Most companies have now moved well beyond only using AI chatbots — such as OpenAI’s ChatGPT and Google’s Gemini — in their day-to-day operations.
They’re now using agentic AI, which rather than just generating text, can search databases, send emails and analyse company data to automate workflows and complete tasks.
Agentic AI startups in Europe have raised a staggering €7.9bn this year, already surpassing the €7bn raised last year, according to Sifted data.
Startups are driving this innovation across Europe in a wide range of sectors, and include Swedish legaltech startup Legora, healthtech Tandem Health and UK-based fintech Cleo.
But how can companies successfully integrate agentic AI models into their everyday work?
Building the agentic layer
For an AI agent to execute tasks, it must be able to interpret often messy and unstructured data from internal systems, as well as the web.
Building a reliable agentic workflow requires turning this data into a structured “perception layer” that agents can reason with, says Emma Burrows, cofounder of Rezonant which develops software for organising a company’s data before they can start automating work.
This perception layer involves training AI models to “organise context from a lot of different inputs,” ranging from communication platforms such as Slack and email, as well as documents and files, she says.
Rezonant supports companies in organising their data by building a tailored ‘context graph’, which maps out scattered information and data into connected business categories.
“This context isn’t organised by every document put together,” she says. “But categorised by features or customer insights.”
By structuring data around categories within the business — such as sales, operations or finance — agents can pinpoint exactly what data they need to make a decision.
The next phase for enterprises is owning how their company-specific intelligence improves.
Many organisations mistakenly believe they need to modernise their entire tech stack before even thinking about using AI.
Lindsay Keim, VP of customer success at AI agent startup N8n warns that waiting to organise data stacks often prevents companies from starting at all, leaving them stuck in a “data cleaning purgatory.”
N8n is an AI platform connecting LLMs, data sources and business tools for a range of uses, including to complete business tasks.
The flexibility of platforms like N8n allows users building agentic workflows to “pull data points out of a legacy system, run them through a separate AI model for analysis and move them into a newer system, without actually modernising the whole stack first,” says Keim.
When workflows require external data, challenges often come from the unpredictability of the internet.
To overcome this, agents need an ‘abstraction’ layer, whose job is to “turn the messy, unpredictable web into consistent context,” says Rotem Weiss, founder and CEO of agentic search company Tavily, which provides an API connecting AI agents to web information.
Without that abstraction layer, an AI agent treats struggles to find and process reliable and relevant information from the web.
Governance, security and humans-in-the-loop
Giving AI complete autonomy to execute tasks across business systems that may be sensitive or confidential can introduce operational and compliance risks.
Organisations are increasingly prioritising trust boundaries (which separate a company's internal data from the public internet), governance and human-in-the-loop safeguards to prevent errors.
Context isn’t organised by every document put together, but categorised by features or customer insights.
If an agent does hit a roadblock or makes an error, how it reacts is critical. According to Burrows, agents can be guided to navigate errors, but mustn’t overstep boundaries.
“The permissions context is the most difficult and important part of what we do,” she says, referring to the permissions users can give an agent when it makes decisions on their behalf.
“Each workflow that’s available for an agent [has] specific permission controls that you can configure on top of it.”
Agents that initially rely heavily on human feedback can “over time, learn routing patterns,” Burrows adds.
“For example, anytime there's a decision about design, I normally ping my colleague Sam. Over time, the agent will notice that and might say 'do you want me to ping Sam in the future next time this happens?' So the agent can actually start to take on more work the way you would, but in a way that still keeps humans in the loop for high-stakes decisions.”
When agents start retrieving information from the web, external inputs aren’t always shared each time, so teams need to ensure caution is built in, says Weiss.
“When an agent needs information from the internet it should send only the minimum context necessary [from an organisation’s data] to retrieve that information,” he adds.
Governance also requires strict auditability. “If a chatbot gets one fact wrong, you get an unreliable or incorrect answer,” says Weiss. “If an autonomous agent gets that fact wrong and then writes it into Salesforce, for example, the error escalates.”
While AI can be responsible for “the speed, the scale and the data synthesis”, organisations must design their systems so that humans continue to provide “the judgement, the context and the accountability,” says Keim.
Measuring success and continuous learning
As organisations continue to use agentic systems, the way they measure their success is shifting.
While early adoption has been driven by immediate time and cost savings, the long-term value of agentic AI lies beyond doing things faster.
When companies first implement AI, attempting to automate everything at once may not give companies the outcome they’re looking for, Keim says.
“We rarely see a customer go from zero to fully agentic in one move,” she says. Instead, the fastest returns come from separating workflows and targeting repetitive tasks where automation would bring the highest value.
For example, one of N8n customers first implemented agentic AI by automating the process of updating details for employees who had retired.
Keim admits this example may not sound “super exciting”, but automating these processes reduced human error, freed up capacity and provided value to the customer.
When an agent needs information from the internet it should send only the minimum context necessary.
The team at Rezonant found that as using AI became part of everyday work, the focus needed to shift towards always maintaining high standards.
“We are very deliberate that we want to enable teams to move faster, but still keep taste and judgement,” Burrows says.
Without guardrails and humans-in-the-loop, companies risk producing inaccurate or low-quality results.
Once organisations solve the early basics of automation, the next step should be learning from AI models, says Weiss.
Every action an agent takes, such as “what the agent searched for, which sources it trusted or rejected and which tools it used” should be noted, he adds.
Weiss calls this concept “enterprise learning sovereignty.”
When an AI agent first starts working on company workflows, it attempts tasks, makes mistakes and gets corrected by employees. It learns the company’s standards as well as which internal systems to trust and what a successful outcome looks like for the company and its customers.
Learning sovereignty allows an organisation to own this continuous loop of trial, human correction and improvement, says Weiss. The AI agent isn't just completing tasks. It’s building a record of how an AI-driven business actually operates.
“The next phase for enterprises is owning how their company-specific intelligence improves”, he says.
