The best AI systems know when not to use AI

It's Thursday, 6^th August 2026. Hello and welcome back to Bold Efforts.

Most weeks, I write about how technology is changing work from the outside. This week, I want to share something I learned while building it. Let’s go.

Over the past eighteen months, I have been building Fursa, a platform for finding jobs and hiring people. Much of the difficult work happens far beneath the search box.

Fursa scans roughly two million job records every day. These records arrive from different systems, in different formats, with inconsistent titles, locations, descriptions, and company names.

The same role can appear several times. Closed jobs can remain online. A position described as remote can quietly require employees to live in one country or commute to one office.

Before Fursa can help someone find the right job, it first has to establish which jobs actually exist.

When I began building this system, the obvious solution was to use large language models.

A language model can read an unstructured job description and interpret it. It can identify the function, seniority, location, employment type, workplace policy, skills, and qualifications. It can understand that “work from anywhere across EMEA” is different from “remote within Germany.”

This feels like exactly the kind of problem AI was built to solve. So why not send every record through the most capable model available?

Because the fact that AI can perform a task does not mean AI should perform it.

AI products often look most impressive when viewed one example at a time.

You give a model a complicated job description. A few seconds later, it returns a clean, structured record. The result feels almost magical.

But a production system is not judged by whether something works once.

It has to work millions of times.

The output must remain consistent. The cost must make sense. Failures must be detectable. The system must reveal enough about its decisions for you to diagnose problems. And it must run again tomorrow when the underlying information changes.

At that scale, sending everything through a powerful language model creates new problems.

It is expensive and slower than simpler methods. Similar inputs can produce different outputs. Obvious questions get treated like ambiguous ones. And when a result is wrong, it can be difficult to understand exactly why.

The system appears more intelligent, but becomes less reliable.

The better approach was not to find a more powerful model. It was to become more selective about where intelligence was required.

Suppose a job record contains a recognized country code. There is no reason to ask a language model which country it represents.

If a company provides a stable job identifier through its hiring system, that identifier should carry more weight than a model’s interpretation of the description.

If a page has not changed since it was last processed, most of the work does not need to happen again.

These are not intelligence problems. They are systems problems.

Deterministic logic handles them better. The same input produces the same result. The decision is quick, inexpensive, and easy to inspect.

Other problems are genuinely ambiguous. “Flexible working” can mean remote work, occasional work from home, or simply flexible office hours.

“Based in London” can mean the employee must live in London, can work anywhere in the United Kingdom, or only needs to visit the office periodically.

A title such as “Chief of Staff” can describe corporate strategy, operations, executive support, or some combination of all three.

These problems require context. This is where language models become valuable.

But the architecture is not a simple choice between fixed rules and LLMs.

Between them sits a broad middle layer of fuzzy matching, vector similarity, embeddings, semantic search, and statistical methods. These techniques are less flexible than a language model, but they are often faster, cheaper, and more predictable.

They are particularly useful when the system does not need to generate an interpretation from scratch, but still needs to recognize similarity, classify a record, or compare it with information seen before.

The goal is not to use the most advanced method available. The goal is to use the simplest method capable of solving each problem reliably.

Fursa’s enrichment pipeline combines deterministic rules, classical methods, language models, external information, historical records, and confidence thresholds.

Straightforward cases follow predictable paths. Similarity problems can be handled through embeddings, matching systems, or previous records.

Ambiguous cases go to models that can interpret language and context.

Uncertain cases are not forced into false certainty. They can be routed through additional checks, assigned lower confidence, or left unresolved until more evidence becomes available.

This distinction is easy to miss. We often think intelligence means producing an answer. But sometimes the intelligent act is recognizing that the system does not yet know enough.

A reliable system must know when it knows, when it needs to reason, and when it should remain uncertain.

The result is that Fursa can enrich a record for roughly $0.003 while maintaining more than 95 percent deduplication across its live job inventory.

The cost matters, but not because cheaper is always better. It matters because job information decays quickly.

New roles open. Old roles close. Descriptions change. Employers move between hiring systems. Application links break. Remote policies shift. A reliable map of available work has to be rebuilt continuously.

