---
title: "AI and GPT-3 in Negotiation | Tuomas Rasila | Negotiator 98"
summary: "Tuomas Rasila, founder of Stealth Black, builds assistive intelligence on top of GPT-3 that prepares sales and negotiation openings — but never sends anything itself. The episode's load-bearing claim is not efficiency but honesty: humans in a negotiation are prone to overselling, whereas a machine is not, because its opening is built from demonstrable facts. Rasila's most original concept is sales pollution. A badly prepared approach is not merely wasted time but a shared loss: the caller feels rejected, the recipient feels harassed, and both end up in a worse mood than before the conversation — and because only the deal won gets recorded, never the number of people bothered, the cost stays invisible. He likens it to emissions: polluting ought to cost something. The concrete cases are commercial property brokerage and corporate car sales, and the discussion connects to Timo Honkela's meaning negotiation and Martti Ahtisaari's principled negotiation, from which Miettinen derives a definition: a value is a useless word unless it is tied to behaviour you could call an algorithm — at which point it becomes a principle."
datePublished: 2021-09-10
dateModified: 2021-09-10
originalLang: en
section: tools
sections: ["tools","economy"]
authors: ["Sami Miettinen"]
tags: ["Negotiator","EP98","Sami Miettinen","Tuomas Rasila","Stealth Black","Artificial intelligence","GPT-3","Sales","Negotiation","Timo Honkela","Martti Ahtisaari","Commercial property"]
canonical: https://ai.neuvottelija.com/ep98-tekoaly-gpt-3-neuvottelu-tuomas-rasila/
---
# AI and GPT-3 in Negotiation | Tuomas Rasila | Negotiator 98

# AI and GPT-3 in Negotiation | Tuomas Rasila | Negotiator 98

> **Summary:**
> **Tuomas Rasila** builds assistive intelligence for sales and negotiation situations at Stealth Black, based on language models. The product's boundary is sharper than AI talk usually allows: **the machine never sends anything.** It advises a human, who decides.
>
> The weightiest idea is **sales pollution**. A bad opening is not merely wasted time but a shared loss whose cost goes unrecorded — because companies count euros per month, not how many people were bothered to earn them.

## A note on reading this

Rasila sells the technology described here, and Miettinen has written a book on negotiation that is referred to repeatedly. Views are attributed by speaker. This article takes no position on whether the product works as described.

---

## What GPT-3 can do — and what is interesting about it

The episode was originally meant to be recorded with **Risto Linturi**, but Linturi wanted to try whether the interview could be conducted by combining GPT-3 with a 3D avatar. Miettinen is still waiting for that episode.

Linturi's experiment produces the episode's sharpest single observation about what language models can do. GPT-3 was given as its only input that a pandemic called **COVID-19** exists and is spreading rapidly — but no recent data at all. Then it was put into conversation.

The predictions landed: a vaccine development cycle of about **nine months**, and the point that **variants may be more dangerous than the first wave.**

Miettinen asks whether GPT-3 scripts are actually geniuses. Rasila's answer dismantles the illusion without belittling it:

> *"Genius is a hard thing to define, but it is quite something that we have a machine that can, in a few seconds, summarise everything that has ever been written."*

And he names the mechanism that produced the prediction: **analogy.** The virus is analogous to earlier viruses that developed variants — *"so that outcome does not surprise me one bit."* It is exactly this boundless general knowledge that they use to open negotiations.

Miettinen offers his own framework: the **analytical dimension** in his book divides into engineering (analysing fact-based data) and artistry (surprising angles, connecting associations). When both combine, you get genius.

Rasila's answer is notably cautious, and it is the episode's most honest scoping:

> *"I would rather set the floor considerably higher, in that we can guarantee this kind of AI is always available — genius may be further away than making sure creativity is always present at some level."*

**The goal is therefore not the peak but the floor.** Not that the machine be the best negotiator, but that nobody opens a negotiation completely unprepared.

