banks are more profitable than they have been in years, but their cost bases have hardly changed relative to assets. The next step change will come from redesigning work around AI, not from another round of cost containment.
Brand, regulatory and competitor signals in India don't arrive in English alone. An LLM that scores every mention — sentiment, brand, region, impact — and joins it to sales in that state turns news into something a brand team can act on.
Direct benefit transfer took cash out of the middle. What remains hard is the beneficiary list itself — duplicates, ineligible entries and failed payments found late. Data and agents can flag them; officers must decide.
Delays and cost overruns rarely arrive as a surprise; they accumulate in daily progress reports, site photos, material receipts and contractor bills that nobody reads together. Reading that data on one model gives project directors weeks of warning — and checks bills before they are paid.
Banks spend heavily on financial-crime compliance and still catch only a sliver of illicit flows. Agentic AI offers a different shape for the work — squads of narrow agents assembling evidence, with investigators making the call.
Banks earned more in 2025 than any industry on earth, yet investors still price them at the bottom of the league. The next leg of returns will come from productivity and precision, not balance-sheet heft.
Data-sharing regimes are widening from payment accounts to savings, pensions and insurance, but the rules are moving at very different speeds. The economics of access, not the APIs themselves, will decide who wins.
Parts go end-of-life inside decade-long programmes, certificates expire unnoticed and single-source suppliers wobble. In our delivered work, the risk was rarely in one system — it sat between PLM, ERP, quality and supplier records.
GeM and e-procurement digitised buying, but evaluation is still a pile of long tenders and scanned bids. Agents can extract requirements and line items — if every value links back to the page it came from.
Service desks measure SLAs after the fact. The same ticket data, read by a handful of well-chosen models, tells a team lead which tickets will breach while there is still time to act.
Road freight margin leaks in two places: the empty return leg and the costs nobody reconciles. Telematics, FASTag records and e-way bills already describe every trip — joined per consignment, they let fleet teams fix both while the truck is still on the road.
Finance teams still spend most of their time assembling numbers rather than explaining them. A continuous close, with agents reconciling and drafting commentary as transactions land, turns the finance function from reporter into navigator.
The months around possession — snags, registration, defect claims, the first resident tickets — shape how buyers talk about a developer for years. Treating units, snags and tickets as one record, and triaging them with agents, shortens possession and keeps facility teams ahead of complaints.
Card networks, big tech and standards bodies are building the rails for AI agents to transact on consumers' behalf. For banks, the shift rewrites the rules of authentication, liability, distribution and brand, and it is arriving faster than consumer trust.
Generative AI is shortening credit memos and cash-flow data is widening who can be scored. The rules on explaining and testing those decisions are shifting on both sides of the Atlantic, but the obligations have not gone away.
Cases and arbitrations against government run for years across courts and tribunals. Without one record, exposure and ageing are guessed. A single analytics view — and actions tracked to closure — changes that.
Emergency queues and cancelled operations are often symptoms of a discharge problem. Hospitals that see free beds, discharges due and requests waiting on one board can act hours earlier.
Asset owners and REITs answer to unitholders, lenders and regulators on numbers assembled from leases, rent rolls, valuations and collections. A portfolio model that tags every number by its basis — reported, estimated or modelled — makes those answers faster and more defensible.
Carrier invoices arrive weeks after the shipment and are paid on trust; cash-on-delivery remittances are matched by hand. Both are matching problems with clear evidence — exactly where a supervised agent earns its keep.
The rulebook is moving in two directions at once — EU deadlines deferred, US model-risk guidance rewritten, Asian supervisors naming AI agents explicitly. Banks that build one control framework for models and agents will not need to rebuild it for each regulator.
Most plants already have the machine signals and maintenance history to improve uptime. What they lack is a governed link between MES, CMMS, production plans and cost — and a clear line between what an agent may recommend and what a planner decides.
The same SKU now sells through dark stores, marketplaces, ONDC and the brand's own app — each with its own price, content and stock. Brands that treat each platform as a separate report lose sales in the gaps between them.
The IndiaAI mission is expanding public AI capacity and the DPDP Act changes how personal data may be used. For AI that touches citizens, the governing rule is simple: agents prepare the work, a named official decides, and every step is on the record.
A dashboard can be green while the service is not. Pauses without customer contact, closures bunched before reporting dates, changes that skip the CAB and leavers who keep access are rarely found by sample-based audits. Testing every record changes that.
Energy is a safety-critical business. AI agents can watch more signals than any control room, but a wrong action can black out a feeder or breach a pipeline. The answer is not to keep AI out; it is to give every agent an autonomy level, an owner and an audit trail.
Real estate runs on money moving between buyers, contractors, lenders and regulated accounts. Agents can do much of the reading and reconciling, but the design has to start from what they must never do on their own — and from consent for every buyer's data.
