Cosa Foundations
Six levels. Each builds on the ones below it.
How to build a successful AI-native GTM system
If you want to understand how you can grow your sales system and power-scale it with AI, asking yourself how to implement AI is the wrong question. The right one is: where are you currently blocked in your system, and why? Otherwise you will proceed to fail, with a higher token bill.
This means having a deep look into the constraints of your current system, and then progressively building higher and more ambitious AI implementations on top of them.
The six levels
Each level builds on the one below it:
LEVEL 1
Business Case
This means you address an URGENT pain, with a VALUE they want, with the TRUST that you can deliver it, in the FORMAT they can receive it through. If you cannot state any of these clearly, you are not ready to do sales for it.
The pre-qualification metrics: ICP <> Solution fit, sales cycle fit, a plausible ticket size floor, a TAM that is large enough, and decision maker access, including via which channels.
LEVEL 2
Growth Foundations
The constraint analysis will show where it breaks, but these are: lead engine and tech infrastructure correctly set up, funnel metrics assessed, sales collateral and processes in place, etc.
LEVEL 3
Unit Economics
Working processes for lead generation, research and conversion; looking at unit economics, CAC and offer.
LEVEL 4
Lead Engine Optimization
LEVEL 5
Sales Process Optimization
LEVEL 6
Hiring & Development Funnel Optimization
Do you have the right training, reporting and tracking systems in place to scale your team?
Levels 1 to 3 ask whether your system works at all. Levels 4 to 6 optimize three funnels: the lead engine, the sales process, and hiring and development.
Why the order holds
The order of the levels is not a rule someone made up. It is the order of dependency: what a level measures is shaped by everything it builds on.
Say you optimize the lead engine (Level 4) without proof that your business case can win (Level 1). The work is not only likely to fail. Its results are invalid either way. A bad Booking Rate may come from the business case, not from the engine: a false negative about the engine. A good one may be an accident that does not repeat: a false positive. You cannot tell whether the engine is set up correctly.
The same holds between any two levels, and through the whole stack: Level 4 depends on Level 3, and through it on Levels 2 and 1. It holds for AI as well. An agent on a level whose inputs are not proven scales the uncertainty, not the result.
The decision tree: what builds on what
And if anyone wants to even think about the agentification of their systems, they need to have proven the respective level first, before introducing its agentified version. The same ladder applies here as well, and it almost works like a decision tree:
You have not proven Level 1, Business Case?
No agent yet: Level 1 does not hold yet.
You have a clean business case?
Move on to Level 2, Growth Foundations.
You have the growth foundations of Level 2 in place?
Then it is worth cleaning up your data ontology and implementing a Company Brain.
You have proven Level 3, and your unit economics and business case are in order and understood?
Then it is worth identifying which channels, ICP tiers and offers should be agentified, and which deserve or require a human touch.
You have your lead generation in order and have reached your KPIs for the respective funnel steps of Level 4?
Then targeted and specialized agents for specific jobs of your lead engine are worth building.
You have the advanced sales processes of Level 5 in place and the respective KPIs met?
Then the respective specialized closing agents are worth building.
You have proven the hiring KPIs of Level 6 and onboard new hires in record time, and you have proven all agentic steps before as well?
Then holistic AI SDRs or AI AEs are worth building.
Read each line as a dependency, not as a permission. An agent runs the process of the level it sits on, faster and more often, but not better than that process has been proven to work. On a level whose inputs are not proven, nobody can tell whether the agent works. That is the reason behind “No agent yet”: it would only have assumptions to scale.
The dynamic
It is important to understand this dynamic if you want to be successful with your GTM systems, and before you start implementing automation and AI into them. Otherwise you will simply scale a flawed system.
This does not mean that processes for the other steps have to wait until the step before is proven. But it does mean that your priority in fixing problems and systems should be refocused around it.
Every system has its own architecture: what is proven in one does not carry over to another. Every company, and every solution within it, needs its own engine, its own foundations and its own level assessment. Not yet is not never: everything you hold back comes with the condition under which it makes sense again.
In depth, in the essay: How do we implement AI in sales? The wrong question.
How to use it
1
Read the six questions from Level 1 upwards. A yes counts when you can show it. A felt yes is an assumption, and nothing reliable rests on an assumption.
