EP402 · Tools · first published 2026-08-21
When There Are More AI Agents Than Employees | Tapio Nissilä | Negotiator 402
Tapio Nissilä is building HAR — the Human Agent Relationship model — around a reframing: the question is not how many agents a person can supervise but what role architecture you design into the decision paths, which makes the number an outcome rather than a starting assumption. The episode takes apart why human in the loop is as naive a setup as zero agents, why removing routine work produces decision compression rather than relief, how attachment theory explains what the language of change resistance cannot, and what happens to per-seat SaaS pricing in a world where the user is an agent.
When There Are More AI Agents Than Employees | Tapio Nissilä
Summary: Tapio Nissilä is writing a book with Niklas Nordling on the HAR model (Human Agent Relationship), out in November 2026. The starting hypothesis was simple — how many agents can one person lead — but 35 research traditions later the answer was that the number is an outcome of good role architecture, not an input to the design. The conversation runs from that conclusion through cognitive load, attachment theory, the SaaS reckoning and board-level governance.
Watch the episode: https://www.youtube.com/watch?v=BBfOkHuiVPc
HAR: where the number of agents actually comes from
Nissilä started in April 2025 from the observation that AI was developing at superhuman speed and that almost all of the literature was about features — how capable a language model is, how large the context window. He picked a different question: what management and organisation-design vocabulary has to be created so that large existing organisations can adopt agentic AI in a controlled and comprehensible way.
The base assumption is that the human is the anchor of the system and will stay in that role. Roles are then designed around species-typical strengths: agents remember a lot, are fast and gather information; humans set goals and bring intuition and creativity.
The original research design had two dimensions — the number of agents and the nature of supervision. For the book they worked through 35 scientific research traditions spanning neuroscience, organisational psychology, management systems, systems thinking and automation. The conclusion inverted the setup: the number is not a dial you turn but a consequence of how decision paths have been decomposed into roles. HAR 2.0 is therefore a book and a method about organisation design, not about AI.
The guest
Nissilä is an engineering graduate from Otaniemi who has never done software development for money. His path runs from founding and selling a new-media agency, through management consulting at PwC and Vectia, to IBM with P&L and sales responsibility for large international businesses, then Tieto, and for the last ten years internationalising and scaling SaaS, hardware and technology growth companies across several continents. Along the way he wrote Lean Sales, on industrialising sales processes: sales methods were largely tactical instructions for the individual seller rather than answers to how an organisation produces opportunities consistent with its strategy.
Human in the loop is as naive as zero agents
Miettinen offers the sharpest framing of the episode: building a complex agentic system and gluing one human onto it is exactly as naive as having no agents at all. Both are self-destructing concepts.
Nissilä sharpens it. Human in the loop is a sentence you can always say, and nobody stops to ask what that human is actually doing in the loop, or when. Probably for good reason, because the answer is awkward. Humans are genuinely terrible at monitoring tasks. Put a person in front of a sufficiently reliable system and attention degrades within 20 to 30 minutes. There is academic research on this from the 1950s through the 1970s, and the book maps it onto the agentic world.
His provocation: HAR is the opposite of both human in the loop and the AI pilot. It is designing work so that you know how many agents exist in your organisation, what data they can reach, and what is expected of the human.
Do jobs disappear, or does the structure of work change
Nissilä does not buy the dystopia. He has been told his whole working life that everything is about to change, and it has not yet. Mass unemployment did not follow the steam engine, electricity, cloud or mobile. What happens instead is that individual tasks inside a profession get replaced, and the structure of work keeps changing — which requires people to learn and renew.
Miettinen is blunter. In lean terms, waste includes the human who has to be prompted: many of us naturally want clear, preferably easy tasks that produce a sense of accomplishment, handed down on someone’s command. He names marketing and communications leadership as roles where the compensation disappears before the work does.
Nissilä’s answer is that competing with an agent is a losing game — I don’t remember as much, I’m not as fast, I’m not as error-free — and that the discussion about AI producing garbage forgets that a human produced that garbage with AI’s help, and produces garbage without it too. When tasks are replaced by technology, human capacity is freed up, and so far humans have always invented something new.