The UK's mandatory reimbursement regime is two years old, the EU has made payee verification universal for euro payments and Australia is legislating shared duties for banks, telcos and platforms. The evidence so far: shifting liability changes behaviour, but banks cannot stop scams alone.
When the operation breaks, recovery is still pieced together by phone and spreadsheet. Agents can assemble ranked options and rebook passengers within policy — but the duty manager must stay the one who decides.
Virtual assistants now handle a large share of routine banking conversations. The public record shows that the programmes that last are those that measure outcomes, not deflection, and make it easy to reach a person.
Many claim denials and insurer queries come down to the same thing: the justification is in the clinical record, but not in the claim. Drafting replies from the treating doctor's record, with finance in control, shortens the loop.
Logistics agents can scan, match and draft at a scale no team can. But three kinds of decision — a customs filing, a carrier payment and a commitment to a customer — must stay with a named person. Here is how we design for that.
Trade spend is one of the largest lines on a consumer company's P&L and one of the least evidenced. Supervised agents can evaluate every promotion and check every claim — provided they never approve money on their own and never use customer data without consent.
Payments revenue growth is slowing just as the rails underneath are being rebuilt. Instant account-to-account payments are becoming the default in much of Asia and Latin America, digital money is moving from trading to treasury, and banks must decide which layer they intend to own.
Manufacturers and trading houses sell through distributors and dealers, but primary sales live in SAP and secondary sales in a distributor management system. Until the two are joined on one golden record, schemes leak, stock is invisible and channel profitability is guesswork.
Models are no longer the constraint. Metadata, a shared semantic layer, knowledge graphs and lineage are — and banks that already invest in BCBS 239 compliance are closer to AI-ready data than they may think.
Access requests are one of the most repetitive jobs in any capability centre — and one of the riskiest to automate. The answer is not to avoid agents, but to design them so that nothing is granted without a named approver and everything is logged.
Shippers judge logistics providers by how early they hear bad news. A self-service platform and proactive alerts only work if every provider's milestones land in one governed record.
Instant payments are now mandatory across the euro area and growing fast in the US, and 2023 showed how quickly deposits can leave. Treasury needs a live view of intraday liquidity, and the data architecture to support it.
Southeast Asia has built some of the world's most advanced real-time and QR payment systems and licensed a new generation of digital banks. Payments have been won quickly; profitable lending has not. The next phase will be decided by data, credit and cost to serve, and incumbents still hold more cards than the headlines suggest.
Tickets, coupons, EMDs, settlements and interline billing are still reconciled across many systems. Matching, integrity checks and commentary are where agents pay back first — with controllers approving every entry.
The best use of AI in claims today is the work around the decision: intake, documents, status and evidence. Coverage, settlement and fraud referrals should stay with people who can be held accountable for them.
Manufacturing has safety, quality and money decisions that no agent should take alone. The answer is not to keep agents out, but to give each one an explicit autonomy level, an owner, and a ledger of the value it claims.
Europe's Digital Operational Resilience Act has produced its first register of critical tech providers and its first year of incident data. With the UK's regime now live as well, the lesson is that resilience depends on third-party dependencies and system failures more than on dramatic cyberattacks.
Assistants, early-warning scores and denial tools are only as good as the record underneath them. Hospitals that connect EHR, lab, pharmacy and billing data on one governed model build once and reuse everywhere.
Core replacement programmes have a long record of overruns and outages. A progressive path — data layer first, capabilities peeled off behind APIs and events, AI to read the legacy code — lowers the risk without lowering the ambition.
By the end of the decade the best banks will be felt more than seen: embedded in the journeys where customers live and work, run by humans supervising a digital workforce, and built on data they can trust. Here is the shape of that bank, and the operating model behind it.
Airlines can't let an opaque model decide a maintenance finding, a crew assignment or a passenger's refund. The answer is not to avoid AI but to give every agent an autonomy level, an owner and an audit trail.
Hospitals do not need a separate rulebook for every AI tool. They need one framework: what each agent may do on its own, who owns it, what each role may see, and a record of every action.
The Bank of England, the ECB and the EBA are all reshaping how they collect data from banks. The banks that benefit will be the ones that pair common data definitions with traceable lineage and AI that interprets, reconciles and explains their returns.
IFRS 17 asks for contract-level data, grouped and measured consistently, with a trail from source to the reported number. The insurers that move fastest treat that trail as a reusable data product, not a reporting project.
The standards are published and the deadlines are converging on 2030–2035. For banks, post-quantum migration is a multi-year programme whose hardest first step is not cryptography at all: it is building a trustworthy inventory of every key, certificate and algorithm in the estate.
AI can take the re-keying out of underwriting and sharpen risk selection. In regulated lines it will be judged on whether every flag and recommendation can be explained, tested for fairness and overridden by an underwriter.