Agents Solve Problems. Workflows Solve Them Once.

For eight years running, I brought a new cohort of graduates into a fintech business. Same intake every autumn: bright, keen, and slightly terrified of the client reporting. My job was to train them, empower them, and point them at the real work: report writing, QA, documentation, configuring how clients got their numbers.

The best ones always did the same thing. They didn't just get through the task in front of them, they automated it. A grad who spent a fortnight untangling a fiddly client reporting configuration would build it so the fiddly part never came back, then go looking for a harder problem. That is what good looks like: solve it once, properly, and bank the result.

Now imagine the opposite. Imagine that instead of banking anything, I hired a brilliant new graduate every single week, briefed them from scratch, and asked them to solve the exact same problem. The same reconciliation, the same reporting config, the same QA pass, over and over, forever. Nobody would let that run for long, and rightly so, because it's an absurd way to burn talent, time and money.

An agent is that brilliant graduate on their first day. You hand it an open problem, give it some tools, and let it work out the path for itself. It reasons, it adapts, and it improvises its way to an answer. That is exactly what you want when the problem is genuinely new: messy inputs, no fixed route, and a destination you can describe but not yet map.

Think investigation, research, or the first triage of something nobody has seen before. When the value is in figuring out how, agency is the entire point, and that is where an agent earns its keep.

Here is the slightly less glamorous truth: most business is not new. It is the same shape, over and over. A refund, a KYC check, a month-end client report, an incident response. You already know the steps, you know who signs off, and you know what "done" is supposed to look like when it fails as well as when it works.

That is not a place for improvisation. What you want is the same correct outcome every time, an audit trail you can stand behind, and a human in the loop at the points that matter. That is a workflow: a known business process, encoded once, and running reliably ever after.

Asking an agent to reason its way through a settled process on every single run is just the every-week-graduate again in a smarter suit. It is slower, it is less predictable, and as we will get to, it is not cheap. Less predictable is not a figure of speech either. When researchers at Carnegie Mellon and Salesforce measured how often AI agents actually completed multi-step office tasks correctly, the answer came out at around a third. That is perfectly fine when you are exploring a new problem. It is not fine for something that has to be right every time, like a client's month-end report.

None of this is really agents versus workflows, and the most interesting systems happily use both.

Pull the intent out of a scruffy client email, summarise a case file, or classify the odd request that refuses to fit the usual buckets. That is agency applied to the single subtask that needs it, sitting inside a structure that stays predictable everywhere else.

There is a deeper reason to build it this way, and it comes down to who is actually in control.

This is not a hypothetical worry. Gartner expects more than 40% of agentic AI projects to be scrapped by the end of 2027, and the reasons it gives are worth sitting with: escalating costs, unclear business value, and inadequate risk controls. Notice what is missing from that list. Nobody is saying the models are not clever enough. The projects fall over because nobody kept hold of the reins.

When you build one big, overarching agent and hand it the keys to everything, you are letting a single model run your business logic. It decides the steps, it decides the order, and it decides when it is finished. That feels like magic right up until an auditor asks why it did what it did, or the monthly bill lands.

A workflow turns that arrangement around. You own the logic and the sequence, and the LLM becomes a component you call at specific steps, on your terms. You decide which model handles which task, how much it is allowed to spend, and what happens to its output before anything moves forward. The AI works for the process, rather than the process bending around the AI.

That matters commercially as much as operationally. If your whole solution is one enormous agent wired to a single provider's model, you are locked in. Prices move, models get deprecated, terms change, and you are simply along for the ride. When AI is instead a set of discrete, swappable steps inside a workflow you control, you can route the cheap high-volume tasks to a small model, save the expensive frontier model for the genuinely hard ones, and change providers without rebuilding everything from scratch. You keep your options, and you keep your margin.

There is a more everyday version of the cost problem too. Every time an agent reasons from scratch, you pay for it in tokens, and solving an already-solved problem from first principles is about the most expensive way there is to get an answer you already had.

A workflow encodes that answer once and then runs it a thousand times for little more than the cost of running it. That frees you to spend your AI budget on the problems that are actually new, rather than re-deriving your reporting logic every month-end. This is starting to show up on real balance sheets: McKinsey's 2026 survey flags AI running costs, tokens included, as a growing consideration for organisations as they scale. Think of it as caching your best graduate's work instead of re-hiring them each week and hoping they reach the same conclusion.

None of this is me talking the technology down. The most sobering finding of the last year came from MIT, whose research put around 95% of enterprise generative AI pilots in the "no measurable return" column. The small handful that actually delivered had one thing in common: they were not generic, do-everything agents. They were narrow, built around a specific process, and embedded into a real workflow. The winners, in other words, put AI inside a frame.

To be fair to the other side of the argument, this is a fast-moving field and the agents are getting better quickly. Adoption is climbing with them: McKinsey found the share of large enterprises scaling AI agents rose from 27% to 40% in a single year, and today's failure rates will look dated soon enough. But a more capable agent does not change the underlying logic. A better graduate is still wasted re-solving the same problem every week.

That is the whole point. AI is genuinely astonishing, and the temptation is to point it at everything and let it loose. But raw capability with no structure around it is just a very confident intern with no oversight and access to the company credit card.

The organisations getting real value out of this are not agonising over agents versus workflows. They treat the workflow as the frame that holds everything together, giving them determinism, control, auditability and human sign-off, and then they place AI agency exactly where it earns its keep. It is the whole idea behind how we think about workflows at Next Matter: give the process a solid structure first, then let AI do the genuinely clever bit inside it.