The idea of introducing AI across a business is an extremely seductive one. The idea of increasing efficiencies and lowering costs is extremely enticing, and rightly so.
Commonly, an organisation will hand the project to an IT lead, and ask them to produce and execute a plan to further explore the opportunity. Worse still, they might say ‘get on with it’.
To do so is not only likely to miss the best opportunity, but to create instability, cultural concerns and non-compliance.
Our philosophy at NITA is to always take a strategic approach, lead by a qualified business analyst and with all appropriate steps taken.
This article investigates the five challenges; a sister article looks at solutions.
You can use the list at the end of this article for explanations of some of the terminology in this article, or our full glossary of terms here
AI Isn’t the Goal. Better Outcomes Are.
Let’s start with ‘The Goal’. What is the strategic end-point of an AI-led business transformation?
There’s a question echoing through boardrooms right now that few leaders are willing to ask out loud: Why isn’t our AI investment paying off?
It’s not for lack of ambition. Most enterprises have launched pilots, hired data scientists, and signed platform deals. Yet the return remains stubbornly elusive. Chatbots frustrate customers. Automation projects stall in proof-of-concept purgatory. Dashboards multiply while decisions don’t get faster.
The problem isn’t the technology. It’s the approach.
AI transformation is not a technology project — it’s a shift in the operating model itself. The goal was never AI for its own sake. The goal is better decisions, happier customers, and new revenue streams. In other words, better outcomes.
Reaching those outcomes demands two things in equal measure: technical expertise and cultural change. Leaders who treat AI as a procurement exercise will continue to be disappointed. Those who treat it as a catalyst for rethinking how their organisation works will pull ahead.
Across our work with business leaders, five challenges surface repeatedly:
- Operational Efficiency
- Data-Driven Decisions
- Customer Experience
- Innovation Enablement
- Risk Management
These five challenges cut across industries, sectors, roles and cultures. Here is an explanation of each.
1. Operational Efficiency
The instinct is to throw AI at existing workflows. But if those workflows are fragmented — riddled with handoffs between disconnected systems, manual compliance checks, and zero visibility into where time actually goes — automation through AI only accelerates the dysfunction.
The top issues observed by NITA when it comes to operational efficiency are as follows:
- Fragmented workflows
- Handoffs across systems (CRM/ERP) causes latency and rework
- Unobserved Processes
- No end-to-end telemetry, no cycle-time/queue-time visibility
- Manual Controls
- SOX/GxP evidence gathered by people; brittle and slow
- Non-deterministic AI automation
- Drift and unbounded actions
- Shadow tools
- Teams adopt unsanctioned tools that create compliance gaps
2. Data-Driven Decisions:
Every organisation claims to be data-driven. Few actually are. The symptoms are familiar: conflicting numbers across departments, duplicated features nobody owns, batch pipelines that deliver yesterday’s insights for today’s problems, data that cannot be understood or consumed by departments, and governance policies that exist on paper but aren’t enforced where it matters — in the data pipeline itself.
The top Data issues observed by NITA are:
- Messy lineage
- Conflicting truths across marts/lakes; unclear provenance
- Feature debt
- Duplicated, undocumented features; re-computed ad hoc
- Decision latency
- Batch ETL prevents real-time insights and interventions
- Bias/representativeness
- Skewed samples degrade model utility and equity
- Governance ‘theatre’
- Policies exist but aren’t enforced in pipelines
3. Customer Experience:
Your customers don’t care about your AI architecture! They care about whether your chatbot gives them a wrong answer, whether they have to repeat themselves across channels, and whether your recommendations feel relevant or random. It is normal to have a chatbot at the start of the customer experience. Think of them as Front Of Desk! So, they have to be trained with absolute assiduousness.
Here are the main challenges we have seen when it comes to CX – perhaps the single most common and damaging of all:
- Siloed channels
- Inconsistent intents/state across chat, email, voice, app
- Hallucinating assistants
- Incorrect answers erode trust and Net Promoter Score (NPS)
- Cold-start & staleness
- Recommendations fail for new users/items or drift
- Latency budgets
- Sub-200ms expectations clash with heavy LLM pipelines
- Safety & brand voice
- Uncontrolled generations harm tone/compliance
4. Innovation Enablement:
Most organisations have no shortage of AI experiments. What they lack is a mechanism to move the promising ones into production — and the courage to kill the rest. From our experience at NITA, here’s where the main sticking points are:
- Prototype graveyards
- Proof of Concepts never ‘cross the chasm’ to production
- IP & data leakage
- Risks when experimenting with third-party models
- Undifferentiated heavy lifting
- Teams reinvent infra and scaffolding
- No ROI signal
- Experiments lack stage gates, metrics, or kill criteria
- Talent bottlenecks
- Scarce Machine Learning engineers block domain teams
5. Risk Management:
Risk management in AI cannot be an afterthought bolted on before launch. It must be designed in from the start as it can lead to business-critical issues
The five most common examples that we see are as follows:
- Model drift
- Performance degrades over time as live data changes, creating unreliable outputs unless monitoring and retraining are built in
- Bias and unfair outcomes
- Skewed training data or poor sampling can produce decisions that disadvantage particular groups and create regulatory and reputational risk
- Data leakage and privacy exposure
- Sensitive information can be exposed through prompts, outputs, or weak controls when teams use live data inappropriately
- Third-party model risk
- External models and suppliers may introduce unclear accountability, limited transparency, and compliance gaps if they are not properly governed
- Hallucinations and opaque decisions
- AI can generate confident but incorrect answers or decisions that are hard to explain, undermining trust, auditability, and operational control
The Real Question
The question for leaders isn’t whether to adopt AI. That ship has sailed. The question is whether you’re building the foundations — in your data, your operations, your culture, and your governance — to make AI actually deliver on its promise.
