OnClarity Insights
AI models are becoming commodities. The harness is what you’re buying.
What an AI harness is, why the model should stay swappable, and the questions to ask any vendor who says they built their own model.
6 min read
The short answer
Leading AI models are converging in quality and falling in price, and new ones arrive every few months. What decides whether an AI agent works for your customers is the harness around the model: approved knowledge, guardrails, actions, handover, quality checks, monitoring and the freedom to switch models. Buy the harness. Keep the model swappable.
Are AI models really becoming commodities?
At the top, mostly yes. As of March 2026, four AI companies sat within 25 Elo points of each other on the leading public leaderboard, which the Stanford AI Index 2026 describes as “top model performance is converging”. Price is moving the same way: the cost of running a model at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024 (Stanford AI Index 2025).
Commodity does not mean identical. Models still differ by task, by language, by speed and by where they can run. The top closed model led the top open-weight model by 3.3% in March 2026 (Stanford AI Index 2026), so the model you can host yourself is not always the one at the top. In voice, models that go straight from speech to speech are catching up with the classic chain of speech-to-text, a language model and text-to-speech.
What it does mean: the best model this quarter will probably not be the best model next year, and the gap between the leaders is small. A choice you make once and lock in will age. A choice you can change will not.
Why do so many vendors say they built their own model?
Because it sounds like a moat. In a crowded market, “we built our own model” is an easy line to say and a hard one to check. Look closer and many of these products turn out to fine-tune an existing model or run on another company’s service.
That is a normal way to build software. Fine-tuning a strong model for a narrow task is good engineering, and building on a third-party service is how most products start. The problem is only when the pitch says something else. If you buy on the model story, you may be buying something the vendor does not own and cannot change.
What is the harness?
The harness is everything around the model that turns it into an agent you can put in front of customers. In customer service it has eight parts:
Approved knowledge. The agent answers only from current, owned articles and policies, not from what the model remembers.
Guardrails and validators. Rules on what it may say and do, checks before an answer goes out, and confidence thresholds below which it does not answer.
Tools and actions. The connections that let it do the work (look up an order, change a booking, issue a refund), with approvals on the sensitive ones.
Handover with context. When it is unsure or the case is high-risk, a person takes over with the full conversation, so the customer does not repeat themselves.
Quality checks on every conversation. Each conversation scored against your own standard, with the evidence, so you find failures before customers report them.
Monitoring. Live views of what the agent is doing, and alerts when something changes.
Model choice and switching. The ability to run the right model for each job, and to change it, with tests run on your own past conversations before each change.
A personal data layer. Personal details removed before a model outside your boundary sees them, where your rules require it.
Image: the model sits in one slot that can change. Knowledge, guardrails, actions, handover, quality checks, monitoring and personal data protection stay in place.
Source: OnClarity analysis.
Most of what goes wrong in production happens in the harness, not the model: a missing or outdated article, an instruction that rewards guessing, an action with no approval, a handover that drops the context. That is also where most of the fixes are. We set out the controls for one of the most common failures in How do you stop an AI agent from making things up?
Why should the model stay swappable?
Because almost everything around the model will change faster than your contract:
A better model arrives for your language, your channel or your price point.
A model is retired or its price changes.
A rule changes about where data may be processed, and the model has to move with it.
Your volume grows and a smaller, faster model becomes the better fit for part of the work.
If your knowledge, guardrails, actions and tests live in the harness, switching the model is a change you test and ship. If they are tied to one model or one provider, it is a rebuild. That is vendor lock-in, and it is the risk worth paying to avoid.
Switching is not free, even with a good harness. A model change is an agent change: rerun your tests on real past conversations, compare, then switch.
What should you ask a vendor who says they built their own model?
None of these is a trick question. A good vendor answers each in a sentence.
Did you build the model from scratch, fine-tune an existing one, or call someone else’s?
Which parts are yours: the language model, speech-to-text, text-to-speech?
Where does it run, and who operates it?
Can we switch to another model? What do we have to rebuild if we do?
What happens to our agent when the model underneath is retired or changes price?
How do you test a model change on our own conversations before it goes live?
What do we get from you that we would not get by calling the model directly?
The last answer is the one to listen for. It should describe the harness.
The honest trade-off
Models are not interchangeable for every job. A fast model matters a lot in voice, where callers notice silence. Some languages and dialects are handled much better by some models. Hard reasoning still favours the strongest models, and a model tuned for a narrow task can beat a general one. So model choice still deserves care, task by task, on your own tests.
A harness also takes work. Someone has to own the knowledge, write the rules, connect the systems and review the quality scores. Swapping models needs a test set and discipline. The point is not that models do not matter. It is that the model is the part most likely to change, so it should be the easiest part to change.
What to do this week
Ask your current or shortlisted vendor the seven questions above, in writing.
List which of the eight harness parts you have today, and who owns each.
Build a test set of 20 real past conversations, including ones that went wrong, to run before any model change.
Check your contract for what happens when the underlying model is retired or changes price.
OnClarity is model-agnostic: you can use a choice of leading LLMs, a self-hosted open-weight model, or bring your own model, with a PII removal layer that strips personal data before a global model is called. See AI agents and deployment options.
Questions people ask
What is an AI harness?
An AI harness is everything around a model that turns it into a working agent: the knowledge it may use, guardrails and validators, the tools and actions it can take, handover to a person, quality checks, monitoring and the ability to switch models. The model generates the words. The harness decides what it knows, what it may do and how you check it.
What does model agnostic mean?
A model-agnostic product is not tied to one AI model or provider. It can run different models for different jobs and switch between them without rebuilding the agent’s knowledge, rules, integrations and tests. For customer service teams it lowers the risk of vendor lock-in, because the model can follow price, quality and data rules as they change.
How do you avoid AI vendor lock-in?
Keep the parts you invest in, such as knowledge, rules, integrations and test sets, independent of any single model. Ask vendors how a model switch works and what it costs, check your contract for model retirement and price changes, and keep a test set of real conversations so you can compare a new model before you move.
Sources
Stanford Institute for Human-Centered Artificial Intelligence, The 2026 AI Index Report, Chapter 2: Technical Performance (data as of March 2026).
Stanford Institute for Human-Centered Artificial Intelligence, The 2025 AI Index Report, top takeaways (inference cost, November 2022 to October 2024).
General information. The vendor questions are a checklist, not a judgement on any company.
