AI agents are everywhere. They are in the headlines, dominating LinkedIn feeds, and likely being discussed in every boardroom. But for an enterprise leader responsible for a complex application or a mission-critical workflow, the noise can be overwhelming. How do you actually start? Read on for our guide to starting your Enterprise AI journey.
In this interview series, Estafet founder and CEO Adrian Wright talks with enterprise architect Branimir Angelov (Brani) about how enterprises can start with AI, using proof of concepts (PoCs), workflow analysis, and the shift toward agentic AI. Brani is a hands-on software architect adept at stripping away the jargon and providing a pragmatic roadmap for AI adoption.
We recorded this interview because many enterprise teams are being pushed to “do something with AI” before they have a clear starting point or a safe route to production. As an established UK and nearshore software delivery and consultancy company, Estafet take an engineering-led approach to AI adoption:
- Start with the right workflow
- Define success metrics early
- Build a PoC that stands up to governance and security expectations, and
- Create a clear path to production.
That blend of strategic thinking, delivery discipline and hands-on AI expertise helps enterprises move faster without gambling on quality, compliance, or long-term maintainability.
If you prefer to watch and listen, see the full video 45-minute on our Estafet YouTube channel here.
Phase 1: Agentic AI – where to start with AI in the Enterprise
Adrian kicked off the conversation posing the question of where a project manager or architect needing improvements, workflows or applications might want to start with AI. How would they know if AI is applicable to their workstreams?
Brani draws a sharp line between applications you sell and workflows that sit inside a wider value chain. Customer-facing apps are sensitive and quality drops can hurt your brand immediately. Brani suggests starting with workflows – the internal processes that form your larger value chain. Internal workflows are lower risk, are easier to roll back and therefore offer a safer lab for experimentation. Workflows also provide cleaner measurement, ideal for making a business case.
Key points:
- Are the problems you want to solve bounded or unbounded?
- Unbounded: Problems requiring human-like reasoning or processing unstructured data (e.g., summarizing a complex legal brief). This is where Generative AI shines. Examples are: drafting, interpreting, summarising, handling cases
- Bounded: Problems with clear parameters (e.g., classification, OCR, or identifying a specific error). Traditional ML models or specialized AI are often more efficient here. Examples are: classification, OCR, extraction, identification
- The Data Factor: Don’t let a lack of data stop you. Modern Large Language Models (LLMs) come with fundamental knowledge that allows you to start with very little data via prompting or fine-tuning.
Deterministic workflows can still be a strong AI candidate.
Where AI can help:
- automating steps that still take manual time
- benchmarking the before/after in a clean way
- producing a business case that stands up to finance scrutiny
Where AI may not pay back:
- automation is already high
- no human involvement
- quality is already good, and there is no cost or performance pain
Watch this part of the interview: Where to Start in the Enterprise.
Phase 2: Making the business case for AI
A PoC isn’t just a technical demo; it’s a business case. The goal is to prove value, measure results and identify a clear road to production. Brani suggests starting with risk analysis and domain understanding.
Key Steps for a High-Impact PoC:
- Use pre-trained models: Don’t reinvent the wheel. Start with models available on public or specialized cloud platforms.
- Establish a sandbox: Run your PoC in an environment that mimics your IT governance but is isolated enough to allow fast iteration.
- The feedback loop (MLOps): You must be able to benchmark performance. If you inject AI into a document extraction step, you need to know exactly how it compares to the manual baseline in terms of speed, cost, and accuracy.
“The PoC is agile in nature. You want to have a fixed scope and fixed budget to create a clean business case. Within that timeframe, you rely on the seniority of the team to utilize AI-assisted coding to move faster.” – Branimir Angelov
Watch this part of the interview: Making the business case for AI.
Phase 3: Agentic AI – Proof of Concept Governance & Security: Protecting IP and Data
The biggest fear for any enterprise is losing control and leaking sensitive data. Brani emphasizes that security cannot be an afterthought – it must be part of the PoC.
- Protecting IP: Ensure your managed MLOps platform explicitly states that any fine-tuning or custom models remain your property and are not used to train third-party models.
- Data Sovereignty: Use your existing cloud landing zones. This ensures that the AI components follow your established network protections, encryption standards, and access controls.
Working on these types of projects, Estafet builds trust through traceable decisions.
Watch this part of the interview: Protecting ip and data.
Phase 4: The team: roles, responsibilities, and best practices
An AI project team looks different from a traditional software team and needs a complete squad – a cross-functional group of senior people. The team must be built to move fast, because the PoC must be short and credible.
The Core Roles:
- The Product Owner (PO): Must have high autonomy and a deep understanding of the business impact.
- The Data Scientist/Developer: Someone who can handle architectural layers, Python coding, and benchmarking machine learning pipelines.
- The MLOps Engineer: Responsible for bringing pre-trained models into the sandbox and managing the deployment lifecycle.
- The Specialized Tester: Traditional testing isn’t enough. You need frameworks to measure hallucinations, response structure, and the quality of GenAI outputs.
Watch this part of the interview: Roles, responsibilities, and best practices
Phase 5: Transforming workflows, strategy, and business growth
In this part of the interview, Adrian pushes the discussion beyond a single business case towards Agentic AI – systems that can proactively interact with one another to manage entire value chains.
Adrian notes that AI acceleration is on such a steep curve that we may soon see tiny teams generating massive revenue by orchestrating AI agents.
New roles are appearing, but AI cannot sit in a silo. On the discussion of Chief AI roles and fractional AI leads, Brani’s warning is simple: AI cuts across the organisation.
- CTO, CEO, operations, and information security all own part of it
- siloing AI blocks adoption, and blocks control
- regulation and ethics become sharper when AI touches customers, not internal operations
In conclusion, the top tips are:
- Don’t wait: If you aren’t learning now, the gap between you and your competitors will become insurmountable within two to three years.
- Think Global, Act Local: Boards need to discuss the “art of the possible” regarding AI transformation, but also stay grounded in emerging AI regulations and ethics.
Watch this part of the interview: Transforming workflows, strategy and business growth.
Thanks to Adrian Wright and Brani Angelov for this insightful interview!
Ready to turn AI talk into measurable progress?
If your board is asking “Where do we start?” you don’t need another slide deck. You need a clear path from idea to evidence to production.
Estafet is an established UK and nearshore software delivery and consultancy company. We bring the thinking that de-risks your choices, the engineering discipline that keeps delivery tight and the practical AI experience that shows exactly where AI will add value in enterprise workflows and systems.
If you want to identify the right use case, prove it quickly and move forward with confidence, book a chat with Adrian Wright. You will come away with clarity on what to test, how to measure success and what it takes to get it production-ready.
To find out more about Estafet and the types of work we deliver, you can explore our case studies here.
Get in touch if you’d like to find out how Estafet can bring your team a safe space for AI development at pace.
