AI models are commodities. Your harness is the advantage.

Every AI-enabled product, including LLM apps such as ChatGPT and Claude, are models with harnesses. None of them have a harness tailored to your specific needs. But you can create your own harnesses. 

Every AI feature you have bought or built is two things stacked. A foundation model that reasons, and a harness that tells it what to work on, what to read, which tools to call, and which limits to hold. The model layer is becoming a commodity, interchangeable and cheaper every quarter. The harness is where durable advantage lives, and almost every harness running in production today was engineered by a vendor and calibrated to an average buyer.

Anything an AI vendor can sell you, your competitor can buy at the same price and the same depth. The one capability no one can sell you is the one built from your rubrics, your standards, your institutional history, and the way your firm actually decides. When you engineer that logic into a harness, you gain a massive competitive advantage. 

Whose logic runs your AI?

A harness is the engineered environment around a model: the instructions, retrieval, tools, memory, evaluation, guardrails, and human checkpoints that turn a general reasoner into a capability that does highly specific work.

A vendor's harness encodes a vendor's idea of best practice for a representative company. That is the right design choice for a product sold to thousands of buyers, and generic logic produces generic results. Logic built from how your firm actually operates produces results calibrated to your firm, and that caliber of capability is not for sale anywhere.

The one layer that doesn't commoditize

Pull a harness (or agentic AI system) apart and you find some combination of dozen or so working parts: triggers, schedulers, automations, classification, retrieval, tools, state and context, observability, guardrails, human-in-the-loop checkpoints, orchestration, and logic.

All agentic/harness componenets are commoditized infrastructure, available to anyone and rapidly improving. All except one. The logic layer that carries encoded judgment, is a massive moat when engineered correctly. It is the one part a vendor cannot ship to you, because a vendor has never sat inside your operation. 

Why owned capability compounds

Not only do commercial AI products have generic logic, but they also create "islands of intelligence" and tech sprawl.  Each arrives with its own login, its own contract, its own pricing. Stack ten of them and you have ten islands rather than a system. A harness you own can architect across these islands, can read from and write to them, or can stand apart. You have endless flexibility.

As you build successive custom logic layers to power different use cases, these can pull from a common infrastructure and can inform each other. This is a core aspect of Talbot West's Cognitive Hive AI (CHAI) framework.  

Each capability sits on shared architecture and inherits from other logic layers, so the second build is cheaper than the first and the tenth is cheaper still. Capability compounds while marginal cost falls. The exact proportions depend on how many capabilities you eventually deploy and how much of your existing stack the harness composes against rather than replaces, and the direction holds under any reasonable assumptions.

The work with no compile step

Some AI work has cheap ground truth. Code either compiles and passes its tests or it does not, so a harness for a coding agent can catch failure after the output exists. Most enterprise work has no such check. If you think about a brief, a disclosure, an analysis, or a proposal, nothing automatically tells you the output is shallow, off-standard, or subtly wrong.

The failure mode is fluent shallowness,  and a thin tool will produce it with confidence. A harness for this kind of work has to structure the reasoning before the output exists, moving the model through the standards and steps your experts would apply. This is the harder discipline, and it covers most of the valuable judgment work in a company.

A well-built harness moves a model reliably upward. On judgment-bound work, the last stretch still belongs to a person, and anyone promising full autonomy on taste-bound work is selling against a wall that has not moved yet and is unlikely to.

Where a harness lives

The discipline of harness engineering is agnostic to tech stacks and environments. Harnesses can take infinitely many configurations, and this flexibility intimidates the uninitiated while providing those with mastery an endless set of options.

Here are some of the many instances of a custom harness.

1. The harness exists completely inside a single piece of software. For example, business logic encoded as Skills inside the Claude app. 
2. The harness lives inside a single cloud service or productivity suite (e.g., AWS Bedrock, Microsoft 365, Google Cloud) and spans multiple apps or software surfaces in the same family. 
3. The harness spans several dissimilar software surfaces, with handoffs or integrations between them. Example: a set of documents in Sharepoint, reasoned over with Claude, with Zapier triggering actions and events. 
4. The harness lives in a dedicated surface, and integrates to external systems. Example: A complex build-out in AWS Bedrock that reads from (and selectively is able to write to) Salesforce, Workday, Sage Intact, and other systems. 