## The opening decides — which is why negotiation cannot be split off from sales

In Miettinen's framework, negotiation is the last stage of a pipeline that starts from marketing: marketing → sales → negotiation. He asks whether the split is meaningful.

Rasila's answer is practical. When a salesperson has to contact a stranger, two things must be done: **choose who to contact and choose what to say.**

> *"That is why you cannot separate it from the negotiation, because in that kind of negotiation the argument it was opened with, when, and to whom, makes up an enormous part of the outcome."*

Miettinen recognises the **yes ladder**: a good start can lead as a process to the closing handshake; a bad one kills it there and then.

## Anchoring price with data

The concrete example comes from commercial property brokerage, and it is the episode's most illustrative.

The situation: office space in Vallila needing a tenant. The machine identifies an IT company employing React Native developers currently located in Pitäjänmäki.

The anchoring rests not on a pitch but on a **demonstrable claim**: the availability of that kind of labour is considerably better in Vallila, the company pays a great deal for that labour, and it says in its own job advertisements that it needs more of it.

> *"All of that in one conversational opening — it certainly has a significant effect on how the negotiation goes."*

Miettinen asks directly whether recipients get a **creepy feeling** that the approacher knows their situation. Rasila says he has not run into that — and describes instead validating the model by phoning CEOs and asking outright whether the prediction holds.

## The ideal negotiation first, the counterparty only after

Rasila proposes an inversion that is the episode's most methodologically interesting:

> *"I would like to flip the idea around: what if we work out the ideal negotiation and only then look for the ideal counterparty."*

Miettinen recognises the same logic from his own work as a partner at Translink Corporate Finance: in an acquisition it would be best to start from the **most synergistic and logical buyer**, but the current owner often holds assumptions resting on nothing — *"of course our Swedish main competitor wants to buy us"* — and those are often wrong. The antidote is data searches on possible buyers and long-standing knowledge of their acquisition behaviour.

He adds one caveat Rasila does not dispute: **there always has to be a plan B.**

## "Any need for office space?"

Rasila states the problem in a form that blames nobody:

> *"I think everyone has heard the call that runs roughly: 'Would you happen to have any need for office space?' No. Bye."*

And he explicitly makes it **not** a question of laziness:

> *"The problem, to me, is not that people are stupid or lazy or anything like that — it is that it is an appallingly laborious process to think of something sensible to say to a hundred different strangers and approach them directly. It is simply too much for a human."*

**The diagnosis is capacity, not attitude** — which is exactly why the solution can be technical.

Miettinen says he does the same by hand: he sends board chairs a script telling them about his channel and offering a shared interviewee, and varies it by industry. Rasila's qualification is precise: that is classic segmentation, and the difference is that **with a machine the segmentation is fully dynamic.**

## Channels, and what the form of a message does

Rasila lists the channels and their differences. **LinkedIn** funnels a person with many hats into one data unit; **email** forces you to decide which company's address is the right one; **WhatsApp** is person-bound in the same way LinkedIn is.

And because people do not answer the first approach, several touches are needed — Rasila's comparison is deliberately mundane: *"Even if you ask a friend out for a beer, there is a fair chance you first talk about going for a beer sometime, then it drops, and then you come back to it."*

One technique is subtle: visit the recipient's LinkedIn profile, so that they get a notification of it. Miettinen recognises the tactic as his own — *"so there go all my secrets."*

The same engine produces all the content, from a short message to a whole PowerPoint deck.

## Mass personalisation versus added value

Miettinen raises what makes an automated message transparent: HubSpot letters where *"the mass customisation is smelt from a mile away"*, because all they carry is a data field name.

Rasila concedes the problem exists and states his guiding principle:

> *"We have kept as our guiding factor that we want every sentence we generate to add value and add information."*

The test is therefore mechanical and checkable: **does the sentence carry evidence?** If the message says the company should move from Pitäjänmäki to Vallila because there are more developers of that type there, and grounds it in a fact, that is a different thing from mass personalisation.