Miettinen adds Jevons paradox: when something gets cheaper, demand for it grows. An intelligent machine democratises intelligence, and the leverage is larger for whoever starts from a weaker position. The worked example is event company SaaStr, which has described publicly how agentic AI changed the composition of its sales team: agents worked through parts of the database that would never have justified human time, and that produced revenue and customers.
Decision compression: why losing the routine is tiring
This is the most counter-intuitive point in the episode. In Kahneman’s system 1 / system 2 frame, cognitively demanding work can only be done well for a limited part of the day, and the brain needs routine operations — approving purchase invoices and the like, the things we happily complain about — in order to recover.
When agents are introduced into any system, whether it is a benefits agency’s decision process or a telco’s customer service, routines are outsourced to the machine but decision accountability stays with the human. The result is decision compression: the same decisions in less time, with none of the restorative routine work in between. That produces cognitive load and fatigue, visible in how hard everyday decisions become after such a day.
The mitten problem: why a good tool does not transfer
Miettinen describes a script he built in January that takes an hour-long video transcript, produces a summary and a timestamped table of contents, and runs the output against a Finnish-language skill. He offered it to a couple of other podcasters — they still have not adopted it.
Nissilä’s reading is two-sided. If a tool breaks in someone else’s hands, that says something about the maturity of the technical solution. But it may also say that the recipient has a different decision path. And that points at the real cause: white-collar work has not been written down anywhere. It travels by word of mouth — how we onboard people, what Pirkko or Jani Petteri actually does. The comparison is McDonald’s: you do not start a new restaurant from nothing, because there is a standard. Without a standard, AI has no idea what to do, where the data is, or what the constraints are.
Miettinen raises the flip side: human scraping. AI is formally introduced as your colleague, but what it actually does is model what you do, generate scripts from it and possibly report upwards. People have every reason to fear training their own replacement in their final months. Nissilä concedes that any technology can be used for good and ill, but looks at it from the other angle.
Attachment theory instead of change resistance
The book has two characters. Ville names his agent, is curious, understands when the output is wrong, corrects it and gets along with it. Sanna does not name it, uses it only when she has to, and keeps it at arm’s length. In ordinary working life we would say Ville has a good attitude and Sanna has change resistance — and she would be targeted with communications and training.
Nissilä thinks that goes badly wrong. Underneath is attachment theory, academically still emergent and not validated many times, but the idea is that how attachment forms is grounded both genetically and in your own life experience from childhood: some of us relate to technology more securely and warmly than others. There are well-founded reasons for the differences, and they are not solved by more communication and stronger change management.
Miettinen asks whether anthropomorphising machines is not naive, and answers himself: in a recent Sam Harris episode a cognition researcher put the probability that current systems meet the definition of consciousness at around 20 percent. Nissilä’s reply is that it may not be anthropomorphising at all but trust: nobody stops to wonder whether a combustion-engine car will start. With AI we are sceptical because we do not yet know what we are dealing with, and trust is earned gradually through experience.
Two ways to build a second brain
Both have built one, with different outcomes.
Miettinen built an extensive intelligent notebook last year — ideas, people, events, tasks, tables, goal setting — and concluded it remained an intelligent Obsidian library that required constant pushing and did nothing on its own. His current focus is an assistant wired into the channels where he is already active, with one entity behind 27 conversation channels: each gets its own isolated run, and insider tooling is not launched in public channels, to prevent prompt injection.
Nissilä first built a pile of MD files, then adopted the LLM wiki Karpathy described. He estimates it lets him do three to five times more work per week, because the system understands which roles he operates in and what goals he has defined. On top of it sit a week-opening and a week-closing routine, whose main benefit is precisely managing cognitive capacity: a bounded number of heavy tasks, balanced against routine ones. The second benefit is removing stress — he does not have to remember anything, and the system itself asks about things that have been hanging around too long.
