You have to still ‘keep taste and judgement’: How companies are actually putting agentic AI to work
Sifted ● Covered by 2 sources
Companies are moving past chatbots and using AI to do real work like search, email and analysis. The catch: it only works if humans keep the judgement and guardrails.
Based on reporting by Sifted — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Companies are no longer treating AI as a chat window on top of existing work. They’re pushing into agentic AI: systems that can search databases, send emails, analyse internal data and carry out workflows with less hand-holding than a chatbot. In Europe, that shift has already attracted serious money. Sifted data puts agentic AI startup funding at €7.9bn this year, above last year’s €7bn total.
The hard part isn’t getting an agent to act. It’s giving it something useful to act on. Emma Burrows, cofounder of Rezonant, says companies need a structured “perception layer” that turns messy internal data into something an agent can reason with. That means pulling context from places like Slack, email, documents and files, then organising it into business categories such as sales, operations or finance. Rezonant does that with a tailored context graph, which maps scattered information into connected categories so the agent can find the right piece of context instead of rummaging through everything.
There’s also a warning here for companies waiting to tidy up their entire stack before they start. Lindsay Keim, VP of customer success at N8n, says that can leave teams stuck in “data cleaning purgatory.” N8n’s pitch is more pragmatic: connect large language models, data sources and business tools, then let teams pull data from a legacy system, analyse it in a separate model and move it into a newer one without a full overhaul first. For web data, Rotem Weiss of Tavily says agents need an abstraction layer that turns the messy internet into consistent context.
That still leaves the ugly bits: permissions, governance and the risk of an agent doing the wrong thing at speed. Burrows says permissions are the most difficult and important part, while Weiss warns that an error becomes much more serious once it’s written into a system like Salesforce. The companies getting this right are not handing over the keys. They’re using AI for speed, scale and synthesis, while keeping humans for judgement, context and accountability.
The most interesting shift may be in how companies measure success. Keim says the best results usually come from automating one repetitive workflow at a time, not trying to become fully agentic in one leap. Weiss goes further, arguing that companies should track what an agent searched for, what it trusted, what it rejected and which tools it used. That’s less glamorous than the hype around autonomous AI, but it sounds a lot more like something businesses can actually live with.
My take — AI-written commentary, not fact-checked reporting
The fantasy is full autonomy; the reality is permissions, logs and a human who still knows when a thing looks off. That’s not a weakness, it’s the product. The companies pretending otherwise are usually the ones about to create a very expensive typo.
Read more about this at: Sifted
