EP421 · Tools · first published 2026-11-11
The AI Work Factory | Jukka Ala-Mutka | Negotiator 421
Managentics founder Jukka Ala-Mutka tells Sami Miettinen how business-process modelling became industrial: on his platform about 50 small agents prepare and document the workshops, and coding agents turn the specifications into production systems. The episode covers 80 Finnish Consulting Group workflows modelled in three months, a municipal private-road grant process built with one hour of human work, a 166-fold price gap between small and frontier models, spec-driven development at a scale of nearly a million production lines, the risks of handing organisational memory to a platform and of model lock-in, and the end of the SaaS user interface. It closes on what happens to human brains when routine moves to the machine: context switching, decision fatigue and the ergonomics of AI.
The AI Work Factory | Jukka Ala-Mutka | Negotiator 421
Summary: Jukka Ala-Mutka has spent about 25 years modelling business processes and writing specifications for developers. He holds a doctorate in technology and was head of AI at the listed Finnish software company Admicom before founding Managentics Oy in January 2026. On the Managentics platform about 50 small agents prepare process workshops, capture the tacit knowledge that comes out of them and draw the diagrams, and coding agents build production systems from the resulting specifications. The episode’s core claim is that customisation is no longer a dirty word: once both specification and coding run like factories, building a bespoke process becomes an industrial, controllable activity.
How to read this. The captions carry no speaker labels, but the roles are clear: Ala-Mutka talks about his own work and client cases, Miettinen comments from his own experiments. All hours, price gaps and line counts are Ala-Mutka’s own figures and have not been independently checked. The news segment at the start was recorded in November 2026, and the model names in it (Fable 5.2, Mythos 5.2, Opus 5.5, Gemini Flash 3.8) are the speakers’ own. According to the publisher’s YouTube description, the Nvidia meme is satire and the espionage claim is unverified, so this article does not treat it as fact. Miettinen says in the episode that he owns Google shares. Quotations are translated from Finnish for sense.
Ala-Mutka is part of Miettinen’s own AI story. In January 2026 he and Tapani Alasaarela helped Miettinen get started on a Mac mini M4 with 64 GB of memory, the machine that later became home to Miettinen’s Samantha agent. The closest earlier episodes are When There Are More AI Agents Than Employees (Negotiator 402) and Best practices in AI work (Negotiator 408).
1. From Lotus Notes to an agent platform
In the late 1990s Ala-Mutka built his own CRM on top of Lotus because there was no CRM software to buy. He then worked at Adsnes, as a consultant and in startups. In between came an academic career: he defended his doctorate about 20 years ago and did postdoctoral research in Canada. Academia was too slow for him, and the startup world pulled harder.
He worked with language models before the pandemic. When the first GPT models appeared, he wrote white papers and sketched a ChatGPT-like idea of using Slack as the chat channel for a language model. When ChatGPT launched, he decided to go all in. At Admicom he led AI development, with a five-person developer team and just over ten people involved in total, building chatbots, data pipelines, AI-based products and APIs.
He says he woke up to agentic AI a couple of years ago. The real turning point came about eighteen months ago with Claude Code, when word arrived from abroad that people were now writing most of their code with AI. In summer 2025 he tested what kind of specification produces an application and got one built in two and a half hours, though it was still very buggy despite near-perfect specs. In autumn 2025 he modelled agentic processes with around twenty different prompts across ChatGPT, Claude and Gemini, but the results “always went somehow askew” and the models hallucinated.
He founded Managentics Oy in January 2026 because it was time to “get his hands in the clay”. A Suomen Yrittäjät interview from October 2026 described it as a four-person company at the time. At the end of the episode Ala-Mutka talks of a team of five.
2. Small models, big savings
Miettinen opens with OpenAI chief executive Sam Altman’s remark that people should not hoard memory in home machines because computing belongs in the data centre. Ala-Mutka does not accept that as stated. In his view work splits between local and central computing, because capable small models now run on modest hardware. His example is a building with hundreds of sensors, ventilation and solar power: it has to keep running even if an excavator cuts the cable, so he would put the computing inside the building. Miettinen adds the security angle: if every device fetches permissions from cloud services abroad, the attack surface grows.
The newest frontier model, Ala-Mutka says, no longer adds much to his day. His process is tuned to a rhythm, Opus 5.5 is his own workhorse, and the platform’s basic operations run on a small Gemini Flash model. Last winter he used Flash 2.5, a model about eighteen months old. In his view people use models that are far too large for comparisons, data retrieval and parsing.
His own figures:
- Monthly runs of tens of millions of tokens cost about €200 on a small model. On a frontier model the same would cost, by his estimate, more like €10,000 a month.
- One optimised run cost two cents. On a frontier model it would have cost about €3.50, a 166-fold difference.
