// System architecture

The most successful AI systems use the least AI

Choosing where AI belongs is an engineering decision with the same trade-offs as any other: accuracy, cost, complexity and risk. The organisations getting returns are the ones being disciplined about it.

Avanti Technologies · 6 min read

It is remarkably easy to fall into the trap of using AI for everything. Teams are wiring agents across their systems, and the impulse is understandable — agent systems are genuinely powerful. But power is not the same as fit. AI is not magic, and it is not the best solution for every problem. Choosing to use it, and at what scale, involves trade-offs across accuracy, cost, complexity and risk, exactly like any other engineering decision.

Two numbers frame the cost of getting that decision wrong.

95%

of enterprise generative-AI pilots delivered no measurable impact on the P&L, across 300 public deployments and roughly $30–40bn of spend. MIT's research points to integration and workflow fit, not model quality.

MIT NANDA — State of AI in Business, 2025
40%+

of agentic AI projects will be cancelled by the end of 2027 — because of escalating costs, unclear business value or inadequate risk controls. Model capability is not on the list.

Gartner, June 2025

Four kinds of intelligence

Almost every problem in a software or data system can be solved in one of four ways: human judgment, rules and code, machine learning, or generative AI. These are not a ladder you climb from humans up to agents. They are distinct options, each with a specific role, and the goal is to pick the right one for the problem.

Diagram 01
Four options, four distinct roles
Humans
Rules & code
Machine learning
Generative AI
Best at
Complex judgment, ambiguity, edge cases
Clear, stable logic — exact outputs at scale
Patterns in structured data, prediction
Unstructured input, reasoning, synthesis, language
Use when
Stakes are high and someone must own the outcome
Logic is known and explicit; errors unacceptable
Patterns exist but rules are too complex to write
Flexibility matters more than precision
Examples
Hiring, diagnosis, legal interpretation, strategy
Payments, validation, transforms, access control
Fraud, churn, demand forecasting, recommendation
RAG Q&A, summarisation, code gen, agent workflows
Cost profile
Expensive, slow, hard to scale
Cheap, fast, scales cleanly
Moderate; ongoing monitoring cost
Significantly more costly at scale
Fails when
Volume is high — millions of decisions per second
Conditions shift and logic gets too complex
Drift sets in, or scope moves beyond training
Correctness must be guaranteed run to run

Humans: judgment and accountability

People handle ambiguity and edge cases well, make decisions from incomplete or conflicting information, and apply context, ethics and accountability in ways machines cannot. Humans should drive high-stakes decisions, anything requiring ownership or liability, and cases of significant ambiguity: hiring, medical diagnosis, legal interpretation, large-scale strategy and financial decisions. If something goes wrong, accountability must sit with a real person.

The trade-off is that human judgment is high quality but expensive, slow and very hard to scale. You can screen CVs automatically; the final hire involves team fit and communication skills you would not want fully automated. Equally, you would never ask humans to review every transaction on a card network — that is millions per second.

Rules and code: determinism

Traditional software is good at executing clear, stable logic and producing consistent, exact outputs at scale. Use it when the logic is known and explicit — if x, do y — requirements are stable, and errors are unacceptable: payment processing, input validation, data transformation, and above all security and access control. Code is fast, cheap, reliable and interpretable; straightforward to test, debug and scale.

To process a card payment you want a code path that checks whether the balance is sufficient and approves or declines. Deterministic, as fast as possible, no room for error. AI here introduces risk and buys nothing. Same for validating an email address format: you need a rule, not a model.

Where rules break down

Rules fail when conditions change constantly and the logic becomes too complex to maintain. Take fraud detection. A static rule — flag any transaction over $1,000 — is quickly useless against attackers whose behaviour evolves around it.

Diagram 02
A static rule decays; a learned model adapts
HIGHLOWFRAUD CAUGHTattackers split paymentsnew channels, new geographiesrule is noiseRULE SHIPPED18 MONTHS LATER
Static rule
Fixed threshold. Every adaptation by the attacker is a permanent step down.
Learned model
Retrained on new behaviour. Holds its ground, at the price of monitoring.
Illustrative — shape of the problem, not measured data

Machine learning: patterns and prediction

Machine learning excels at learning patterns in structured data and making probabilistic predictions. Reach for it when patterns exist but are not obvious, when rules would be too complex to define by hand, and when what you actually need is a prediction: fraud detection, churn, demand forecasting, recommendation.

For demand forecasting, a human brings intuition and bias; a rules-based approach is too simplistic to pick up the nuances of a trend. A model learns the complex patterns in the data and predicts more accurately. Models scale well and behave predictably, but they need monitoring and ongoing maintenance — drift is real, and flexibility outside the original problem scope is low. Explainability varies, and if you need a written report on what it thinks the societal causes of a demand trend are, you are out of luck.

Generative AI: language and interpretation

Analysing trends across massive structured datasets is not the natural home of an LLM. Unstructured data is. Use generative systems when inputs are text, documents or other media, when the task involves interpretation or transformation, when flexibility matters more than precision, and when some error is tolerable: retrieval-augmented question answering over document sets, large-document summarisation, code generation, multi-step agent workflows.