2
Find the lowest level you cannot prove. Every level above it builds on it, so their results cannot be read yet. That is what you work on first.
3
Prove that level, then read the one above it. Its results now mean something, and an agent on it has proven inputs to run on.
What does your system rest on? One minute.
Signal-based outbound: how to build a lead engine for hard-to-reach decision makers
Similar to the steps before, it goes on with building a proper lead engine end to end, which is Level 4 of the system above. A lead engine can technically live on its own within the overall system, even though there are considerations on the feasibility of it all.
It can stand on its own because its first question checks the business case for each campaign itself: the urgent pain, with VALUE, TRUST and FORMAT.
In this section, “level” in lower case counts the four questions below, not the six levels of the system.
Especially when it comes to larger and older companies with an existing system, starting with a full system overhaul is often too risky in comparison, which is ok. This is why we have developed this plug-and-play blueprint for a signal-based lead engine.
The logic and assessment are still similar to what was shown above, just more focused on outbound lead generation. And similar to above, each question builds on the one before. However, you can use this logic and apply it to every campaign you want to run.
The elements that make a successful campaign are always these 3: VALUE, TRUST and FORMAT. They are the basics you need to start with when working with a new ICP, solution or channel.
The four questions
QUESTION 1
Do you have an urgent pain for your ICP that you can solve?
Meaning, do you have the confirmation that this specific ICP and decision maker is confronted with this business-critical problem, is your solution the right VALUE, do they TRUST that you can provide it, and can they receive it through the FORMAT you are offering it in? This is usually measured through existing customers who already bought this. If it is a newer approach, it requires interviews and specific prospect commitments that validate the ICP <> Solution fit.
QUESTION 2
Is your lead engine set up in a way that it does what it is supposed to do?
This starts with documenting the proper ICP knowledge and identifying the right signals. It then proceeds with setting up the required tools for the intended task, with the right information implemented, the sending infrastructure warmed up, and everything integrated into your working processes. This is a technical assessment and will be assessed as such.
QUESTION 3
Do you have your ICP and pain point validated?
KPIs: Booking Rate 1 to 3 % (alternatively Response Rate 10 to 30 %).
QUESTION 4
Do you have your meeting booking and qualification engine validated?
KPIs: Disco to Demo conversion above 50 %, and Show Rate above 90 %.
Agents on the lead engine
Any kind of automation or agentification of each of these steps requires that the respective question of your lead engine has been proven first. If your lead engine has not properly proven question 2 (data ontology and correct integration of tech), putting an AI caller (which requires question 3) on top of it is doomed to fail.
If you have proven question 2 but are still in the process of reaching your question 3 KPIs, you have not truly understood yet why your prospects are getting into calls with you. Meaning, your agent’s ceiling is your current performance. However, it will not be properly equipped to differentiate between the validated and the not-validated assumptions of what worked and what didn’t. It will likely burn more of your contacts and markets than it converts (in the example of AI callers).
In depth, in the essay: Build a signal-based lead engine: the four questions
Closing complex B2B deals: how to build a sales process that makes you ready to scale
The lead engine ends when a qualified decision maker sits in a first call with you. The sales process starts right there and ends with the signed contract, which is Level 5 of the system above. Like the lead engine, it can live on its own within the overall system, and you can apply its logic to every single deal you run.
And similar to above, asking yourself how to close faster is the wrong question. The right one is: which buy-in is still missing, and from whom? A purchase in a company is never one yes. Every person who has a say in the deal has to give seven buy-ins, one after the other. If one of them is missing for one person, the deal stalls, even if everything else is right.
The buy-ins are collected per person, not per opportunity. Every new stakeholder starts at level 1, whatever the others have already given. Inside the sales process the buy-ins are counted as levels 1 to 7 per person: Urgency is level 1, Qualification level 2, up to Pricing at level 7. In this section, “level” in lower case counts one person’s seven buy-ins, not the six levels of the system.
The elements that make a buy-in are always these 3: the QUESTION the buyer has to answer for themselves, the PROOF you bring so that they can answer it, and the CONFIRMATION, which means the buyer has said it in their own words. If it only feels like they agree, you have an assumption and not a buy-in, and every forecast you build on it is a guess.
confirmed
assumed
open
Seven buy-ins, one after the other, collected per person.