At millions of records, even a small unnecessary cost compounds rapidly. A pipeline that is affordable to run once but too expensive to refresh is not a reliable system. It is a snapshot.

Reducing the cost of understanding each record allows Fursa to keep understanding the market again and again.

Much of the discussion around AI still treats the model as the product. Companies announce which model they use. People compare benchmark scores. New products are judged by how impressive the first response appears.

But a model is only one component. A production system also needs rules, memory, verification, fallbacks, monitoring, confidence thresholds, and a clear response when the model is wrong.

The most capable model does not automatically produce the most capable system.

A powerful model used on every record can make the product slower, more expensive, and less predictable. A smaller model applied to the right subset can create more value. A matching algorithm can outperform both when the task is recognizing similarity. A simple rule can outperform everything when the answer is already known.

The quality of an AI system depends less on how much intelligence it contains and more on how carefully that intelligence is allocated.

This distinction will become more important as AI spreads through every company.

The easiest way to add AI to a process is to place a model in front of it. The harder work is redesigning the process itself.

Which decisions are repetitive enough to become rules? Which require interpretation? Which require comparison with past information? Which need a person? Which mistakes are acceptable? What should happen when confidence is low?

These are questions of architecture, incentives, risk, and judgment. They are also increasingly questions of management.

For most of the industrial era, companies divided work between people.

One person handled finance. Another managed sales. Another wrote software. Another made hiring decisions.

AI introduces a different division of labour.

Work will now be divided between deterministic software, probabilistic models, and people.

Deterministic systems are good at following known rules. Statistical and similarity-based systems are good at identifying patterns across large volumes of information.

Language models are good at interpreting messy language and operating within ambiguity.

People remain responsible for deciding what matters, defining acceptable trade-offs, and acting when the situation does not fit the system.

The value will not come from replacing one of these with another.

It will come from assigning each type of work to the right kind of worker.

A company that uses people for predictable, repetitive decisions wastes human judgment.

A company that uses AI for decisions requiring accountability, values, or deep context creates risks it may not understand.

A company that uses fixed rules for a changing and ambiguous world becomes brittle.

And a company that reaches for the most powerful model before understanding the problem will spend more money to produce less reliable work.

The advantage lies in knowing the difference.

This may become one of the most valuable skills in the future of work: not simply knowing how to use AI, but knowing how to redesign work around different forms of intelligence.

There is also a personal lesson in this. We often treat repeated deliberation as evidence that we are being thoughtful.

We reconsider whether to exercise. Whether to save money. Whether to protect time for focused work. Whether to check our phones during dinner. Whether to accept another meeting.

But many of these are not new decisions. They are recurring versions of decisions we have already madee.

Using judgment every time does not make the outcome more intelligent. It makes the outcome more vulnerable to mood, fatigue, convenience, and distraction.

Some parts of life benefit from becoming deterministic.

You exercise on certain days. You save a fixed share of your income. You reserve hours for important work. You decide what you will not compromise on before the pressure to compromise arrives.

You do not negotiate with yourself each time. Rules reduce the number of decisions competing for your attention. But life cannot become entirely deterministic.

No rule can tell you which person to trust, which opportunity deserves several years of your life, when persistence has become denial, or what kind of work feels worth doing.

These decisions cannot be automated cleanly because their value lies in context, uncertainty, responsibility, and personal meaning. They deserve judgment.

Freedom does not come from making every decision again each day. It comes from deciding the obvious things once, building systems around them, and protecting your attention for the decisions that remain genuinely uncertain.

The objective is not to remove thinking from life. It is to stop wasting it.

AI makes intelligence cheaper and more accessible.

That does not mean we should apply it everywhere.

The strongest systems will use intelligence selectively. They will rely on rules when the answer is known, statistical methods when patterns can be recognized, language models when the world needs interpretation, and people when a decision requires responsibility or meaning.

The same is true for companies. And it is true for individuals.

As intelligence becomes abundant, the scarce skill will be knowing where it belongs.

P.S. Fursa launched on Product Hunt today. You can explore the product and support the launch here.

Thank you for reading.

Best,
Kartik

I write Bold Efforts every week to think clearly about where work and life are actually headed. If you want these essays in your inbox, you can subscribe here.