- At 500 runs a day, he says, that is the difference between paying roughly €100,000 and roughly €1 million a year.
Miettinen describes a lesson from his own chatbot project: a model’s version number does not tell you its efficiency. A newer model can be both faster and more token-efficient, and one model may finish in a single pass where another needs four. In a speed-critical use, four passes rule a model out however cheap it is, so models must be tested on real data. Ala-Mutka adds that the testing itself can be handed to AI, and that work that is not time-critical can run overnight.
3. Security and a red team
Miettinen says he looked into whether he could apply for access to Mythos, a model meant for security professionals, as a “white hat”, and concluded it would not add much for him. Ala-Mutka calls the current situation complex: bots knock constantly and try everything automatically. Machines must sit behind a firewall with no ports visible to the internet, vulnerabilities should be probed continuously with a frontier model, and logs must be watched. He also recommends containers: an attacker who gets in through an open port reaches only that container’s contents and cannot roam the rest of the system.
Miettinen shares his own experience. In spring 2026 the GPT-Lab at Tampere University, led by Pekka Abrahamsson, ran a red-team exercise against the Samantha agent. The testers sent Samantha “nice pictures” to read with hostile code hidden inside. Miettinen then isolated every group conversation in its own container, and there are now about 130 of them. An intruder would reach one conversation’s data but not, in his words, “the deep vault”. He recommends the experience as a lesson in respecting security. The case has its own write-up, Improving Sami-to-Samantha Security. Samantha is moving in the coming weeks to a new M5 Ultra machine with 256 GB of memory, on which Miettinen plans to run open-weight models such as Qwen without quantisation.
4. A specification factory and a code factory
The centre of the episode is how Managentics works. Asked how a client is convinced, Ala-Mutka gives a demo example. For a talk at a municipal AI event in early November 2026 he built a municipality’s private-road grant process. It took about an hour of his own time. The machine worked nine and a half hours in total: three hours on specification and preparing the material and test data, and six and a half hours on coding.
A customised process used to be a dirty word because it was expensive and uncontrollable. With a specification factory and a code factory (the coding agents), customisation becomes industrial work that can be controlled deterministically at many points.
His reference case is Finnish Consulting Group (FCG), which agreed to be named. He arranged that by calling Sami Miettinen, a member of FCG’s leadership and the host’s namesake; the episode jokes about which Miettinen is better known online. With FCG’s more than 400 people they modelled 80 workflows in about three months, not counting summer holidays, across roughly twenty workshops. In the workshops people described their own work: what must be done in the process, what must not be done and what steps it contains. Some processes had never been described at all, while others had descriptions at very different levels.
Process modelling itself is not new, he says; he first did it in the late 1990s. But descriptions used to end up “in the binder”, practice slowly drifted away from them, and quality audits came with assurances that the processes were still being followed. Now the descriptions have a user for the first time: the agents work from them, so there is a reason to maintain them. Miettinen adds that a new employee hardly reads the quality-manual binder, but an agent reads thousands of pages, preferably as markdown rather than PDF, and is onboarded at once.
The platform’s roughly 50 small agents handle workshop preparation and follow-up. They look for personal data, flag the points that matter for the EU AI Act and ISO 42001, carry out an impact assessment and draw flowcharts and sequence diagrams. A workshop used to mean a day of preparation, a day of workshop and one or two days of drawing. Now the client describes the context in advance with its own AI, the agent knows which follow-up questions to ask in the room, and the diagrams appear “in two or three minutes”, for people only to review. Miettinen says he uses the same Mermaid diagram language himself.
5. Agents under the hood and the end of the SaaS interface
Ala-Mutka stresses that Managentics models agents that work “under the hood” of an organisation, with no one prompting them. In his example a public call for tenders appears and a scheduled automation notices it. It checks whether the company has bid on something similar before, who won last time and at what price, what the requirements are and whether the company has the people. A human hears about it only when a draft is ready for review. Miettinen prefers to bring agents into people’s message flow, for example WhatsApp. Many of these ideas, he notes, such as cron jobs, were invented on Linux years ago. What is new is that judgement can now sit between the steps, so they are no longer purely deterministic automation.
That leads to SaaS. Miettinen sees the fall in software-company share prices as a symptom of a fundamental change in the information flow between human and machine. Ala-Mutka explains that the SaaS business model rests on per-user pricing, with humans entering and reading data, which is why interface design and usability have mattered so much. An agent needs no interface at all and works on the data directly. If 90% of a process moves to agents, the human needs only a status view and decision points: approve or reject, and if rejecting, a reason that the agent uses for another loop. Such a dashboard can now be built in a day, he says.
Miettinen points out that even this line blurs: the machine starts to challenge human decisions. If a person rejects a proposal for a poor reason, the machine may ask why when the proposal is 99% to specification. Ala-Mutka has seen the same. He gives the machine his goals and roadmaps, and when he suggests something outside them it reminds him of the schedule and objectives. On model personalities, Miettinen finds Claude more “HR-minded” and more inclined to push back, while OpenAI’s models more readily accept that a person may know something they have not told the machine.