This flexibility accommodates every conceivable use case, enterprise tech stack, budget, timeline, compliance environment, and set of priorities. There's a perfect harness for every situation. Some can be deployed in a day, others may take several months. Some require no code at all, while others involve extensive custom development. Some involve extensive tooling, triggers, sequences, and integrations, while others need only a reasoning layer. 

Some categories of work where harnesses deliver

The most non-commoditizable, high-value harnesses tend to be those that optimize high volume processes that involve nuanced judgment, carry a heavy cost of error, and require senior people whose hours are the bottleneck. Here are a few of the general categories we've seen.

• Regulated and high-stakes writing. Filings, disclosures, regulatory submittals, and capital-markets materials, where voice and accuracy both carry weight.
• Engineering and operational documentation. Briefs, specifications, and technical reports that encode senior judgment and look similar from one to the next.
• Institutional knowledge codification. The rubrics and decision patterns held in a few heads, captured in a form the operation can query after those people leave.
• High-volume document workflow. Contracts, leases, and clause work where most items are routine and the rare exception is what needs a human.
• Revenue and sales reasoning. Account research, call coaching, and outreach that reason from your playbook and your wins rather than generic market data.
• Templated communication at scale. Stakeholder, customer, and community correspondence in your voice and branding.

A harness use cases addendum takes several of these categories and shows what a harness for each one looks like inside a specific kind of business.

What to ask before your next AI purchase

The next AI tool you buy will arrive with someone else's logic already inside it, tuned for a company that is not yours. That serves commodity work well enough. For the work that actually distinguishes your firm, the more useful question is what your own logic is worth once it is encoded into a system that can run it at scale, and who you want engineering that system. Advanced custom harnesses will be the most durable competitive advantage of the next few years for the firms that build them, because they are the one capability a competitor cannot acquire by signing the same contract you did.

Addendum 1: AI harness primitives, and why logic is the moat

A harness is a stack of engineered parts. Most of them are turning into standard infrastructure that anyone can buy. 

The parts that are commoditizing

The working parts of a harness group into a few jobs:

• When it runs. Triggers, schedulers, and automations decide what kicks the system off and on what cadence.
• What it knows. Classification, retrieval, and state and context assemble the right information for the task at hand.
• What it can do. Tools and orchestration let the model take actions and carry out multi-step work in order.
• Whether you can trust it. Observability, guardrails, and human-in-the-loop checkpoints keep the system auditable, bounded, and correct.

These are real engineering, and wiring them well takes skill. But they are built from components everyone can reach. Cloud platforms ship them, open-source projects sharpen them, and every vendor draws on the same catalog. The pieces get cheaper and better every quarter on their own. None of them, by itself, separates your AI from a competitor's, because none of them is yours alone.

The logic layer

The logic layer is the encoded answer to a single question: what does good work look like here, and how does this firm actually decide? In practice it holds:

• The rubrics and standards that separate excellent output from merely passable output in your domain.
• The decision criteria your experts apply, including the tradeoffs they weigh and the lines they will not cross.
• The institutional history: what has been tried, what worked, and what the firm has learned not to repeat.
• The escalation rules that say which cases a person must see, and why.
• The voice and posture the firm holds with the outside world.

This is the part the model cannot supply on its own and the vendor cannot ship in a box. Several forces keep it that way.

There are no economies of scale on it. A vendor turns something into a commodity by building it once and amortizing the cost across thousands of buyers. Your logic layer applies to exactly one firm. There is no one to spread the cost across and no product to package, so the economics that push everything else toward commodity pricing do not act on it.

Its inputs are private and largely tacit. Models commoditize in part because they train on shared, public data and converge toward one another. The logic layer draws on your private context, much of which has never been written down, and it diverges from everyone else's by construction. Pulling it out of senior people's heads and into an explicit, reusable form is the slow, human work of building a harness, and it is exactly the part that cannot be downloaded.

It is a moving target. Your standards and your decisions shift as the firm operates and learns. A logic layer you own keeps absorbing those changes. One you rent freezes at the vendor's last release, and the gap between the two widens with time.

