Finding Your Ideal Customer: The AI-Driven Approach

This week’s blog explores a practical approach to identifying your Ideal Customer Profile (ICP) for targeted prospecting. I will also build on last week’s blog where I argued that the real power in AI lies not with the companies developing Large Language Models (LLMs) but with domain-specific data enterprises. A striking example of this shift was with the launch of the Chinese hedge fund High-Flyer DeepSeek’s R1 model, reportedly trained in just two months at a cost of $6m. While this figure is conceivable given its reliance on a smaller yet highly curated dataset and reinforcement learning—an approach we successfully leveraged in our hedge fund—the actual cost is likely much higher when accounting for infrastructure and expertise. DeepSeek is said to have amassed 50,000 Nvidia GPUs over an eight year period and at a cost of ~$1,000 each, converting the true expenditure would be well beyond $50m—still a fraction of OpenAI’s enormous costs!

The Commoditization of AI & the Rise of Open-Source Models

Could nascent transformer models, including OpenAI’s ChatGPT and Anthropic’s Claude ai, be on a path to commoditization due to the advent of open-source models like Meta’s Llama 3 and DeepSeek’s R1? Particularly as closed-source models require vast amounts of capital to maintain, whereas open-source iterations benefit significantly from community-driven/ ecosystem advancements.

Take ARM Holdings a specialist chip developer, who exemplified how ecosystem-driven collaboration can foster industry dominance. ARM’s partners with over 1,000 companies and its ability to integrate with specialized industry players enabled it to anticipate future technological needs, leading to industry dominance with a 99% penetration rate in mobile phones. By contrast, the forever insular Intel, once dominated the semiconductor space, becoming the most valuable company in the world at $509b in 2000. But Intel currently languishes at a $87b market cap and its gross margins sit at a pathetic 15%. Compared this to ARM whose market cap was $550m (yes m NOT b) in 2000 and now stands at $161b and delivers gross margins of 95%! A true testament to the disruptive power of collaborative ecosystems. Beware; partnerships can also pose challenges, as illustrated by the Chinese proverb 同床异梦 ("Sleeping in the same bed but dreaming different dreams").

My Experience with GPUs & Investing in AI

DeepSeek’s origins as a quant hedge fund strongly resonate with me. In 2015, after repeatedly frying CPUs, we sought a better solution. Upon recommendation from our local CPU shop, we switched to Nvidia GPUs—a move that was transformative. What once took 3-4 days to process was now being completed overnight. The landscape evolved further in 2018 with the introduction of Nvidia’s GeForce RTX series and the advent of parallel processing across multiple GPUs.

I have long adhered to Peter Lynch’s investment philosophy: "Invest in what you know." Peter believed that everyday investors could gain an edge by paying attention to products, services, and businesses they encounter in their daily lives and workplace! Recognizing Nvidia’s potential in 2018 and investing at approximately $5 per share would have paid massive dividends given it trades at $120 per share today. As a matter of full disclosure, I intend to participate in Anthropic’s current funding round. But how does this align with my advocacy for open-source AI? While LLMs may become commoditized, the companies behind them can still become highly profitable due to the differentiation in training parameters, processing speed, features (text, images, video, voice, web search), context windows, proprietary training datasets, but perhaps most importantly due to the vast amounts of capital they have raised; allowing them to blitzscale.

?? Recommended Reading: Blitzscaling by Reid Hoffman (The Lightning-Fast Path to Building Massively Valuable Companies).

The impact of AI assistants on organizational productivity is profound. $20 a month is peanuts, in fact at a cost of $500 a month it is still absurb! So what’s my willingness to pay? Is it $2k a month, is it $5k a month? Probably. What’s the size of the pie when two billion or four billion people come to the same realization? What are the margins then! Hmmm let me plug this into my NPV and future cashflow calculation - thousands of percent of growth, with continual margin expansion! Feels like I was just teleported back to 2000 and given a chance to buy Google again! Add this to the fact that I am absolutely enamoured with Claude.ai and its UI (i.e. Peter Lynch's investment criteria) and you get BINGO!

Footnote1: SoftBank’s rumoured $25b investment in OpenAI is not merely political—it is strategic and it is to make a substantial return on their investment! Softbank are no dummies!

Footnote2: SoftBank acquired ARM for $33b in 2016, relisted it in 2023, and today ARM boasts a market capitalization of $161b. NB: Softbank still owns 90% of ARM.

Disclaimer: The views expressed are for informational purposes only and do not constitute financial, investment, or legal advice. They are my personal views and not intended to be investment advice as they may not suit your own investment criteria or goals.

AI-Driven ICP Analysis

Now on to the practical part of my blog. My approach leverages the complementary strengths of Claude 3.5 Sonnet and GPT-4. Claude 3.5 Sonnet, with its 180 billion parameters and training data through April 2024, brings unique capabilities derived from its Amazon-influenced dataset. GPT-4, trained on 220 billion parameters through June 2024, offers different advantages. While Claude's inability to perform real-time web searches might appear limiting, this constraint actually enhances its reliability by reducing the potential for hallucinations—those plausible but fabricated responses that plague AI systems.