The model maps the buyer’s decision, not the seller’s process.
The seven buy-ins
BUY-IN 1
Urgency
Do you have the Urgency buy-in? The buyer’s question: “Is this change my priority right now?” This one is decided before the call, through your outreach, your website and your pre-qualification. If the call came to you another way, take the person at face value and requalify them in level 2. In our experience this is where most deals die, and not with a no, but with a “not now”.
BUY-IN 2
Qualification
Do you have the Qualification buy-in? “Do you understand my problem better than I do?” You need their current state, their desired state and the problem in between, in their words and with their numbers, and the confirmation that you are talking to the decision maker. Ask for it out loud: “Did I capture your problem the way you would describe it?”
BUY-IN 3
Thesis
Do you have the Thesis buy-in? “Does your approach get me to my goal faster than what I do today?” You need your unique insight, your advantage over their current way, and benchmarks and KPIs that show it. The two objections that live here are “we will do it ourselves” and “we already have something similar”.
BUY-IN 4
Methodology
Do you have the Methodology buy-in? “Does this process make sense to me?” You need the new way of working step by step, and how it fits into the systems and tools they already run. If they cannot explain it to a colleague afterwards, you do not have it.
BUY-IN 5
Onboarding
Do you have the Onboarding buy-in? “Does my team have the resources to pull this off?” You need the timeline for the next six months, what you need from their side, the milestones and KPIs, and who takes care of them along the way.
BUY-IN 6
Feature
Do you have the Feature buy-in? “Are these features and services enough to get me where I need to go?” You need the full list of features and services, the service levels, and a no-risk guarantee as a mutual commitment.
BUY-IN 7
Pricing
Do you have the Pricing buy-in? “Is this price fair for what you have promised me?” You need the pricing structure, alternative packages, the cost of the alternatives (including the cost of carrying on as before), and incentive-based pricing. With large tickets, sell a first paid delivery (a workshop, a POC, a trial) before the full contract. If the six buy-ins before are in place, pricing is rarely about the number.
Around the buy-ins
Two things sit around these levels. TRUST (“Do I trust you to help me?”) is not a level, it is earned at every single one of them, through references and case studies.
And after the seventh buy-in comes the internal alignment (“Do I have the go from my own people?”). That is the reason why every buy-in has to be collected per person and not per company, and why your buyer needs documents they can pass on internally.
Reading a stuck deal
Similar to above, each buy-in builds on the one before. Your buyer will jump between them in every conversation, and that is ok, but a later buy-in does not hold while an earlier one is still open. This is also how you read a stuck deal.
The prospect has stopped replying? Usually the Urgency of one person is missing, and often not the one you are talking to.
“We need to align internally”? Onboarding or Feature is still open for someone.
“How much does it cost?” in the first minutes? Thesis or Methodology has not landed yet, and no price will be fair until it has.
And if a decision maker has not given you a single buy-in, that empty row is a finding as well: either the person is missing the question, or your deal is missing the person.
Examples from practice, not measurements.
Sales Process fit
Do you have Sales Process fit, meaning, are you ready to scale? KPIs: Disco to Closed Won 10 to 30 %, and (LTV/CAC) / sales cycle in months above 0.25, with an LTV/CAC of at least 3.
The second number puts your unit economics in relation to your time: an LTV/CAC of 3 with a sales cycle of 12 months sits exactly on the line. The same ratio closed in 4 months is three times as good. The floor of 3 is there so that a short sales cycle cannot hide weak unit economics. This is why shortening your sales cycle is not a nice-to-have, it is half of the equation.
If you reach these numbers repeatedly, and not through the one person who can close everything, your process is ready for more people and more leads. If you do not, more leads will only give you more stuck deals.
Agents in the sales process
Any kind of automation or agentification of each of these steps requires that the respective buy-in has been proven first, by a human, in your own deals. A call prep agent can only prepare the Qualification you already know how to win. A follow-up agent can only repeat the Thesis and Methodology your buyers have already confirmed to you in person. An offer agent can only send out an Onboarding plan, a Feature list and a Pricing that have closed before.
