Artificial intelligence dominates software industry conversation, but the reality is more nuanced than either the hype or the backlash suggests. Putti offers a balanced view: AI genuinely adds value in specific contexts and fails predictably in others, and knowing the difference is what separates successful implementations from expensive experiments.
The fundamental shift with modern AI is that computers can now operate in "much fuzzier and more human-like realms." Historically, software excelled at precise, repetitive tasks with deterministic outputs. Today's AI systems handle ambiguity, language, and pattern recognition, but accuracy isn't guaranteed in every application. That necessitates evaluating traditional versus AI-driven approaches for each specific problem, not adopting AI as a default.
What We See Working
- Staff leveraging language models to locate information, analyse documents, draft content, and format data
- Context-aware chatbots providing meaningful user assistance
- Coding assistants accelerating development processes
- AI scheduling and note-taking applications
- Prototype development through prompt-driven approaches
- Industry-specific implementations with quality datasets: workflow matching, form field extraction, personalised recommendations
What We See Failing
- Marketing claims promising production-ready products from prompts alone
- Off-the-shelf AI models applied to specialised problems, resulting in hallucinations, inconsistency, and privacy concerns
- Automation projects that underestimate data preparation requirements
A Strategic Approach for NZ Mid-Market Companies
1. Prioritise the problem first
Define the pain point before selecting a technology. Avoid AI if alternative solutions prove more direct, maintainable, or cost-effective. The question isn't "how can we use AI?" It's "what is the most effective solution to this specific problem?"
2. Establish measurable KPIs before you build
Track time savings, conversion improvements, error reduction, and transaction costs. Projects without quantified targets cannot demonstrate value, and cannot secure the internal buy-in needed for ongoing investment.
3. Audit your data quality
Verify cleanliness, representation, and accessibility. Confirm data ownership and control. Addressing data infrastructure before deploying models is non-negotiable: poor data in, poor results out.
4. Implement iteratively
Progress through proof-of-concept, controlled pilot, and production phases. Evaluate reliability, latency, costs, and user feedback at each stage before committing further investment.
5. Plan for operational maintenance
Model performance degrades over time. Monitoring, logging, retraining, and rollback procedures must be designed in from the start, not bolted on after go-live.
6. Maintain human oversight
For most business processes, human-model combinations outperform fully automated decisions in both safety and effectiveness. Design workflows that keep humans accountable for consequential outcomes.
7. Address compliance early
Privacy, intellectual property, explainability, deepfake risks, and bias concerns should guide design, vendor selection, and contracts, not be retrofitted once a product is in production.
Vendor Tooling and Hosted Models
Accessible APIs and hosted language models lower the barrier for prototyping and non-sensitive features. However, applications involving confidential customer data or proprietary information warrant private deployment. Scrutinise contractual protections: clarify output ownership, data storage practices, and restrictions on vendors using your data to train their models.
The Developer Reality
Despite marketing promises, building reliable software remains fundamentally engineering work. Integration, security, testing, monitoring, and deployment pipelines require proper implementation regardless of whether AI is involved. Adding AI introduces new infrastructure, additional testing requirements, and specialised expertise. A strong custom AI services partner provides pragmatic guidance, transforming concepts into sustainable, maintainable solutions rather than fragile prototypes.
Implementation Playbook
- Problem discovery and value hypothesis definition
- Data audit and feasibility assessment
- Lightweight prototype with defined KPIs
- Pilot deployment with selected users and monitoring
- Production release with ongoing model operations and support
Conclusion
Artificial intelligence will reshape software development positively in many contexts. It neither replaces strategic thinking nor eliminates the need for strong engineering and design practices. Organisations that prioritise problem analysis, data ownership, and operational planning are most likely to convert AI implementation from trend into genuine competitive advantage.