The pendulum is swinging back.

For two years the answer to everything has been to point an AI agent at it. Nex belongs to a breed of startup taking control back, handing most of the job to ordinary deterministic code and leaving the agent only the parts that genuinely move.

What they built

Nex builds and runs the complex, high-volume sales and marketing workflows that general-purpose agents struggle with. The company calls it GTM engineering, which is the work of connecting data, tools and outreach so a sales team reaches the right people.

Read their own description again and notice what it admits: the workflows that general-purpose agents struggle with. That is the company's starting position, not a caveat buried at the end.

The founders ran HubSpot's three largest customer platforms. Najmuzzaman Mohammad is chief executive, with Francisco Dias as co-founder, and the company is backed by HubSpot founder Dharmesh Shah. Its site is nex.ai.

Sales tooling is possibly the most crowded market in software. HubSpot and Salesforce are the systems most teams run on. Apollo, ZoomInfo and Clay supply and enrich the contacts, with Clay in particular having made this kind of workflow building popular. Outreach and Salesloft run the sequences. 11x and Artisan sell AI sales representatives outright, and Relevance AI and Lindy sell agent teams. Most of the newer ones answer every question with more agent.

The point: taking the work back from the general agent

Here is how the opportunity was created.

Software opened its gates. Almost everything now has an interface an agent can reach, so a general AI agent can in principle do anything: read the record, enrich it, decide, write it back, send the message.

In principle. At volume it goes wrong in three ways at once.

It is slow, because thinking through each step takes real time and the work is genuinely complicated. It is error-prone, because a system that reasons afresh every time will occasionally reason differently, and across a hundred thousand records occasionally means often. And it is expensive, because you are paying a model to think about the same thing again and again.

Run that across the volume a real sales operation handles and all three problems multiply together.

Most of these workflows do not need thinking. They need doing, correctly, a hundred thousand times. That is what ordinary code has always been good at, and asking a model to do it instead is slow, unreliable and costly for no benefit.

So somebody who actually knows the work comes in, writes the known parts as deterministic code, and leaves the agent only where the job genuinely varies: the awkward record, the judgement call, the thing that needs language. The moving parts get an agent. Everything else gets code, which is fast, cheap and does the same thing every time.

Why this window has only just opened

This could not have been built two years ago, and that timing is the whole point.

First people had to try the all-AI version. They had to point agents at everything, get real productivity out of it, and then hit the wall: the bill arriving at the end of the month, the run that took nine hours, the four hundred records handled wrongly in a way nobody caught.

Only after that does anybody know which parts genuinely needed a model and which were just being done expensively. You cannot draw that line in advance. You learn it by watching the general version fail in specific places.

That is why the founders' background matters here more than the technology. Having run HubSpot's largest customer platforms means knowing exactly which parts of this work are fixed and which actually vary, and that knowledge is the product. An outsider building the same thing would guess the line wrong and end up with either a rigid system or an expensive one.

The formula, and where it ends up

This is my opinion, not something the company has said.

The same move is available across a great deal of software right now, and it makes a usable formula.

Find somewhere people enthusiastically applied a general agent. Wait for them to discover it is slow, costly and occasionally wrong at scale. Then rebuild it properly: deterministic code for everything that is actually fixed, with a few deliberate gates where an agent handles what genuinely varies. Charge for the result.

The candidates are easy to spot. Anywhere an agent is being asked to do the same task tens of thousands of times. Document processing, data cleaning, reconciliation, ticket routing, compliance checks, onboarding flows. All places where the AI version works impressively in a demonstration and hurts at volume.

The product that wins in the end will not be the most AI-driven one. It will be deterministic code and AI combined in the right proportion, with the line between them drawn deliberately by a human who understands the work.

That is a less exciting sentence than "an agent does everything", and it is much more likely to be what people are actually running in five years.

Why it is bigger than it sounds

The short description is AI for sales workflows. The claim is about how software gets built from here.

  • Most of this work does not need thinking. It needs doing correctly at volume, which is what ordinary code has always been good at and what a model is slow, costly and unreliable at.
  • You can only draw the line after the all-AI version fails. Which parts genuinely vary is learned by watching a general agent break in specific places, not by planning.
  • The winning shape is a deliberate mix. Deterministic code for the fixed parts, an agent at a few chosen gates, with a human who knows the work deciding where the boundary sits.

What to watch

The number is cost per completed workflow against what the customer was paying when an agent did the whole thing. That is the argument stated as a measurement, and it should be dramatic rather than marginal if the thesis holds.

The second is how many customers one person at Nex can carry, and whether it is rising. Running the work is a service, and if the deterministic half is real the ratio climbs. If it is flat, this is an agency with good tooling.

The third is where they put the gates. Ask which parts they gave to the model and why, because that answer is the entire intellectual property, and a team that cannot explain it clearly has probably not found the line yet.