Commoditization raises the floor, not the ceiling. Each year the models get stronger and the infrastructure gets cheaper, which lifts the baseline for everyone at once. What that progress never touches is whose judgment the system encodes. As the commodity layers improve, the logic layer becomes a larger share of what separates one firm's AI from another's, not a smaller one.

Put together, this is why a competitor can buy the same models, rent the same cloud primitives, and even hire similar engineers, and still not reproduce your capability. Everything on the shelf is available to them at the same price you paid. The one input that is not on the shelf is the logic that comes from inside your operation. That is the part worth owning, and the part worth engineering with care.

Addendum 2: AI harness use cases

A revenue-reasoning harness for a sales organization

Most sales AI tools land as one more dashboard, one more login, and one more vendor for IT to manage, and the productivity gain pays for the friction. A custom business development harness consolidates many related functions into a single architecture, with each capability built as a module on the same spine, sharing the same source-of-truth documents and the same firm-specific reasoning. 

Because every workflow scores against one rubric and reads from one set of source documents, the system sharpens as the rubric improves, and the firm's selling logic, not a vendor's, does the reasoning.

A harness portfolio for an upstream energy operator

For an upstream operator running a large capital program off a lean senior bench, the work that rewards a harness clusters into six categories.

A proposal harness for a research firm pursuing government contracts

A small research-and-development firm that lives on federal solicitations runs into a hard ceiling: every SBIR, STTR, or agency proposal needs section-by-section compliance, prior-art and past-performance retrieval, and a technically grounded narrative in the firm's own voice, and all of it lands on a few senior people whose hours cap how many bids go out.

A proposal harness runs that front end. It parses each solicitation into a requirements map and a compliance matrix against the agency's review criteria, retrieves the prior proposals, technical write-ups, patents, and publications that match, and drafts first-pass technical sections grounded in the firm's corpus with citation candidates attached. Because the firm bids into a controlled environment, the harness runs inside the same security boundary, logs every action, and traces every claim back to its source, so a reviewer can defend any sentence internally or in front of the agency.

The principal's technical judgment still decides what to pursue and what to ship; the harness removes the lookup and assembly time that used to cap how many proposals the firm could even attempt. The corpus and retrieval layer it installs become the foundation that later capabilities build on, from technical-literature monitoring to internal technical Q&A against the firm's own work.

A contract-and-redline harness for a legal team

For a legal team carrying high contract volume, most of the work is routine and a small fraction carries real risk. The bottleneck is that a senior lawyer's attention is needed on everything to find the fraction that matters. A harness for this work runs inside a regulatory perimeter and reads each incoming document, flags every deviation from the firm's clause library, drafts the routine redline, and routes anything outside policy to a person with the relevant precedent attached.

Privilege-scoped retrieval, action limits, and an audit trail are defaults rather than features. The lawyer stops reading every line and starts reviewing only the exceptions, with the harness's reasoning shown next to each one.

A knowledge-capture harness for an engineering firm

An engineering firm runs on judgment held in a few senior heads, and that judgment leaves when those people retire or move on. A harness encodes the firm's decision patterns, project history, and standards into a reference layer the rest of the team can query.

An engineer asks how the firm has handled a structural problem before and gets an answer drawn from decades of completed projects rather than from whoever happens to be in the room. The capability here is institutional memory: it survives turnover and gets stronger each time a project is added.

A regulated-writing harness for a bank or insurer

A bank or insurer produces a constant stream of writing that has to be accurate, on-voice, and defensible: credit memos, disclosures, regulatory correspondence, and customer notices. The cost of a subtle error is high, and the volume is too large for senior review of everything. A hybrid harness fits this work. Deterministic gates check the required disclosures, figures, and regulatory language, while the model drafts the narrative against the institution's standards. Compliance holds final approval, and every draft carries the citations and checks that let a reviewer trust it quickly.

A correspondence harness for a utility or healthcare system

An organization serving a large population generates correspondence at a high volume: billing questions, eligibility notices, service updates, and complaint responses. Most of it follows known patterns, and a minority is sensitive enough to need a person. A harness drafts the patterned majority in the organization's voice, holds the required regulatory and accuracy constraints, and routes the sensitive minority to a person with full context attached. The team's hours move from typing routine replies to handling the cases that genuinely need judgment.