← All articles

Consulting firm EXL: in enterprise AI, what matters isn't the model, it's what you build around it

Topics: AI, Automation, Small business

Server infrastructure layered around a glowing central core

Short answer: on October 22, 2026, MIT Technology Review's website published a paid, promotional piece titled "Designing an AI serving harness built to last." Right at the end there's a direct disclaimer from the publication: this isn't an editorial article, it's a piece written and paid for by EXL, a company that sells enterprise AI consulting. EXL's main point: the competitive edge in enterprise AI depends less on which model you pick than on the engineering system built around it. That point has a direct commercial interest behind it: this is exactly the kind of "harness" EXL sells. Below I go through what in this piece is worth taking as a working principle, and what remains an ad for a specific service.

Who wrote it, and why that matters

The piece is published on the MIT Technology Review technology platform, but it isn't an editorial article. It's what's called sponsored content: a paid piece from a company that the publication hosts on its site. At the bottom of the page there's a note: the material was prepared by EXL, and MIT Technology Review staff did not write it. Whether the editorial team reviewed the text before publishing isn't stated in the note itself.

EXL is a consulting and engineering firm that helps large businesses build and run AI systems. That directly means the piece has a commercial interest: the more companies believe their AI project won't take off without a properly engineered "harness" around the model, the more demand there is for services like the ones EXL sells. That doesn't make the point wrong, but it does mean it's worth treating as a vendor's point of view rather than independent research.

EXL's main point

According to EXL, the question companies ask themselves about enterprise AI has shifted from "what can the model do" to "can we deliver that intelligence reliably, economically, and at the scale we need," as AI moves from experimentation into real work. This is EXL's own framing of a shift in emphasis, without a reference to any survey or study.

By EXL's account, the answer to that question increasingly depends on what the company calls the "serving harness": the engineering system around the model, which includes the architecture, memory management, request scheduling, and optimization decisions.

EXL's three practical arguments

Argument one: start from the task, not the model. In EXL's view, there's no single "best" model or architecture, only the right solution for a specific business problem. An application that answers a customer instantly and an overnight batch document-processing pipeline, the company claims, need different engineering architectures even if both use the same base model. EXL warns that a model with slightly more accurate answers isn't necessarily the better choice for a production system if it doubles the cost of every request or doesn't meet the required response speed.

Argument two: measure before you improve anything. EXL advises finding the system's main bottleneck, fixing just that one, measuring the result, and only then looking for the next one. An example from the piece: while a model is running, the GPU's memory isn't taken up by the model alone. It's also taken up by a cache that grows with the number of simultaneous requests, known as the KV cache. As load increases, this can leave the system short on memory even when there's still enough raw compute available. Adding more compute without understanding this bottleneck, according to EXL, can simply make an inefficient system more expensive without solving the actual problem.

Argument three: build the system for future change, not today's model. Today's best model, according to EXL, will most likely not stay the best tomorrow: models change, hardware changes, new optimization methods appear. According to EXL, if the architecture around the model is built correctly, switching to a new model can become an ordinary configuration task rather than a reason to rebuild the whole system. That's the company's own promise, not a rule that always holds.

What here is checkable fact, and what is only a vendor's opinion

I don't have an independent source confirming or disproving the technical description of KV cache growth: it's only EXL's own claim from the piece. I treat it as a plausible technical explanation from one interested company, not as a fact I've verified against outside data. The recommendations along the lines of "build exactly this kind of harness, measure bottlenecks exactly this way" remain promotional material: there isn't a single number from a real client project, a study, or a comparison with alternative approaches in it. The text contains no specific example of a company that applied EXL's advice and got a measurable result.

What a small business owner should take from this

  1. Don't pick a model before you've described the task. Before comparing neural networks against each other, write down exactly what your AI tool needs to do: answer a customer instantly, process documents overnight, or something else.
    • Who does it: you, or the person responsible for AI adoption at your company.
    • How to verify: you have a one-page document describing the task and the required response speed before you start comparing models or vendors.
  2. Ask a vendor about the cost of one request in production, not just the quality of an answer in a demo. A demo with one or two examples doesn't show what happens when dozens of users are working at the same time.
    • Who does it: you, when discussing terms with a vendor.
    • How to verify: the vendor has named a price for one of your company's typical requests at expected load, not just an overall monthly or license price.
  3. Test response speed under real load, not in a quiet room. A tool that answers quickly on its own can slow down once several employees are using it at the same time.
    • Who does it: your technical specialist, or the vendor at your request.
    • How to verify: ask for a test run with a number of simultaneous requests close to your real peak-hour load, and record the response time in writing.
  4. Ask the vendor what happens when the model changes in the future. If the whole system is tied to one specific model, switching to a newer or cheaper model can turn into an expensive rebuild.
    • Who does it: you, when signing the contract.
    • How to verify: the contract or technical spec states that switching to another model doesn't require rebuilding the whole system, or names an approximate cost for such a switch.
  5. If you're buying a ready-made service at a fixed price rather than building a system yourself. Points 2 and 3 above, about cost per request and speed under load, are built for a company that configures the architecture itself or commissions it for its own load; you can skip them if your plan's price doesn't depend on the number of requests and you almost never have many simultaneous users. Points 1 and 4, about describing the task and a future model change, apply to any option, including a ready-made service.
    • Who does it: you, when choosing between a ready-made service and building your own.
    • How to verify: ask the ready-made service's vendor whether price and response speed change as your number of users grows, and what happens to your data and settings if the vendor switches the model inside their service.

What this news does not support

  • That this is independent expert analysis from MIT Technology Review. The piece is paid for and written by EXL; the publication's editorial staff did not write it.
  • That EXL's principles are backed by numbers or case studies. The piece contains no measurable example from a specific company, only general engineering arguments.
  • That an enterprise AI project won't work without services from a consulting firm exactly like this one. That's a predictable conclusion of a piece written by the company that sells exactly that service, not a proven fact.

Bottom line

On October 22, 2026, MIT Technology Review's platform published a piece paid for by EXL arguing that in enterprise AI, what matters isn't the choice of model but the engineering system built around it. The general principle is sound on its own terms: start from the task, measure bottlenecks one at a time, build a system that's easy to update. But the piece itself is an ad for consulting services, without a single independent example or figure, and that's worth keeping in mind when reading "expert opinions" like this from companies that make money on them. If you're considering AI adoption at your company, the practical value lies elsewhere: describe the task in advance, ask a vendor for a price under real load, and find out what happens when the model changes in the future.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources