Artificial Intelligence
Start with the process, not the agent.
Most AI projects begin by picking an agent or technology and then looking for something to point it at. The organisations getting real value from AI today started by focusing on the process they were trying to improve, then looked carefully at where AI truly fits, and where it does not.
In practice, that means AI that accelerates and synthesises work at scale, keeps a person in the moments that need one, and leans on workflow and decisioning tools where accuracy is paramount. This is process automation assisted by AI, not AI for the sake of AI.
What we mean
AI where it fits, workflow and rules where the stakes are high.
Customer service work, for example, is a mix of two very different jobs. Reading a request, routing it, and answering the routine questions is probabilistic work (the output can vary, even on the same inputs), and AI is very good at it. Committing to an outcome that has to be the same every time, and defensible later, is deterministic work, and that is work that for decades has been perfectly supported by rules automation solutions (eg. rules engines). Most organisations and service providers tend to forget this, and ask AI to deliver precision that it is fundamentally not designed for. We understand the probabilistic and deterministic worlds very well, including where and how they should converge. The hard part, and the one most miss, is the interface between them: the point where AI hands what it has found to the rule that makes the decision. Anyone can buy the two ends; the work that lasts is building the interface in the right place, where AI hands what it found to the rule that makes the decision. Right tool, right job.
The agentic conversation is backwards
"Agentic" is everywhere, but the market talks agents first and process second, if at all.
Flip it. Name the part of the operation that is holding you back, describe how it runs today and what good would look like, and the question of which tool goes where mostly answers itself.
Skip that step and you automate a broken process faster, which is how organisations end up in the queue of underwhelmed AI adopters. This is the whole argument in the AI conversation is backwards.
A probabilistic tool should not own a deterministic decision
AI is probabilistic by design. Full stop. You can make it more accurate, but no one will guarantee accuracy for that very reason.
A common response to the demand for certainty from AI is to ask a probabilistic tool to do a deterministic job, then stack more AI on top to check the layer below. That reduces the risk of error. It does not remove it.
The buyer's test is one question: would the vendor warrant that the system will always follow policy, through every model change, every data change, every environment change?
Knowing when and where to use AI is important, and avoiding summit fever by trying to fix AI's probabilistic nature with more AI.
We let AI do what it does best, but keep decisions governed by adjacent technology that produces the same answer every time, without fail. We wrote about that difference in AI is brilliant at finding an answer.
A probabilistic tool should not own the whole process
A process has structure: a definite shape, a current state, and a defensible next step. Workflow and case management engines were built to hold exactly that, with decades of patterns for knowing what a process looks like, where a given case is up to, and what has to happen next.
AI is unstructured by nature. It accelerates and supports the work; it's not always the best tool to use to govern it. The common mistake is to bolt guardrail after guardrail onto an AI agent under the guise of "Agentic orchestration" until it approximates what a workflow engine does out of the box.
We keep the process master in the tool built for it, and let AI move the work along inside that structure. Right tool, right job.
The agent belongs with the process, not the app
Bolting an agent onto a CRM or an ERP deepens the silo it already lives in. Real processes cross systems, but an agent trapped inside one application can only act on what that application knows, and its vendor will only ever warrant its own patch of ground. In some cases, this is sufficient. But a truly scalable, organisation-centric agentic process doesn't work this way.
Agentic work is process-centred, so the agent belongs with the process master that spans the systems, governed by the process it exists to support, not tucked inside any single app.
Outcomes you can own
What changes when operations run this way
The gains are operational and you can measure them: routine resolved in seconds, hard cases reaching the right person with the context already gathered, and quality that holds as volume scales, instead of degrading under it. All without losing confidence in your ability to deliver accuracy and compliance.
Routine cleared, people freed for judgement
High-volume, low-judgement work is handled end to end. The cases that need a person reach one with the context already gathered, so your team spends its time where experience actually counts.
Answers assembled in seconds, across systems
AI reads across documents, policies, and the systems where your information is scattered, and returns the answer with its source. Knowledge work that used to mean hunting through several places becomes a single question.
Outcomes that stay consistent as volume grows
Because the decision sits under rules rather than a model, a case is handled the same way whether it is the first today or the ten-thousandth, and you can show why. Quality holds as volume climbs instead of degrading under it.
Work that moves across systems, not stuck in one
Agentic work follows the process wherever it runs, spanning the CRM, the core system, and everything between, instead of stalling at the edge of one application. The process stays the master; the agent moves the work along inside it.
Patterns we have put into production
Patterns defined, deployed, and proven
These are patterns running in production today. Most began in service and customer operations, because that is where high volume met high stakes first. The same pattern carries into any high-volume operation: agents that span systems, with rules governing what they are allowed to commit to.