In customer support, instead of hard-coded responses, an agent can understand intent, pull the relevant knowledge and generate a genuinely helpful reply. Doing that with rules alone would be close to impossible. Generative systems are flexible and reach a first result fast. The trade-offs are non-determinism, correctness that is not guaranteed across runs, and a cost curve that is materially steeper at scale.

That is the question that actually decides whether to build with agents and LLMs: how much non-determinism and uncertainty is this workflow willing to tolerate?

Diagram 03
Three properties, priced differently
Determinism
Flexibility
Cost at scale
Humans
Rules & code
Machine learning
Generative AI
Relative positioning. No single option wins all three columns — which is why the choice is a trade-off, not a ranking.

Hybrid systems win

Say you want an AI agent to analyse a year of spending. You cannot feed in thousands of transactions and expect a non-hallucinated answer, or even correct arithmetic. What you want is a calculator tool using code and maths for determinism, traditional machine learning to identify trends over time, and then an LLM to generate a clean, readable report at the end.

These hybrid systems tend to be the most successful: all four types of solution in conjunction, each doing the part it is actually good at. When someone says make everything agents, it was never meant literally. Every strong modern system is a combination.

Diagram 04
One workflow, three kinds of intelligence — annual spend analysis
Input
12 months of transactions
Thousands of rows of structured ledger data.
Code
Calculator tool
Totals, categories, variances. Exact every run.
Machine learning
Trend detection
Seasonality, anomalies, direction of travel.
Generative AI
Narrative report
Explains the numbers in plain language.
Untouched by the model · Arithmetic never delegated to an LLM · Pattern work, monitored for drift · The only non-deterministic step

A heuristic you can use in the next review

Use humans when you need judgment, accountability and ethics, or the workflow is high-risk. Use rules or traditional code when the logic is clearly defined — if x happens, do y — and it will not change often. Use traditional machine learning when you want patterns and predictions from past data. Use generative AI when you need to interpret, generate or reason over complex inputs and flexibility matters more than precision.

Diagram 05
Which system for this problem
Does someone need to be accountable for this decision, or is it high-risk?
Human judgment
↓ NO
Can the logic be written down explicitly, and will it stay stable?
Rules and code
↓ NO
Is the input structured, and what you need a prediction from past data?
Machine learning
↓ NO
Is the input unstructured, and can you tolerate variation between runs?
Generative AI
Most real workflows answer yes at more than one gate. Split the workflow at those boundaries and give each step the system it deserves.

Where projects actually die

Most failures to reach production do not come from bad models. They come from choosing the wrong system in the first place. That is worth sitting with, because it changes what a CXO should be asking in a review: not which model, but which kind of intelligence, and who owns the output.

Diagram 06
The stated causes of failure — and whether a better model fixes them
Cause
What it really is
Fixed by a better model?
Escalating cost
Generative inference applied to work a deterministic path would do for a fraction of the price.
No
Unclear business value
The problem was chosen because the technology was available, not because the workflow needed it.
No
Inadequate risk controls
Non-deterministic output placed in a workflow that requires a guaranteed answer.
No
No workflow integration
A pilot that demonstrates capability but never enters the system of record.
No
Causes as stated by Gartner (June 2025) and MIT NANDA (2025). Assessment is ours.

The future of technology is not replacing humans, and it is not putting AI everywhere. Great systems are built by making disciplined choices about how decisions get made across humans, rules, models and machines.

The most successful AI systems will not be the ones using the most AI. They will be the ones that consistently choose the right kind of intelligence for the problem — and know when not to use it at all.

// Work with us
We review architectures for exactly this question.

Avanti Technologies builds and supports enterprise systems that stay up. If you have an AI initiative in flight — or one you are being pushed to start — we will walk the workflow with your team and tell you plainly which parts warrant a model and which do not.

Sources
  1. MIT Media Lab NANDA initiative, The GenAI Divide: State of AI in Business 2025 — 95% of enterprise GenAI pilots showed no measurable P&L impact; based on 300 public deployments, ~150 executive interviews and ~$30–40bn of enterprise spend. The report attributes failure to a learning and integration gap rather than model quality.
  2. Gartner, 25 June 2025 — over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Gartner also estimates only ~130 of the thousands of self-described agentic AI vendors are genuine.

// Keep reading

Briefing · AI Infrastructure
Training large language models: an EU briefing

What pre-training, post-training and evaluation involve — and what they mean for compute, data governance and talent decisions in Ireland and the EU.

CXO Guide · Machine Learning
Machine learning, minus the mystique

What machine learning actually is, where it earns money in a 20–250 person European business, what a first project costs, and what the EU rulebook asks of you.

Research Note · Strategy
Build, buy, or partner

Where deep-tech AI startups should own the stack, where they should rent it, and where ownership has quietly stopped being an option — with decision protocols.

All resources →