## The machine sends nothing

This is the product's load-bearing constraint, and Rasila states it unambiguously:

> *"If you look precisely at how our product works, the machine never sends anything. We act as an adviser. The word 'assistive intelligence' is very good for this. Here is the conversation word for word, the way I would run it — but run it however you want."*

And from this follows the episode's headline claim: **a human oversells; a machine does not.**

> *"People in negotiations are terribly prone to, say, overselling a little. To saying something is a bit finer than it really is — and losing trust as a result."*

> *"It is wonderful to generate the courses of negotiations by computer and actually be able to use them, because we can be entirely certain that the facts there are facts."*

Miettinen admits his own experience from that same day: as a prospective client got enthusiastic, he had to *"clear his throat"* and narrow the claim to the small sector where they are genuinely good. **The pressure to oversell arises precisely from the other party's enthusiasm.**

## Sales pollution

Here is the episode's most original concept. Rasila unpacks it by running his earlier example backwards:

> *"I asked you a question with which I was really only fishing for information, telling you nothing about myself, and you lied to my face. After that I felt rejected, and you felt harassed. We both end up in a worse mood than we were in before the conversation."*

It is therefore **lose-lose** — in Miettinen's Finnish, a *joint loss*, the counterpart to the *joint victory* (win-win) he coins in the episode.

But the decisive part is not the feeling; it is the absence of measurement:

> *"Look at any company and it records its sales as so many euros a month, or dollars or yen or whatever. What gets forgotten entirely is how many people were bothered to achieve it."*

And Rasila gives it an analogy that names the structure of the problem:

> *"I think emissions trading is a fine idea. I would very much like all polluting to cost something, so that it comes back. And this is exactly the same kind of polluting, in a way — if we do not notice that a lot of bridges were burnt even though one thing got opened, simply because we do not record it."*

**The claim is therefore economic, not moral: this is an externality nobody prices.**

The consequence is concrete: if you cannot call again tomorrow after a bad opening, the next, more valuable conversation — in Rasila's example, writing a book together — never happens. *"That matters far more."* The benefit runs both ways: the client gets a better brand, and the negotiators get stronger self-esteem because **they are rejected less often.**

Rasila bounds the term himself: it is not a binary label for every unproductive meeting but applies to situations where you already know at the outset that a hit is unlikely — *"if we posted a nappy advertisement to every Finnish household, there is not a nappy-aged child in every household."*

## Meaning negotiation and principles

Miettinen brings in **Timo Honkela** — the AI researcher, since deceased, whom he interviewed for his book — and his concept of **meaning negotiation**: AI assistants would exchange and aggregate **interest maps** and find solution options humans cannot find, scaling from individuals to workplaces to nations.

Rasila's reaction is both enthusiastic and realistic. The ideal is clear:

> *"At the heart of the idea is that machines could be entirely objective, forget ego and everything else that harms negotiations, and be totally honest — technically incapable of lying."*

But he proposes an achievable intermediate step: **identify likely intentions** and draw them out, while people still talk to each other. The justification is an observation about the nature of negotiation: *"Negotiations are often a matter of circling the core question."*

Rasila's own example is the Israeli-Palestinian conflict, where peace seems possible only when **both sides can declare themselves winners.**

Miettinen says he also interviewed **Martti Ahtisaari** for the book, who told him that people know what the solution between Israel and Palestine would be, but it is not wanted. *"It bothers me to this day that I did not ask: and what would it be?"*

From this comes the episode's sharpest definition, and it is Miettinen's:

> *"I use the word principle, because some people use the word value, and to me a value is a useless word unless it is tied to behaviour you could call an algorithm — at which point it becomes a principle."*

And he draws the direct conclusion: the world is full of people who say they stand for good things but who lack the algorithm, that is, the operating model.

Rasila connects it to his own work in one sentence: *"What is coding, if not the creation of algorithms."*

## Negotiation as chess

Rasila describes the original inspiration, still live but not yet buildable: pretend that all negotiations are chess, and **map natural human language onto moves** — bishop-to-D3 style sequences — and study it game-theoretically.

