Our POV · POV 04 · AI strategy

What does AI need around it to do real work?

Eugene Yap, Managing Partner3 min readOctober 2026

The short answer

Two things you build around it. The model is the part you buy, and increasingly everyone buys the same one. Results depend on the harness, which sets the AI up to do a specific job, and the infrastructure, which gives it reliable data and systems to work on. Think of the model as a brilliant new hire: the harness is how you set them up, and the infrastructure is the company they join.

The harness

52.8→66.5%

Same model. Changing only the harness moved it from the top 30 to the top 5 on a demanding coding benchmark.

The infrastructure

87%vs43%

of data leaders say their data is AI-ready, and the share who call data readiness their biggest barrier.

Sources: LangChain, Terminal-Bench 2.0, 2026. Drexel LeBow and Precisely, 505 enterprise data leaders, 2026.

Both numbers point the same way. The model matters less than what surrounds it. In a separate 2026 study, the same model’s score swung by almost ten points depending on the harness alone, a wider spread than switching between leading models. And while most data leaders believe their data is ready, Gartner expects organisations to abandon 60% of AI projects that lack AI-ready data.

The star hire

Same genius. Different results.

Select any part to see what happens when it is missing.

The modelThe hire

Raw intelligence, rented by the hour. The same one your competitors use.

A brilliant graduateGPT, Claude, Gemini

The harnessHow you set them up

Turns raw intelligence into reliable work on one job. Built in weeks.

The infrastructureThe company they join

Makes the work possible, safe and repeatable across the business. Built over years.

What happens when it is missing

Select a part of the harness or the infrastructure.

What happens when each part is missing

The job description (instructions). It improvises the role. Tone, scope and policy drift from one day to the next.
The briefing pack (context). It answers from general knowledge, fluently and confidently, about a business it has never seen.
System access (tools). It can advise but never act. Someone copies every answer into the real system by hand.
A notebook (memory). Every conversation starts from zero. It greets a client of five years like a stranger.
A manager who reviews (checks). Mistakes go straight to customers, and nobody sees them until someone complains.
Accurate records (data). Confident answers built on the wrong numbers. A perfect harness on messy records still gets it wrong.
Connected systems (integration). It sees one system at a time, and every new use case begins with a new integration project.
The right permissions (security and access). It sees too much and legal stops the project, or too little and it cannot do the job.
IT that works (compute and platform). Pilots that shine on a laptop stall at scale, on cost, speed or reliability.
Oversight (monitoring and governance). Quality slips quietly. Nobody can say how well it performs, or explain a decision to a regulator.

Two investments, two clocks

Harness

Infrastructure

What it does

Sets the AI up to do one job well

Lets that work run safely across the business

Time to build

Weeks, improving fast

Months to years, compounding

When it is weak

Confident answers nobody checks

Pilots that never reach production

Who owns it

The business team and its AI builders

Technology, data and risk leaders

Models are converging, so the advantage moves to the two layers you own. They run on different clocks. A harness can be built and improved in weeks, one use case at a time. Infrastructure takes years and compounds: every clean dataset and connected system makes the next use case cheaper. The practical path is to start the harness on data you already trust, and build the infrastructure in the order your use cases need it.

Everyone can hire the same genius. The advantage is in how you set them up, and the company they walk into.

What to do next

1
Start with one job.
Pick a valuable, well-bounded use case and build its harness: instructions, context, tools, memory and checks.
2
Fix data in use-case order.
Clean and connect the records that first use case needs. Each one makes the next use case cheaper.
3
Measure the system.
Keep a fixed set of test questions and re-run them after every change to the model, the harness or the data.

Questions leaders ask

What is an AI harness?

Everything around a model that sets it up to do a specific job: its instructions, the context it is given, the tools it can use, its memory, and the checks on its work.

Is a better model enough?

Rarely on its own. A 2026 study found the same model’s score varied by almost ten percentage points depending on the harness alone, more than switching between leading models.

What makes data AI-ready?

Data that is accurate, current, connected across systems and governed, so the AI can find the right information and is allowed to use it. Gartner advises aligning data work to specific AI use cases.

The harness results come from software tasks, where the evidence is strongest today.

Sources

LangChain (February 2026). Improving Deep Agents with harness engineering. Terminal-Bench 2.0.
Zhang et al. (2026). Stop Comparing LLM Agents Without Disclosing the Harness. arXiv preprint.
Gartner (February 2025). Lack of AI-Ready Data Puts AI Projects at Risk.
Informatica (2025). CDO Insights 2025. 600 data leaders.
Drexel LeBow and Precisely (2026). Data Integrity and AI Readiness. 505 data leaders.

Put it to work