Multi-channel customer service orchestration
One conversation across web, voice, kiosk, and in-venue. Context follows the customer. The same AI agent handles the routine across channels; the same person handles the escalation when it is needed.
Where we have shipped thisEuropean banks on our AI agent platform, including Banca Sella (alongside a regional cooperative, a mobile-first fintech, and a specialty finance bank). A major Mediterranean ferry operator. A global food and dairy multinational.
High-volume self-service, agent-run
Agents handling high-volume, low-judgement work end to end: eligibility checks, status enquiries, bookings, amenity reservations, applications. Routine cleared in seconds. Hard cases routed to people ready to use their time well.
Where we have shipped thisState government contexts in regulatory oversight and education service delivery. Citizen service at Sei Toscana, with 100,000 appointments booked end to end. Condo and residential portfolio self-service on the platform.
In-venue and digital kiosk experience
Kiosks and avatars in physical locations handling service the same way the digital channel does. People on the floor freed for the moments that matter: the welcome, the recovery, the recommendation.
Where we have shipped thisA 5-star Italian luxury resort. Hotel and restaurant deployments on our conversational AI platform across Europe.
Accessibility as a first-class capability
Multilingual support including languages with limited digital presence, and sign language. Accessibility designed into the service from the start, not added at compliance review.
Where this mattersRegulated, public-service, and community-facing contexts where serving everyone is the standard, not the exception.
Where this applies
What this looks like in your industry
AI shows up differently in every sector. The constant is where it belongs: accelerating the work, and staying clear of the decisions that have to hold. A few pictures of where it earns its place.
Insurance
Claims triage and first notice of loss, broker and policyholder service across channels, and AI reading the document-heavy submissions that slow underwriting, all sitting on top of the rating and eligibility that has to stay exact.
Financial Services
Origination and servicing journeys, identity and document checks synthesised in seconds, and customer service across channels, with the credit decision itself governed by rules an auditor can follow.
Government
Citizen self-service for routine matters, case triage that routes work to the right officer, and eligibility bound to policy, multilingual and accessible by design rather than by retrofit.
Logistics
Consignment and exception handling at volume, shipment status pulled together across the systems that each hold a piece of it, and support for customers and drivers that does not queue.
Industrial & Manufacturing
Operating and maintenance knowledge on demand at the line, procedure and manual lookup in plain language, and warranty and service support that spans the systems behind it.
Retail & Hospitality
Guest and customer service in any language, in venue and online, with bookings, orders, and amenity requests handled around the clock and the front desk freed for the moments that matter.
Related thinking
The thinking behind this
Questions we get asked
What is agentic AI, and how is it different from a chatbot?
A chatbot follows a script. An AI agent gathers context, uses tools, and carries out multi-step work across the systems you already run, through integrations. Business rules govern what it is allowed to commit to.
What kinds of work do you apply AI to?
Three broad kinds: customer and citizen service across channels; internal knowledge and synthesis, reading across documents and systems to return an answer with its source; and process automation, where an agent moves work along inside a governed workflow. Consequential decisions in all three stay under deterministic rules - rules that integrate with AI, but are not made by AI.
Where should an agent live?
With the process it serves, not inside a single application. An agent bolted into one CRM or ERP can only act on what that system knows. Real processes cross systems, so the agent belongs with the business process, or user journey, that spans them.
How do you stop an agent making expensive mistakes?
Rule-bounded design: the agent proposes, and deterministic rules govern what can actually happen, no matter what the model suggests.
Where has this been proven?
In production across European banking, hospitality, transport, and food and beverage, with the decisioning layer underneath drawn from a decade of rules and automation work.
When is AI not the right answer?
When the same inputs must always produce the same outcome, when you need to trace or explain how the outcome was reached, or when the cost of getting it wrong outweighs the benefit. Those are decisions for deterministic rules. AI still does the reading and the routing; the rules make the call.
What do "deterministic" and "probabilistic" actually mean?
Probabilistic means the output can vary, even on the same inputs - that is how AI language models work, and it is the right property for reading and finding. Deterministic means the same inputs always produce the same output - if the applicant is over 65 and the cover exceeds the limit, then refer to a human. Rules are deterministic, which is why the decision that has to hold sits with them. We go deeper in AI is brilliant at finding an answer.
Can the same AI serve customers in any language, including sign language?
Yes. The platform we deploy is multilingual, including languages with limited digital presence, and runs a signing avatar in production at acquevenete, an Italian water utility, so deaf and hard-of-hearing customers self-serve directly. Accessibility is designed in from the start, not retrofitted.