Every year produces a list of AI trends, and most of them describe what has become technically possible rather than what is changing how businesses actually operate. The distinction matters. A trend is only interesting to a business when it moves a number that business already cares about: revenue, cost, speed, quality or risk.
This piece looks at the shifts we see doing exactly that, with examples from systems we have built, and ends with a practical view of what to do about them.
1. From chat to agents
The first wave of business AI was conversational. You asked a question and received an answer. That was useful, but it left the actual work to a person: reading the answer, deciding what to do, and doing it.
The shift now is toward agents: systems that carry out a sequence of steps against real tools and data. An agent does not just suggest an email; it researches the recipient, drafts the message, sends it on schedule and sorts the replies. The value is not in the conversation. It is in work that no longer has to be done by hand.
Our AI Sales Development Representative is a clear example. It runs prospecting, enrichment, personalization, outreach and reply handling as a single pipeline, and took most of the manual prospecting off the team and put more worthwhile meetings in the calendar.
The lesson from building agents is that reliability comes from structure. Each step must validate its output before the next one runs, or errors compound silently. Agents that work in production are usually more constrained, not more autonomous, than the demos suggest.
2. Retrieval that understands structure
Retrieval-augmented generation, in which a model answers from documents found at question time, has become the standard way to ground AI in a company’s own knowledge. Plain retrieval finds passages that look similar to a question. That works for many questions and fails for an important class of them.
Businesses increasingly need answers that depend on how things relate: which regulation applies to which product, which supplier affects which shipment, which commitment depends on which asset. Knowledge graphs combined with retrieval, often called GraphRAG, make those relationships explicit, so a system can follow them rather than guess at them.
That is the approach behind Nexo, which made answers far more precise and cut the time analysts spent hunting for them, in an import and export compliance setting.
3. Synthetic data becomes a serious tool
Synthetic data has moved from a research curiosity to a practical way to get answers faster and protect privacy. In market research, AI-generated respondents can stand in for survey panels that are slow and expensive to run, allowing ideas to be tested in hours rather than weeks.
Our Multimodal Synthetic Surveys track closely with live benchmarks and replaced panels that cost a great deal more. The important word is benchmarks. Synthetic data is only credible when it has been measured against real results and the gap is understood.
4. AI that works on more than text
Voice, images and video are now practical as both inputs and outputs. That opens use cases that text alone could never reach.
- In healthcare, voice-led assistants such as Megan handle routine patient calls and save each provider about five hours.
- In retail, virtual try-on lets shoppers see a garment on themselves, and they stay on the page longer because of it.
- In document-heavy work, multimodal search finds information that exists only in figures and scans.
5. Ethics and governance become engineering problems
Ethical AI is moving out of policy documents and into system design. The organisations making real progress treat fairness, transparency and accountability as requirements to be engineered and tested, not principles to be stated.
In practice that means evaluation harnesses that measure performance across groups, audit trails that show how an answer was produced, human review for uncertain cases, and clear escalation paths. The organisations that treat governance as an engineering requirement will ship faster, because they will not have to retrofit trust later.
6. Measurement over demonstration
The most important shift may be cultural. Businesses have seen enough impressive demonstrations that fail in production. The question has changed from “can AI do this?” to “what will this change, and how will we know?”.
That is healthy. It favours projects scoped against a clear metric, measured against a baseline and evaluated in production rather than in a controlled test.
What to do with this
None of these trends is, on its own, a reason to start an AI project. The reason to start is a specific problem with a number attached. The trends only tell you which tools have become good enough to solve it.
A practical approach:
- List the problems where work is repetitive, slow or error-prone, and write down what each one costs.
- Pick one where a clear metric would show whether AI helped.
- Measure the baseline before building anything.
- Start narrow, prove the result, and expand from there.