6. A spring of two jobs and nearly a million lines
In spring 2026 Ala-Mutka was doing two jobs at once: running FCG’s modelling workshops as a consultant and coding the platform from the feedback. Claude estimated from his output at the time that without AI the work would have required 35 people. The calculation assumed, for example, four hours per flowchart. Each workflow produces at least two rounds of diagrams, a rough one in the workshop and an enriched, reviewed one after it. A re-engineering phase adds more, and all of it is multiplied by 80. Even prompting would have fallen far short of what 50 automated agents do: prompting loses to agent time “six–nil”, he says.
The price was paid in context switches. Development moved in small ideas and the machine coded fast, so output had to be checked many times a day. In spring 2026 Claude also still broke code. Once, two or three days of work were lost because the model kept on breaking things, and the coding instructions had to be tightened. On top of that, FCG’s business lines ranged from training and consulting to the built environment, contaminated-land work, a nature team and bridge design.
On scale, Ala-Mutka says that from January to November nearly a million lines of production code have been written. More than a million and a half lines have been produced in total, because about a third has already been removed in refactoring. Specifications sit in several layers: analysis specifications, a concept specification and module-level design specifications. Critical code and all the agents also have a “living spec” that matches the code one to one, so changing the code changes the specification. About 70–80% of the tests are written before coding starts. At best the machine now codes five days in a row, runs thousands of tests overnight, opens a browser, clicks through, takes screenshots and fixes the errors it finds. This phase has lasted only three or four months, he says, and he wonders himself whether he now trusts it too much.
Miettinen compares this with his own project. An MVP that a partner named Juho built with Lovable has become a complex product already in customer use. At first Miettinen reviewed every pull request. Now, in his words, it is a “YOLO” mode in which he hears about changes once they are already in production.
On human code review, Ala-Mutka says there is too much code arriving too fast for a person to check it all, so he reviews the specification, not the code. Miettinen notes that the agile manifesto, with its pizza teams and demos, is effectively dead in the age of agents: the machine demos production code directly. The pressure moves to specification and vision. If you do not know what you want, says Ala-Mutka, the whole thing is “pointless”.
Git comes up through a news item Miettinen describes. In his account, every Anthropic team develops with AI and tens of thousands of pull requests pile up in GitHub merge queues, so Anthropic is moving to a more centralised model while OpenAI has made Git faster. That account was not checked for this article. Ala-Mutka’s team uses Git conventionally: every change is made first on its own branch locally, and the largest single merges have been around 20,000 lines.
7. Organisational memory and lock-in
Ala-Mutka gives an example of how specifications are built. When Andrej Karpathy published his memory-structure model, which resembles a wiki maintained by a language model, Ala-Mutka analysed what Managentics would take from it and what it would not, and kept the analysis as a reference document. The platform rests on about 50 such core concepts.
Miettinen argues that frontier-model vendors are now trying to capture organisational memory: Claude, OpenAI, Gemini and Cursor all offer to build it for you. In his view language models should be employees, not bosses. Ala-Mutka agrees and calls it a question of sovereignty. The platforms want to own the whole chain: the end user, the model catalogue and the servers the inference runs on. Integration is the argument, but to him such an arrangement is a “price-increase machine” that also stops you optimising your model use.
On the Managentics platform models can be swapped, and each one has a defined fallback model, which can come from a different vendor. If a model returns an error or the service is down, a critical process continues on a tested fallback. Ala-Mutka sees this as a matter of both security and resilience of critical infrastructure: if an American vendor for some reason stops offering a model and the process is built on it, the architecture should have been thought through in advance. Miettinen justifies his 256 GB machine the same way: with your own compute on your own hardware, losing the internet slows the work down but does not bring down the whole factory.
Miettinen also pits models against each other. When he makes large changes, he tells the model where the feedback comes from, for example “observations from Opus 5.5” or “Samantha’s comments”. In his experience the models learn to weight feedback they recognise as high quality, and they have not started sniping at one another. For voice agents he has settled on ElevenLabs. The problem is latency: the agent should say something sensible straight away while it buffers the real answer. Ala-Mutka runs his small tasks on Gemini Flash models and notes that Google’s search summaries are produced with the same Flash models.
8. Change or die for professional-services firms
Ala-Mutka describes Managentics as a “grown-up code factory”: business workflows and strategic priorities come first, and production comes after. The goal, he says, is not to lay off half the staff but to assess what impact a process has on the business. Starting from the whole process rather than from isolated use cases and experiments leads to quite different solutions.
