What's Possible with AI Today That Wasn't Possible a Year Ago
Every month brings another announcement claiming AI can now do something remarkable. Most of them are marketing. Some of them are real. The challenge for a business owner is distinguishing between the two — and understanding which real capabilities are mature enough to build on today versus which are still research demos.
Here is an honest account of what has genuinely changed in the past year and what it means practically.
Long-context reasoning has matured
A year ago, AI models had practical context limits that made them awkward for anything beyond short documents. You could summarise a single page, but a 50-page contract required chunking the text, summarising chunks, summarising summaries — each step introducing error and losing coherence.
Today’s models handle tens of pages in a single call with good attention across the full document. This is not a marginal improvement. It changes the class of tasks that are automatable. Document-heavy workflows — contract review, due diligence, policy analysis, customer email thread handling — that required significant pre-processing workarounds now work cleanly.
Structured output is reliable
Models can now produce structured data (JSON, specific formats) reliably enough to integrate with downstream systems without extensive error handling. A year ago, getting a model to consistently output a specific JSON schema required significant prompt engineering gymnastics and still produced frequent deviations that needed catching and correcting.
Today, structured output modes on major APIs produce format-conformant responses at a high enough rate to build production workflows on. The practical impact: the integration glue between AI and your existing software (CRM, ERP, databases) is much simpler to write.
Agentic workflows are buildable
AI agents — systems where a model takes multiple steps, uses tools, and makes decisions across a longer task — were largely research territory a year ago. Today they are buildable for well-defined business workflows.
A practical example: a document arrives by email. An AI system extracts key fields, checks those fields against your database, identifies discrepancies, drafts a response requesting clarification for the specific discrepancy found, and queues the draft for human review before sending. Each step involves a model making a decision. The human only sees the output ready for approval.
These are patterns we build and run in production. A year ago they would have been fragile prototypes.
Multimodal is genuinely useful
AI models now read images, screenshots, and scanned documents at a quality that makes them useful for real workflows. A year ago, vision capabilities existed but were unreliable for the kind of dense, complex documents businesses actually work with — tables, charts, mixed text-and-image layouts.
Today, processing a screenshot of a spreadsheet, extracting data from a photographed form, or reading a scanned invoice works well enough for production use in many cases. This opens up workflows where your data arrives in visual formats that previously required manual data entry or expensive OCR pipelines.
What has not changed as much as claimed
Real-time accurate information. Base models have training cutoffs. Retrieval-augmented generation helps but introduces its own failure modes. For workflows where accurate, up-to-date facts are critical — legal, financial, regulatory — AI needs to be paired with reliable data sources and the output needs review. Do not assume a model knows your current pricing, your regulatory status, or today’s news.
Complex relationship and context knowledge. Models do not know your specific industry nuances, your company’s particular situation, or your client relationships. They can be given context, but loading sufficient context to replace genuine domain expertise is hard. AI augments expert work; it does not replace it.
Consistent complex reasoning at high volume. For sufficiently complex multi-step reasoning tasks, error rates are lower than a year ago but not zero. High-stakes automated decisions — credit approval, medical triage, legal conclusions — still need human oversight.
The practical implication
The threshold of “mature enough to automate” has moved meaningfully downward in the past year. Workflows that required too much engineering investment to be worth it twelve months ago are now straightforward builds.
The question for your business is not “can AI do this?” (the answer is increasingly yes for a wide range of tasks) but “does the ROI justify the build, and does our team have the capacity to adopt the output?”
Want to find out which of your workflows now clear the automation threshold? Book a Free Consult and our AI specialists will give you a direct answer.
Want to put AI to work in your business?
Book a Free Consult with our AI specialists. One session, concrete next steps.