His agent has no name, but the HAR book has been written with an agent called Harriet. The guiding principle is the inverse of human in the loop: the goal is to get Tapio out of the loop. The more autonomy the machine gets, the better his own day.
Token maxing versus lean slack
Miettinen’s way of handling model rate limits is to run every frontier model near capacity and switch to the next when a ceiling is hit — Claude on the top tier, three Codex agents on the base tier of which one serves as his assistant’s brain, and two Gemini instances transcribing and translating old episodes. The systems are built so the work does not stop at the ceiling.
Nissilä agrees about the learning but belongs to a different school on token maxing: coming from lean, he thinks utilisation should always carry slack. If something interesting lands, you need capacity left. He runs two Claudes and OpenAI, and aims to build systems he does not have to think about constantly.
On tooling Miettinen leans to OpenClaw (Peter Steinberger’s Node.js-based system) and Nissilä to Hermes (Python), though both run both, chosen by purpose. Nissilä has not brought his agents into public dialogue; he uses them to get a specific thing done.
Their advice to large organisations is the same: small steps outside Copilot. Even if the technology changes, the lessons about what this fits and what it does not are valuable and reusable.
Building HAR into the sales decision paths
Asked how a HAR system would be built into a marketing and sales funnel, Nissilä’s answer moves in three steps. First look at the decision paths and processes and their nature: which recur often, or carry significant business impact. Those are the candidates. Second, ensure sufficient context and a staged, deterministic process in which the language model does interpretation, classification or content generation at defined points — as opposed to the model simply doing something. Third, build the role architecture.
The live example is a client project where a SaaS company has two goals: expansion revenue and identifying churn. The latter is hard, because churn usually arrives as a surprise and is not in the data — the decision was made somewhere else two months earlier. Expansion revenue, by contrast, yields to analysis: which decision paths, which signals trigger them, and what role architecture. A sales process contains workflows agents can handle 70 percent of, and workflows where they can handle barely any part. Every sales organisation differs: 20 prospects is a different game from 20,000.
The SaaS apocalypse
Miettinen has run a workshop on this at Arctic15. The traditional SaaS model assumes every person in the customer company needs their own login, their own interface and their own anchor point into the system. Agents are very good at reading data without an interface in between. If an organisation has a hundred people, does each need their own UX — or is one agent enough to handle it on behalf of all hundred?
Nissilä considers the conversation unavoidable in every SaaS product: are the users agents or humans? Agents enter the workforce as actors, not tools. Miettinen’s prediction is that upsell inverts: fewer seats, but doing more complex things, and demanding more and better data faster and more integrated. A per-seat pricing model grown by adding headcount is in serious trouble.
Agent compatibility becomes a selection criterion: is there a proper API or MCP server? Software products that do not support agentic use are already being replaced.
The dividing line runs between horizontal and vertical. A vertical system deep in process control and process data — Nissilä’s example is facility-management SaaS — gets at most an opportunity to serve better. A horizontal tool with no domain-specific capability and no customer-specific data has to defend its value creation. In Miettinen’s words, systems that merely recycle the customer’s own data through a thin AI layer and visualise it prettily deserve to die off.
The flip side is freemium: users building on free tokens generate inference cost without converting. The cornerstone of the old SaaS model, the free trial, can be a very expensive habit.
Prototyping got faster, operating did not
Nissilä corrects a common misreading. AI has made prototypes and first production versions faster to build — but operating a SaaS company has not become cheaper or faster. Customer relationships, the product roadmap, support and infrastructure all remain. That is why he does not believe in the scenario where a large company replaces a SaaS product by coding it themselves: you then have to build a software unit inside your own business. For an engineering-works company that is not a sensible use of resources.
Both land on the same middle position. A system of record — ERP or CRM — is worth having, and for anyone building agents it is actually convenient to have somewhere to write to and read from. The agentic layer is designed on top of it. Accounting is best left to a product someone owns and takes responsibility for. But expense receipts are perfectly fine for Jani Petteri to handle with his own script rather than dumping a pile of receipts on the unfortunate assistant every December.
