Putting AI Agents to Work: 8 Problems Every Company Solves From Scratch

Getting AI to do productive work in a company takes more than a prompt.

Most people's experience of AI is a chat window: you ask, it answers, and it is genuinely useful. Moving from that to real work inside a business is a different problem entirely. The work has to run on the company's data, inside the company's systems, under the company's rules — and it has to produce the same result on the hundredth run as on the first.

None of that is a prompting problem. It is not a model problem either; the models are remarkable, and there is a better one every few weeks. The problem is how little there is around them.

Think about how websites were built in 1995. Every team solved the same problems from scratch, because there was nothing to build on. Then frameworks arrived, turned those problems into shared building blocks, and building a website stopped being a specialist craft.

Putting AI agents to work in a company today sits somewhere before that point. No standard stack, no shared building blocks, every company solving the same hard problems on its own.

“No standard stack, no shared building blocks, every company solving the same hard problems on its own.”
Jörg Bröker

And the problems are genuinely hard. Anyone trying to put AI agents to work in a business runs into the same obstacles, and these are only a few of them.

Choosing the right model

Every task needs a model with enough capability to do the job and a price that makes the job worth doing. Get it wrong in one direction and the output is unusable; get it wrong in the other and the economics collapse. The difficulty is not making the decision once. It is that new models appear every few weeks, and each release quietly invalidates part of the mapping you built. A choice that was correct in March is merely expensive by June.

Budgets

Every agent action costs money, and unlike a software license the bill scales with use. The first question is what you are spending and on what; the second is making sure nobody runs up a large bill by accident, with an agent caught in a loop or a workflow triggered far more often than anyone expected. Answering both means budgets at more than one level — a global ceiling, but also limits per task, per person, per department, per whatever unit the money actually belongs to — and it has to stay simple enough to administer, because a cost control system that needs its own team to run will not get used.

Data protection

The strongest models are, for the most part, hosted outside Europe. That creates a standing tension between capability and compliance that every European company has to resolve for itself, case by case, dataset by dataset. Some organizations go further and require fully local hosting, which narrows the field again. None of this is optional and none of it is solved by the model providers.

Reliability

A demo is relatively easy. A process is not. Getting consistent, reproducible results from agents and AI-driven workflows — the same input producing an equivalent, correct output on the thousandth run as on the first — is a different discipline from getting one plausible-looking answer. Most organizations discover the gap only after the demo has already convinced someone.

Maintenance

Nothing in this stack holds still. Models are replaced, prompts are improved, the software underneath is updated, and every one of those changes is an improvement in isolation that can quietly break a workflow which has been running correctly for months. Conventional software has an answer to this: you write tests, and the tests tell you when a change broke something. With agents there is no fixed output to compare against, so the equivalent has to be built from scratch — a way to confirm that a workflow still does what it did before, checked before the change reaches production rather than after. Most companies work this out the first time something breaks without anyone noticing.

Integration

An agent that cannot reach the systems where the work actually lives is a toy. Connecting agents across a company means giving them safe, structured access to existing applications through APIs, and the calibration is unforgiving in both directions. Grant too much and you are one bad decision away from a headline nobody wants to be in: company lets AI agent delete production data. Grant too little and you have a system that can look at a few things and change nothing, which leaves the entire benefit on the table.

Governance

Every agent needs an explicit answer to three questions: what is it permitted to do, when must it stop and ask a person, and what did it actually do. Permissions have to be narrow enough to be meaningful and every action logged well enough to survive an audit — ordinary enterprise practice for human users, built from nothing for agents. The middle question is the hardest. Set the threshold too low and everything needs approval, which defeats the purpose and turns the human into someone who clicks yes all day without reading; set it too high and decisions that needed judgment get made without any, occasionally with consequences you cannot undo.

Development

Software development is the one area where agents already deliver massive productivity gains. That is the clearest early win AI has produced, and it is not seriously disputed by anyone who has tried it.

But the gains are not free. For every tech stack, someone has to glue the tools together and build the workflows that let agents work effectively across the whole cycle — code, build, test, review, commit, document, release and deploy. What you assemble for one stack does not carry over to the next. And it is work that has to be done before the real work can even start, again and again, on every project.

So companies hire experts

Each of these demands real expertise and hard-won experience. So companies do what they did at the dawn of the internet: they hire experts. Internal specialists or external consultants are brought in to work out what AI could do for the business and how to implement it — and they reinvent a bespoke solution to every one of these problems, on every project.

The experts are scarce. The engagements are expensive. The results vary wildly with the skill of whoever happens to be in the room. And every hour spent solving these problems again is an hour not spent on the application itself — on the actual business problem you set out to solve.

This is precisely the 1995 pattern. In 1995 you hired a web agency because there was no other way to get a website, and what you got depended entirely on who you hired.

Then it got easier, in steps. Visual editors like Dreamweaver meant you no longer had to hand-write HTML. Server-side scripting with a database behind it — the LAMP stack — meant pages could be generated from data instead of written one at a time. Content management systems like WordPress meant you no longer needed a developer to publish a page at all. Frameworks like React meant developers stopped rebuilding the same foundations on every project.

Each step took something that had required a specialist and turned it into something everyone could simply assume was there.

“Each step took something that had required a specialist and turned it into something everyone could simply assume was there.”
Jörg Bröker

If only...

The list above is not complete. These are only the very obvious ones, and more will appear as AI applications mature and expectations rise with them. Compare a Dreamweaver site from 1998 with a modern responsive React application: nobody in 1998 had responsive layouts on their list of problems, because nobody expected a website to work on a phone. The bar moves, and the list of things you have to get right moves with it.

The need is obvious. A reliable, repeatable way to meet it is not yet available. So every company putting AI agents to work today pays for the same solutions to be invented again, from scratch, by someone slightly different each time. It is slow, it is expensive, and none of it is the work anyone actually set out to do.

If only someone would build a platform that solved all of this for you...