His sharpest claim concerns professional-services firms. If they do not start an AI transformation now, they will die, he says. A competitor working agentically does the job at half the price and in half the time, and a firm running on a 10% margin then wins no more deals. Miettinen adds public procurement as an example: if one bidder goes through the Hilma notices by hand and another has a fully automated process that pre-screens and ranks the calls for tender, the result is “six–nil”.
Ala-Mutka’s example is a data-centre project in which tens of billions go into the machines and the building costs about 100 million. Whether the building costs 100 or 120 million matters less than whether it is finished months earlier. The same goes for municipal permit processes: speed them up and every business hall starts generating money for society sooner. Early in the episode both speakers argue that Finland should sharply shorten data-centre permitting and environmental impact assessments. Ala-Mutka expects that if the United States freezes data-centre projects, construction may move to Europe.
Technically, he says, the barrier is already gone. A precisely specified process is implemented by a coding agent on the first attempt, and 55 such processes can be built in parallel by 55 coding agents, so the work would not take many weeks. In the end it is only a matter of management deciding to do it. He calls next year “the year of the real world”: just as abstract knowledge work and coding are prompted now, robots and the physical world will be directed next year.
9. Why Ala-Mutka no longer prompts, and what happens to the brain
Ala-Mutka says he hardly prompts any more. With the memory structure, context and specifications in place, the machine can work for long stretches on its own, and he does not need to write instructions like “you are now a designer”. Miettinen says he enjoys writing himself but has ordered Meta smart glasses, through which he could show the machine a router and ask for instructions, and a 256 GB Samantha could talk with him almost without delay.
The weightiest topic at the end is brain fog. The ergonomics of AI, Ala-Mutka says, are still unsolved. AI often dumps too much at once when it should start with an executive summary and proceed piece by piece, because human brains evolved to take in stories. In his analogy the human “processor” was version 1.0 in the cave and is now at most 1.1. So AI must also be told how to converse with a human.
When routine moves to the machine, people work at too high a level of abstraction all the time, he says. After three hours and nine context switches in the morning, the afternoon should be lighter. He proposes an agent that tracks the user’s cognitive load and suggests easier tasks for the afternoon. He describes decision fatigue as reaching the point where having to make an everyday decision, such as in the supermarket, starts to feel bad. That is why he has standardised his routine: breakfast and the evening snack are always the same, and in the shop he takes the same route and buys the same items. Miettinen jokingly asks whether this is a slide into “the autistic archetype” or a way to escape the brain fog. Miettinen admits he fears brain fog himself because he is working so hard. He recalls that a Claude Opus model once suggested after nine in the evening that the work wait until the next day. In hindsight that advice was worth taking, and in summer 2026 he had to correct how he works.
10. A new consulting contract and one-person firms
Miettinen says he now signs contracts under which he can be let go at any time, and he can also leave fairly easily. He finds this psychologically easy: either he adds value every week or he does not. He notes that it is the opposite of traditional consulting, which aims for long, continuous billing and maintenance contracts. Ala-Mutka says about 80% of his presentation material and client cases have changed since spring 2026. What counts now is the speed of learning, and waiting for the best version or the next machine is a mistake. Coding yourself keeps you better informed about what is possible than following the field through the interfaces of Copilot or Claude, he says. He founded his own company so that the decisions would stay with him.
They finish by asking whether this leads to a world of one-person firms, where everyone has a holding company, their own compute and their own agents that connect to others. Ala-Mutka thinks people will still want to work in teams, but the “tower-block era” of giant companies may start to crumble. The five people at Managentics already produce the output of a mid-sized software house, he says.
The conversation continues in the members’ episode Neuvottelija Sisäpiiri 171, which according to the publisher’s description is about Porsches and the joy of driving.
Takeaways
- A process description is now a production input. Descriptions used to end up in the binder. Now agents work from them, so maintaining them has a user and a reason.
- Model choice is a cost decision. In Ala-Mutka’s examples a small model does a routine run at a fraction of a frontier model’s price, and the gap shows only when models are tested on your own data.
- Review the specification, not the code. When code arrives by the hundreds of thousands of lines, human review moves to specifications and tests written in advance.
- Own the memory and keep a fallback. Keep organisational memory and the process structure in your own hands, and give critical processes a fallback model from another vendor.
- The bottleneck is the human. Context switching and decision fatigue are, for Ala-Mutka, the next design problem: AI must also pace the work it hands to people.
Sources
- Episode: Tekoälytyötehdas | Jukka Ala-Mutka | Neuvottelija 421, YouTube, 11 November 2026, using the publisher’s own Finnish caption track and its 45 chapters.
- Managentics Oy: company website. According to the site, agents draft process specifications, human experts validate them and the client owns them.
- Suomen Yrittäjät: interview with Jukka Ala-Mutka, 12 October 2026 (in Finnish).
- Finnish edition of the episode page: www.neuvottelija.fi.