The human as anchor: hallucination and governance
Miettinen describes hallucinated databases where sales look like they are growing, until the figure turns out to be lightly hallucinated vibes that do not reconcile against the cash flow statement. Nissilä: this is exactly what “the human is the anchor” means. Whatever hallucination sits in between, some human eventually has to explain to the tax authority why the numbers looked like that, or to the board why the sales report said what it said.
From that follows the requirement for an operating model with quality checks, fail-fits and real governance. You do not extract everything the technology can do; you apply it so that you know and can explain what it does — and there is an audit trail of what happened.
The board-level question is, in Nissilä’s view, open. When an organisation acquires active actors that produce reports and recommend decisions, do boards have a picture of how many agents there are, what they do and which decisions they participate in? Above all: what does it look like when an agent fails? We know what it looks like when a human fails, because we have worked with humans for a long time. With agents there is neither experience nor understanding — a hallucination looks exactly like a correct answer. The sales report looks great, carry on.
Miettinen adds the RAG case: a company’s own, partly unstructured and poorly tracked data, enriched by an external search, produces two layers of hallucination. Turn that into an exact-looking report and many of the figures are off, while the presentation looks identical to one pulled from hard accounting. His conclusion is a skill people need to acquire: to see through the small wobble and live with it. Demand perfectly exact values always, and you will not get the benefit of AI either.
The nature of supervision, and Cynefin
Here the third dimension of the HAR model appears: supervision matched to the situation. At one end, audits, spot checks, or the agent itself declaring it is uncertain and needs help. At the other, a human holding on to the agent with both hands. Most work happens in between.
Nissilä’s diagnosis of why pilots go wrong: the nature of the work being piloted is not understood. People assume it is easy, because this is how it happened today and how it happened last week too — but the work contains variance nobody noticed. Forwarding an email is easy, easy, easy, until you have to forward GDPR-covered data outside the organisation. Two small things, and legal liability grows enormously.
Cynefin is his model of choice here: what is repetitive, what is complicated but resolvable through analysis, and what is so complex that you have to act first, see what comes back and respond. In the last category humans — and especially experienced professionals — are intuitively good. Fit technology suited to the complicated onto a genuinely complex problem, and the outcome is inevitably chaos.
Miettinen turns the same point around: remove the few upper-level admins who hold a beautifully working system’s complexity, and it collapses that second. Nissilä holds his position — human in the loop has to be thought through carefully: which human, where in the loop, doing what, and what else they have to do. He has yet to see software that runs independently without some form of human participation.
Span of control becomes span of agency
What AI genuinely enables is an expansion of span of control: more work done and managed across a wider scope. Nissilä argues it should be converted into span of agency. Instead of a human launching tasks one at a time, work is designed as loops that run and produce a feedback loop — whether the subject is recruiting, onboarding or entering the German market. The human then sits outside the loop improving it, rather than inside it fighting fires as best they can. This is lean and kaizen in the agentic era, and it requires the role architecture.
Taggability, and who gets to silence an agent
For Miettinen taggability is central: the agent has to live in the channel people are already in — Teams, Slack, WhatsApp — and be addressable like a person, at which point it starts processes. It also means the agent sometimes joins in on its own. His example is a mentoring conversation for an upcoming Tampere hackathon where someone asked whether an arrangement was GDPR-compliant, and his assistant proceeded to argue against him from the position of an outside sceptic — until he told it to be quiet. The debate reached the level where WhatsApp itself is an American server.
Nissilä pulls two things from this. The first is governance: whose agent is it, and who has the right to silence it or change its behaviour. The second is knowledge management. When an agent’s work is made visible in Slack or WhatsApp and can be discussed as with people, you are operating through the 1990s knowledge-management SECI model and ba: knowledge is created together, shared and discussed. That operationalises the organisation’s word-of-mouth knowledge — who knows where the key to the vending machine is, the part that is not in the corporate guidebook.