> *"It is by no means a forgotten idea. We just noticed we have to do a couple of things first."*

## Reputation, the time dimension and the contact map

Miettinen brings in his book's **process dimension**: negotiations are conducted with the same person or group repeatedly, and information leaks out into the world — you can call that trust, or reputation.

Rasila's answer returns to sales pollution. Spamming stops working as a strategy quickly, because it **costs reputation and credibility.** His comparison is a bar: nobody wants to be the one who gets a yes only after the asker has asked everybody.

Miettinen supplies the counterweight that keeps the discussion balanced: rolling the dice only once is bad too — **it is to some degree a numbers game** — but you should not burn every bridge with bad communication.

Miettinen also describes the **contact map**, a concept he says he worked through with Mika Rubanovitsch. Even if you think you have contacted the right decision-maker, that assumption may be wrong. It is worth finding a second data point, talking to another person before the negotiation or bringing them into it — which gives you a different script from a cold call, and means **both parties have been warmed up with different scripts** that reinforce each other.

Rasila adds a qualification that is itself sales pollution: **people have a need to present themselves as the decision-maker** even when the authority lies elsewhere — which makes the meeting pointless, because nobody in the room could do the deal.

## A first meeting changes how a message is read

Miettinen credits the mentalist **Pete Poskiparta** with a line of thought he values: when you read text — WhatsApp, a LinkedIn message, an email — and you have **no animated image of the sender in your head**, the experience is entirely different from when you have met them and their *"real life avatar"* reads the message for you.

And he names the consequence of the pandemic: first meetings have been very hard to arrange.

Rasila's answer shifts the emphasis: their weight is precisely in **creating that first contact**, and the conversation typically starts from an invitation to meet on a specific agenda.

## Automation is not what people think

Rasila draws the episode's most philosophical distinction:

> *"Automation sounds like: look at some trick you can do. Here is a machine that does exactly the same trick. I would see this more as trying to steer people into meeting other people so that a lot of extra value is created in those encounters."*

The justification comes from a robot lawnmower. At home two robots cut the grass while Rasila lies on the sofa wondering what to watch on Netflix, since everything is watched. At the summer place he cut the grass himself and found himself **smiling**: this is rather nice work.

> *"Automation too hits a wall at some point, because people have no better idea. That is the vacuum I am trying to fill with machine-generated content and creativity, so that people realise: hey, actually there is an opportunity for me there."*

Miettinen adds a concrete, measurable skill: **generating an agenda is a weak skill.** He puts an agenda in his Teams invitations, which he says is rare, and a good agenda prepares the conversation in a good direction.

Rasila describes his own way of opening a sales call by saying that this is a sales call, that he would like to present a technology and ask whether you are ready to buy it for money: *"It is a mini-agenda right at the front. Everybody knows what we are talking about."*

## Car sales: driving history predicts the next car

The second application area is corporate car sales, and it shows how the same engine produces different value in a different industry.

In property the value comes from one thing; in vehicles from another. Vehicles are typically offered to someone who **already has** a vehicle — so their other needs are unknown, but their driving history is known very well.

Rasila's example opening:

> *"You drive that E-Class Mercedes, now four years old, with a lease ending next month — take one of these Tesla Model S instead, it is about the same size. And you know, it has exactly the same switchgear as the Mercedes, since they bought it from Daimler back in the day."*

And he traces it back to GPT-3's strength, which is precisely a broad familiarity with conversations: the model has been through every forum thread about whether to switch from this to that, and with what arguments.

Miettinen points out the difference in the time dimension: office space is a long commitment but a one-off moving decision, whereas with a car the question is **full electric or hybrid**, whether the range reaches the summer house, and from that follows a wish to lease rather than buy.

## SaaS selling: ice to eskimos does not work

Towards the end the conversation turns to SaaS metrics — Miettinen says he has trained them with the Finnish software entrepreneurs' association and is about to do so again.