Technology alone won’t get you there. A practical, structured approach — one that pairs deep technical capability with the willingness to change how people work and how decisions get made — will.
The organisations that thrive in the next decade won’t be the ones that adopted AI first. They’ll be the ones that adopted it well.
What do I do next?
If this article has resonated with you, why not look at our suggestions for how to solve these challenges?
Glossary of Technical Terms and Acronyms
- Architecture: The overall design and structure of systems, applications, and data flows.
- Automation: Using technology to carry out tasks with reduced manual effort.
- Batch ETL (Extract, Transform, Load): A process that collects data in groups, cleans or reshapes it, and loads it into another system at set intervals rather than continuously.
- Bias: A systematic distortion in data or models that can lead to unfair or inaccurate outcomes.
- Brand voice: The consistent tone, style, and personality an organisation uses in its communications.
- Cold-start: The difficulty of making good recommendations or predictions when there is little or no historical data.
- CRM (Customer Relationship Management): Software used to manage customer information, sales activity, service interactions, and relationships.
- CX (Customer Experience): The overall quality of a customer’s interactions with an organisation across all touchpoints.
- Cycle time: The total time taken to complete a process or task from start to finish.
- Data lake: A large repository that stores raw data in many formats for later analysis.
- Data lineage: A record of where data came from, how it was changed, and how it moved through systems.
- Data leakage: The accidental exposure or unauthorised sharing of sensitive data.
- Data mart: A smaller, focused store of data built for a specific business area or team.
- Decision latency: Delay in getting the information needed to make timely decisions.
- Domain team: A team with specialist knowledge of a particular business area, such as finance, operations, or customer service.
- Drift: A change over time in data, behaviour, or model performance that can reduce reliability.
- ERP (Enterprise Resource Planning): Integrated software used to manage core business functions such as finance, procurement, operations, and supply chain.
- Feature debt: The cost and confusion caused by having too many duplicated, poorly documented, or inconsistent data features.
- GxP (Good x Practice): A collective term for regulated quality standards such as GMP, GLP, and GCP, commonly used in life sciences and healthcare.
- Hallucinating assistants: AI systems that confidently produce incorrect, unsupported, or invented answers.
- Infrastructure (infra): The underlying technical foundations needed to run systems, such as servers, storage, networks, and cloud services.
- IP (Intellectual Property): Valuable creations of the mind, such as ideas, content, designs, inventions, code, and proprietary methods.
- Kill criteria: Agreed conditions that determine when an experiment or project should be stopped.
- Latency: A delay between an action and the result. In business processes or systems, it refers to waiting time that slows things down.
- Latency budget: The maximum acceptable delay for a system response before the user experience is harmed.
- Model utility: How useful a model is in practice for delivering value or supporting decisions.
- ms (milliseconds): A unit of time equal to one thousandth of a second.
- Non-deterministic: Not guaranteed to produce exactly the same output every time, even when given similar inputs.
- NPS (Net Promoter Score): A customer loyalty metric based on how likely customers are to recommend a company to others.
- Operating model: The way an organisation is structured and run to deliver its strategy, including people, processes, governance, technology, and decision-making.
- Platform deal: A commercial agreement to buy or adopt a software platform or technology service.
- Proof of concept (P.O.C.): An early demonstration used to show that an idea or technology could work in practice.
- Provenance: The origin of data and the history of how it was created or handled.
- Queue time: Time spent waiting before the next step in a process can begin.
- Real-time: Available or processed immediately, with very little delay.
- Representativeness: The extent to which data reflects the population or situation it is meant to describe.
- ROI (Return on Investment): A measure of the value gained from an investment compared with its cost.
- Scaffolding: Supporting components, tools, or code used to help build, test, or connect a larger system.
- Shadow tools: Unapproved or unofficial tools adopted by teams without formal oversight.
- Siloed channels: Customer contact routes, such as chat, email, phone, or apps, that operate separately instead of sharing context.
- SOX (Sarbanes-Oxley Act): A US law that sets requirements for financial controls, reporting, and auditability in companies.
- Stage gate: A formal review point used to decide whether a project should continue, change direction, or stop.
- Telemetry: Automatically collected data about how a system or process is performing.
- Third-party model: An AI model provided by an external supplier rather than built internally.
- Workflow: The sequence of steps, tasks, and approvals needed to complete a process.