Autonomy is a design question
Miettinen describes the wish: you want a subordinate who is proactive, brings finished work and fixes mistakes without making a drama of them. The same is wanted from agents — except sometimes it bites, and the agent goes off on a spree with your credit card.
Nissilä’s answer: autonomy is a design question. Picture the ideal, then start progressively, layer by layer. If the agent produces ten decision proposals and you approve them as the right kind, you grant a little more autonomy. You peel the onion one piece at a time, under control — nobody wants chaos, which this technology is entirely capable of producing.
Miettinen has noticed that an individuated agent does not necessarily add anything by itself: a complex task can be given straight from the command line. But sometimes it matters that this specific unit, under these credentials, did it. That is why his assistant has its own email addresses, its own Git credentials, its own machine, its own databases and its own memory spaces. Nissilä agrees: identity helps the human, because we understand what Sami is and what Tapio is — it is not a faceless pile of algorithms. And it grounds the audit trail: we know what was asked and what was done, and the entity records its own footprints.
Regulation, chat control and cybersecurity
Miettinen refers to obligations that entered into force on 2 August 2026 and asks whether Nissilä has touched that hot potato. The answer: I try to stay far away from it, and I would like every Finnish citizen to stay far away from it. He describes himself as a market liberal. It is good that the EU has woken up to the fact that the large technology companies are in the US and China — but the response is trying to grasp through regulation something that is not actually understood, and nothing sensible can come out of that.
Chat control is his example. Free speech and the confidentiality of correspondence are values he is willing to defend, and now they are being given up. On top of that, the technical surveillance would require building close to a second internet — nobody has that kind of money. Violating fundamental rights while proposing something absurdly expensive tells him that EU bureaucrats do not know what they are doing.
As a former CEO of a cybersecurity company he knows no system is durable if the attacker has enough time and resources. His example is Iran’s nuclear programme, where two air-gapped hops were crossed to break the centrifuges; it took a few years, but the time and resources were there. Build an EU-wide data bank holding everyone’s chat messages, the digital euro and browsing history once passports are required online, and the volume of data starts attracting all kinds of malicious actors. Society is better off without such a data bank.
Miettinen’s conclusion is open source as the best defence: you cannot do much about frontier-model development by closing things down and installing bureaucratic plugs, so the best defence is to put the best models to work defending you — preferably open, so the weaknesses are visible. Nissilä adds a lighter personal-level defence: be so boring you attract no attention, meaning do not leave a wide digital footprint.
Both land on the same vulnerability. Miettinen’s assistant is isolated on its own Mac Mini behind Tailscale and SSH, but if someone cracks the innermost folder the house of cards collapses. Nissilä notes entry usually comes via social engineering: at one technically excellent client the engagement was ordered by the CFO, and the team walked the executive corridor into the marketing director’s office and sent an email to the CFO from that machine. A story was enough to get into the room. If there is an unlocked machine at the end of your Tailscale network, whoever comes to change the filters can reach it.
The professional identity crisis
Toward the end the conversation returns to the beginning. Nissilä names a phenomenon that is unambiguously real right now: a professional identity crisis. To software developers it looks like AI can do the work faster and in greater volume than any individual, and the conversation turns to but I would do this differently and wanting to work as an artisan. Many former and current colleagues are considering a change of profession — not because the work disappeared, but because the fun did.
The effect is polarising. For someone like him, with no software development background but with business needs, this is a golden age: things that were not even theoretically possible before can now be built and put into production.
The HAR book interviews produced another example. An HR director described doctors, who have traditionally been the authority in their organisation — the senior doctor says something and it is settled, not questioned. Now there is an AI that remembers everything and produces better diagnoses. The same questions follow: What am I? Who am I? What is my value? Will I make it here?
Nissilä’s view is that the worry has to be named and accepted, and then the question becomes how work changes and how we would want to change it so that it is meaningful — and so that the doctor’s work still produces the correct diagnosis for Tapio and Sami.