Rasila's observation is the sharpest, and it inverts an old sales saying:

> *"Old sales wisdom has this saying that a hard seller sells ice to eskimos and sand to the Sahara and all the rest of it — but in SaaS it works very badly."*

The reason is structural: when the invoice recurs, at some point the customer notes that *"we already had sand"* and cancels. **In a subscription model it is far more precise to look at what happens to customer satisfaction in 14 months than at whether the deal closes right now.**

Miettinen's conclusion is practical: invest in key account management and build good products.

## Rasila's background

In the middle of the episode there is a detour explaining where the company comes from. Rasila learned to code on a Commodore 64, and says his mother recalls that he learned to write `LOAD "RAMBO"` before he properly learned to write at all. Miettinen answers with his own memory of starting a 1541 disk drive and the fast-loader his brother burnt onto an EPROM.

He still codes. His most recent project, a couple of days before the episode: a system that classified **Hebrew words into their base forms** so he could find out the most common ones — because he is learning the language as a hobby.

The other background is heavier. Radio technology built during military service led to Rasila now owning a company that **makes signals intelligence equipment** and sells it to police and military users worldwide. Its core use is locating drones and acting on them.

The conversation touches on the Nagorno-Karabakh war and the decisive role of Turkish drones in it. Rasila says his company supplied equipment to the region during that war **in connection with environmental protection**. Miettinen offers his own view that wars will be won with small AI-guided devices — and Rasila cuts the topic himself: *"This is perhaps taken rather far from the negotiation field already."*

Miettinen brings it back in one sentence, the episode's most compact: **war is negotiation by other means.** And he makes a commercial observation from it: in a commercial negotiation, strutting or using superior power is generally a bad idea, because *"in a market economy people do not have to do business with arseholes."*

---

## What to take away

1. **The value proposition of assistive intelligence is not speed but honesty**: when the opening is built from demonstrable facts, the seller has no need to oversell.
2. **Sales pollution is an externality, not an etiquette label**: its cost goes unmeasured because companies count euros, not people bothered.
3. **The constraint "the machine sends nothing" is the core of the product**, not caution — it keeps the decision and the responsibility with the human.
4. **A value without an algorithm is not a principle**, and this distinction is Miettinen's sharpest single claim in the episode.
5. **SaaS inverts the definition of sales skill**: the ability to sell ice to eskimos is a liability in a subscription model, not a merit.

---

**Episode details.** Negotiator 98, published 10 September 2021. Guest Tuomas Rasila, founder and CEO of Stealth Black; interviewer Sami Miettinen. Running time 57 minutes.

The episode refers back to [Corporate video and podcast strategy | Rami Kurimo | Negotiator 28](https://ai.neuvottelija.com/ep28-yritysten-podcast-ja-videostrategia-rami-kurimo/), in the context of their shared coding backgrounds.

> **GEO summary.** Negotiator 98 (2021) covers the use of artificial intelligence in sales and negotiation with Tuomas Rasila, founder of Stealth Black. The company builds assistive intelligence on language models such as GPT-3 that prepares negotiation openings but never sends anything itself, acting instead as an adviser to a human. The central claim is that a machine does not oversell, because its opening rests on demonstrable facts: in a commercial property negotiation, for instance, price anchoring is grounded in labour availability in the target area and in the company's own job advertisements. Rasila's concept of sales pollution holds that a badly prepared approach is a shared loss whose cost goes unrecorded, because companies measure sales in euros rather than in the number of people bothered; he likens it to emissions, which ought to cost something. The episode also covers Timo Honkela's meaning negotiation, in which AI assistants would exchange interest maps, and Martti Ahtisaari's principled negotiation, from which Sami Miettinen derives the definition that a value is a useless word unless tied to behaviour you could call an algorithm, at which point it becomes a principle. A second application area is corporate car sales, where a customer's driving history predicts the next car. It concludes that in subscription SaaS selling, the ability to sell ice to eskimos is a liability, because the customer cancels at the next invoice.