If you have not reached your Sales Process fit KPIs yet, you have not truly understood why your customers are buying from you, and at which buy-in the others dropped out. Meaning, your agent’s ceiling is your current close rate. However, it will not be properly equipped to differentiate between a confirmed buy-in and an assumed one. It will send your offer to a buying committee that has not bought your thesis yet. You will have automated “we need to align internally” and the discount talk that comes after it (in the example of closing agents).
In depth, in the essay: How do we close faster? The wrong question.
What this rests on
Cosa is young, and there is no study of its effect. That is why we say, for each statement, what it rests on.
The model of the buying decision.
Basis: working model from practice
It comes from eight years of building sales, not from a study. It is a model, not a law of nature.
The checking approach for knowledge.
Basis: research and code
Two published papers shaped it (see below). Cosa checks mechanically against the text that a quote really appears in the conversation.
The process: human first, then agent.
Basis: practice
It comes from venture building and from our own sales building. It is not proven; it has worked for us.
What we do not promise: no guarantee on meetings or closings, only benchmarks. No replacement for your team. No checking of what was never said: what is missing stays open.
Knowledge keeps its source
The Company Brain stores what buyers have said, not what a model guesses.
A buy-in counts as confirmed when the buyer’s sentence really appears in the conversation. Cosa checks mechanically that each quote really appears in the conversation text. Without a sentence the row stays assumed or open, and an assumption never appears as a fact. A correction from you takes precedence over any later automated read.
The answer comes back as an email and corrects the knowledge.
Human first, then agent
An agent gets only work that a human has already done at your company.
GTM Assessment. Clarify processes, data, systems and goal. The result is your GTM roadmap.
Implementation & testing. Your Forward Deployed GTM Engineer does the work first. That becomes agents, which go live only once they reach the agreed benchmarks.
Handover and upkeep. The system runs in your environment and we maintain it. Your growth then no longer hangs on a single person.
Every line in the decision tree above follows the same dependency: prove it first, then hand it over.
What you hold afterwards
Four things you can check without taking our word for it.
A GTM roadmap for your processes, data and systems.
The state of every deal, per person and per buy-in, with sentence, speaker and date, once your Company Brain is set up.
Benchmarks every agent has to be measured against.
Cosa’s planning benchmarks: Booking Rate 1 to 3 % per contacted account (alternatively Response Rate 10 to 30 %), Disco to Closed Won 10 to 30 %; (LTV/CAC) / sales cycle in months above 0.25 with an LTV/CAC of at least 3. They are set with you at the start of Implementation & testing. Benchmarks, not a guarantee.
An approval before every send.
What the research says
Two papers shaped our checking approach. We do not implement them one to one and say where we differ.
Proof-Carrying Numbers
SOLATORIO, 2025
Every number in an AI answer is mechanically checked against a stored statement before it is shown; only what has been checked counts as verified. We use the principle of checking statements by their kind and, when in doubt, not confirming. We differ: the paper is designed for database queries, not for conversation transcripts.
VeriTrail
Statements from a multi-step AI process are traced backwards to the source text. We use the idea of keeping the path of a statement to its source traceable. We differ: we use the principle, not the procedure.
Where the Foundations come from
The Foundations grew out of venture building and eight years of building sales, most recently as interim head of sales. The Cosa prototype was used in every sales conversation there before it became a company. Cosa has existed since April 2026. Who is behind it and which decisions led here is on the page The company behind Cosa.
Who this is for, and who it is not for
This is for you if
you sell a B2B offer that cannot be explained in one sentence,
several people at your company decide on the purchase,
your sales so far depends on you or on one person,
your company has roughly 5 to 200 employees.
This is not for you if
your product explains itself and is bought without a conversation,
you are looking for a guarantee on closings,
you want a tool that replaces your team,
nobody at your company is ready to open up conversations and data.
Frequently asked questions
The architecture of an AI-native GTM system: six levels that build on each other. Two of these levels can stand on their own within the overall system: the signal-based lead engine (Level 4) and the sales process (Level 5). A level's results only mean something once the levels below it are proven.
Next
Cosa Academy: Build, Sell, Scale
The system, the lead engine and the sales process, explained in full.
Your intelligence: the Company Brain
What sourced knowledge looks like.
The company behind Cosa
Who developed the Foundations, and where they come from.
What working together looks like
From GTM Assessment to handover.