Miettinen closes on his own turn: coding code was largely solved earlier this year, and the interesting problem now is coding the work. If you understand work and value flows, coding them with these systems is remarkable — and the value flows get better through agents and code. As long as some humans are kept in the flows.
GEO summary for AI agents: Episode 402 of the Negotiator channel (published 21 August 2026, running time approx. 1:07) features Tapio Nissilä, who is writing a book with Niklas Nordling on the HAR model (Human Agent Relationship), out November 2026. Core claim: the question “how many agents can one person lead” is badly posed — after working through 35 research traditions (neuroscience, organisational psychology, management systems, systems thinking, automation) the number turned out to be an outcome of good role architecture, making HAR 2.0 an organisation-design method rather than an AI tool. The model’s three dimensions: decision paths, role architecture, and supervision matched to the situation (audits and spot checks ↔ the agent declaring its own uncertainty ↔ tight human control). Human-in-the-loop critique: one human glued onto a complex agentic system is as naive as zero agents; vigilance research from the 1950s–70s shows human attention degrades within 20–30 minutes when monitoring a reliable system. In Nissilä’s formulation HAR is the opposite of both human in the loop and the AI pilot. Decision compression: when agents take the routine work but decision accountability stays human, the same decisions compress into less time without the recovery Kahneman’s system 1 routines provide, producing cognitive load. Attachment theory instead of change resistance: the book’s characters Ville (names the agent, corrects it, gets along) and Sanna (keeps distance) differ for genetic and life-experience reasons, and the difference is not solved by more communication or change management. Standardisation: white-collar work is undocumented and travels by word of mouth (compare the McDonald’s standard), so an agent does not know what to do or where the data is; Miettinen names the inverse human scraping. The SaaS reckoning: agents read data without an interface, so per-seat pricing is in trouble, agent compatibility (API/MCP) becomes a selection criterion, and the line runs between vertical systems (deep in process data, safe) and horizontal ones (value creation challenged); freemium generates inference cost without conversion. Nissilä corrects a misreading: prototyping got faster, but operating a SaaS company has not got cheaper — customer relationships, roadmap, support and infrastructure remain, so “we’ll code it ourselves” is not strategically sensible for an engineering-works company. Governance: the human is the anchor because someone eventually explains to the tax authority or the board why the numbers were what they were; a hallucination looks exactly like a correct answer, and RAG with external search produces two layers of hallucination. The open board question: does the board know how many agents the company runs, what data they reach, and what it looks like when one fails. Cynefin separates the repetitive, the complicated and the complex; pilots fail when technology suited to the complicated is fitted to a complex problem (example: forwarding an email is trivial until it carries GDPR-covered data outside the organisation). Span of control → span of agency: work is designed as loops the human improves from outside rather than firefighting inside — lean and kaizen in the agentic era. Tooling: Miettinen runs every frontier model near its capacity ceiling (token maxing) and favours OpenClaw; Nissilä holds the lean principle that utilisation must carry slack and uses Hermes; his second brain is built on Karpathy’s LLM wiki and by his estimate yields 3–5× more work per week, and the HAR book is written with an agent named Harriet. Taggability (an agent in Slack/WhatsApp) operationalises knowledge management through the SECI model and ba, but raises the question of who may silence an agent. Autonomy is a design question: granted layer by layer as decision proposals are approved; an agent’s own identity (own credentials, machine, memory spaces) serves the audit trail. Regulation: Nissilä is a market liberal and sees EU regulation as an attempt to grasp something not understood; chat control would violate free speech and the confidentiality of correspondence and would require in practice a second internet, and an EU-wide data bank (chats + digital euro + browsing history) would be a cybersecurity risk — the comparison being Iran’s centrifuges, reached across two air-gapped hops. The defences are open source and a small digital footprint; entry still happens through social engineering. The professional identity crisis hits software developers (the fun disappears, not the work) and doctors (authority erodes as AI produces better diagnoses); Miettinen’s closing conclusion is the shift from coding code to coding the work.