The relationship between context windows and hallucination rates presents an interesting paradox. Initially, we found that narrower context windows reduced hallucinations by forcing more precise, focused inputs. For instance, when analysing market trends, providing annual reports from competitors would introduce noise that degraded model performance. However, our subsequent testing revealed that larger context windows, when properly utilized, can actually enhance accuracy by maintaining more comprehensive context throughout the analysis process (i.e. reduced the need to dump its memory and commence hallucinations). This is particularly relevant when comparing Claude's 200,000-token context window with GPT-4's 128,000-token capacity.

The keys lie with optimal prompt instructions and in strategic data presentation: Providing comprehensive but carefully curated information that maximizes the model's analytical capabilities while minimizing irrelevant inputs. This approach allows us to harness each model's strengths while mitigating their respective limitations. For prompting I follow three very simple rules:

1. Ensure the LLM takes on a persona.
2. Be clear and precise.
3. Maintain a professional and respectful tone. Treat it as an assistant, fostering a constructive exchange.

So, let’s start:

Step 1: Set up a new project in Claude (we also trained ChatGPT in a similar fashion – however I am only including Claudes training for this blog).

Note: To access Claude Projects you need the $20 a month version for this

Note: A new Claude project will only train on the data contained in this specific project, it will not train on any other data previously sent to train it in the main window or any other project.

Step 2: Give the project a name & describe what you are trying to achieve:

Name: Ideal Customer Profile

Purpose: Find my ideal customer profile to ensure my marketing efforts are more targeted and more successful. Please provide names of employers or companies I can target.

Step 3: Commence feeding the model

Commence with the following prompt:

You are a HR technology sales and marketing expert with 15 years of experience in developing ideal customer profiles.

Refer to this guide on building an ICP: https://fullfunnel.io/ideal-customer-profile/ ((Note as Claude does not have the ability to search the web, you’ll need to copy, screenshot or catalogue the website and paste)

Please also refer to:

Onesto Employer Presentation.

Onesto website (Note as again Claude does not have the ability to search the web, you’ll need to catalogue the website and paste for the most up to date data source)

Onesto Proprietary Dataset V14256 (Note: Preprocessing of your data is paramount in this step, ensure that it conforms with the SUM IMPACT framework remove all specific details about the client. In a future blog I’ll be discussing the data analytic techniques used to preprocess the data into an optimal and useable state.)

- Target companies with >1000 employees

- Target with operations in Australia

- Target company revenue of >$25m

- Exclude financial services companies

- Exclude all government workforces

Please create an ideal customer profile using the above & pasted resources. Please return in bullet point form the criteria for a Sales Development Representative to use when identifying potential target companies

AI-Driven ICP Analysis: Leveraging Data for Precision

Rather than defining my Ideal Customer Profile (ICP) manually, I tasked Claude to analyse my dataset to identify patterns and insights. Instead of imposing assumptions, I allowed the model to determine our ICP based on actual data.

Fortunately, we maintain a meticulously catalogued, clean, and up-to-date dataset spanning five years. This dataset contains key ICP components, including:

• Sales cycle length (both overall duration and engagement touchpoints such as emails, calls, and meetings)
• Organisations that align philosophically with our product—particularly those prioritising employee well-being
• User activation rates that correlate with successful launches and referrals
• Demographic insights such as employee location, income, workplace, spending patterns, age, and gender

Through our historical analysis we uncovered critical behavioural patterns:

• Employees paid weekly were less likely to engage with our solution, which correlated with large unionized workforces—thus, they are not a priority target.
• Employees earning over $125K per year or aged over 50 were less likely to adopt our service.

So we were eager to see if Claude was able to flesh these out.

Our structured data is securely stored in our database, fully anonymized, and seamlessly fed into our AI models via a CSV file direct from our proprietary dashboard. If your data is not yet formatted for AI training, don’t worry I will walk you through a step-by-step approach to structuring and optimizing data for effective preprocessing & model training in future blogs.

Output from Claude:

I'll help you create a comprehensive ICP framework for SDRs based on the provided information about Onesto. Here's a structured ICP targeting criteria:

Primary Qualification Criteria:

Company Size: >1000 employees

Geographic Focus: Australian operations

Annual Revenue: >$25M

Workforce Demographics: Predominantly aged 18-50 years

?

Industry Focus:

Service sector companies, particularly:

Retail chains

Hospitality groups

Quick service restaurants

Healthcare providers

Contract cleaning/facilities management

?

Behavioural Indicators (Strong Signals):

High employee turnover rates (>20% annually)

Multiple current job openings on job boards

Regular pattern of recruitment advertising

Lower average wage rates compared to industry standards

Fortnightly or monthly pay cycles

Recent initiatives around employee wellbeing/benefits

Mentions of employee retention challenges in company communications

?

Immediate Disqualification Criteria:

Financial services sector

Government organizations

Heavily unionized workforces

Currently using competing employee benefit solutions (implemented in last 6 months)

Operations primarily outside Australia

Companies in severe financial distress

?

Ideal Company Pain Points:

High recruitment costs (>$3500 per employee)

Struggling with employee retention

Experiencing wage inflation pressure

Limited employee benefits program

Manual/inefficient reward & recognition processes

Employee financial stress impacting productivity

High administration costs for employee benefits

?

My response:

Thanks. Can you please find a list of companies that adhere to this

?

Claude’s response:

Ideal Target Companies for Onesto

NB: All Companies names have been removed for this Blog)

Employer 1

Employees: ~200,000

Payment Cycle: Fortnightly

High proportion of 18-50 age demographic

Multiple locations across Australia

High turnover retail environment

Already in Onesto's partner network for Shop & Save

?

Employer 2

Employees: ~107,000

Multiple retail brands

Regular recruitment needs

Mix of full-time and casual workforce

Strong focus on employee benefits

?

Employer 3

Employees: ~100,000

Young workforce demographic

High turnover industry

Multiple locations

Regular hiring needs

Focus on employee benefits and development

Hospitality Groups

?

Employer 4

Employees: ~20,000

24/7 operation with shift workers

Multiple locations

High turnover industry

Mix of salary levels

?

Employer 5

Employees: ~15,000

Multiple venues nationwide

High proportion of casual staff

Regular recruitment needs

Healthcare & Aged Care

(NB: I've only listed 5 of the hundreds Claude uncovered)?

Claude

All meet the minimum size requirement (>1000 employees)

Australian operations with significant workforce

Service sector focus

History of investing in employee benefits

High proportion of employees in target age demographic

Regular recruitment needs indicating turnover

Non-unionized or partially unionized workforce

No competing solutions recently implemented

Fortnightly/monthly pay cycles

Multiple location operations

Demonstrated need for employee retention solutions

?

Prioritization Factors

Employee turnover rate

Current recruitment activity level

Existing benefits programs

Geographical spread

Workforce demographics

Payment cycles

Industry wage levels

Current economic pressures

?

AI-Driven Insights: Prioritization Factors Defined

Claude efficiently analysed our clean dataset and precisely identified key prioritization factors within minutes. The model provided invaluable insights, highlighting companies that:

• Have a proven history of investing in employee benefits
• Demonstrate a clear need for employee retention solutions
• Maintain a high proportion of employees within our target age demographic

Additionally, Claude effectively filtered out negative factors, including:

• Unionized workforces, recognizing that these groups predominantly follow weekly pay cycles. (Claude easily identified a preference for forthnightly & monthly paid employees).
• Companies that have recently implemented competing solutions, ensuring we focus only on viable prospects

Ensuring Data Security & Adherence to the SUM IMPACT Framework

Our results strictly adhered to the SUM IMPACT framework, with a strong emphasis on data security and privacy. To safeguard proprietary information, we implemented stringent measures:

• ChatGPT Privacy Controls: We disabled the "Improve the model for everyone" option in the profile settings under "Data Controls," ensuring that none of our data could be used to train the ChatGPT model in the future.
• Claude Projects Privacy Mode: Our model was trained exclusively in a secure environment within Claude Projects.
• Data Anonymization: All sensitive information—including names, addresses, and dates of birth—was removed before processing.

Protecting proprietary datasets is crucial, as exposing them to the public domain could provide competitors with strategic advantages. By maintaining strict security protocols, we ensure our data remains confidential and exclusive to our organization.

Conclusion

The strategic identification of an Ideal Customer Profile (ICP) is essential for targeted prospecting and business growth. By leveraging AI-driven analysis and the SUM IMPACT framework, we have refined our approach to ensure data-driven decision-making and precise targeting. Our exploration highlights the shift in AI power from model developers to domain-specific data enterprises, reinforcing the value of proprietary, well-structured datasets. Additionally, our commitment to data security and privacy ensures that sensitive information remains protected while maximizing analytical efficiency. As AI continues to evolve, organizations that embrace structured data methodologies and ecosystem-driven innovation will maintain a competitive edge.

In the next instalment, we will dive deeper into data preprocessing techniques to ensure optimal output, and I will detail the next steps in prospecting and marketing armed with our latest ICP target group.

#AI #DeepSeek #GPT4 #Nvida #ArmHoldings #Intel #Chatgpt #Claude #AILeadership #ArtificialIntelligence #MachineL#AI #DeepSeek #GPT4 #Nvida #ArmHoldings #Intel #Chatgpt #Claude #AILeadership #ArtificialIntelligence #MachineLearning #OpenSourceAI #IdealCustomerProfile #ICP #Softbankearning #OpenSourceAI #IdealCustomerProfile #ICP #Softbank