What It Really Takes to Build a $100M AI Agency : Nate Herk Playbook
The first automation Nate Herk sells in his business-building narrative is rarely presented as a technological breakthrough. It is more likely to qualify leads, update a customer relationship management system, produce a report or remove several hours of repetitive administrative work. That apparent lack of glamour is central to the model.
Businesses may be fascinated by autonomous agents, but they generally pay for shorter response times, lower operating costs and fewer employees copying information between systems. An AI agency becomes commercially useful when it translates rapidly changing technology into those measurable outcomes.
Herk, a Chicago-based automation entrepreneur and online educator, says he started an AI automation agency in September 2024 and reached $100,000 in monthly recurring revenue nine months later before selling his interest to his business partners. He has since converted the experience into courses, online communities and a playbook for prospective agency owners.
The figures have not been independently audited, and the value of the sale has not been publicly disclosed. His model should therefore be examined as a founder’s account rather than a verified case study of agency performance.
Even with that qualification, the playbook reflects an important change in the AI services market. Low-code platforms and commercially available language models have reduced the cost of building prototypes. The difficult work has moved elsewhere: identifying a process worth changing, integrating systems safely, proving a return and supporting the automation after it reaches production.
That is where an agency can create value. It is also where many new entrants underestimate the work.
Sell the operational result, not the AI
Herk’s most commercially credible principle is that an agency should sell an outcome rather than a tool.
A customer does not normally need an “n8n workflow”, an “AI agent” or a connection to a particular language model for its own sake. It may need to respond to every sales enquiry within five minutes, reduce the time required to prepare proposals or prevent customer-service requests from disappearing between email and a ticketing system.
This distinction determines whether the engagement produces a durable business relationship or a demonstration that looks impressive but remains peripheral to operations.
An agency assessing a customer-service process, for example, should calculate the present cost of receiving, classifying and routing enquiries. It should establish how many are repetitive, how often staff correct errors and what a delayed response costs the company. Only then can it determine whether an automated system is worth building.
The technological architecture follows from the problem. Some tasks may require a language model because the inputs are unstructured or variable. Others are better handled through conventional rules, database queries and application programming interfaces. Adding generative AI where deterministic software would work introduces cost and unpredictability without necessarily improving the result.
This is particularly relevant as the market moves from experimentation to financial scrutiny. AI adoption is widespread, but the proportion of projects producing the expected return remains much lower. The agency that can identify when AI is unnecessary may be more valuable than one that attempts to place an autonomous agent in every workflow.
Begin with one process that can be measured
The broad promise to “automate a business” is difficult to sell and dangerous to deliver. Herk’s more practical content begins with individual workflows that have an identifiable owner and quantifiable cost.
Lead qualification is a common starting point. A system can collect enquiries, enrich records, classify prospects and prepare an initial response before assigning the opportunity to a salesperson. The result can be measured through response time, qualified meetings, conversion and hours saved.
Other suitable processes include extracting information from documents, preparing recurring reports, routing support tickets, updating customer records or drafting standard communications for human approval.
These tasks share several qualities. They occur frequently, follow a recognisable structure and produce an output that can be checked. Errors are usually detectable before they create severe financial or legal consequences.
A first engagement should avoid processes in which a model can transfer money, make binding commitments or communicate unsupervised on sensitive matters. High-risk automation can be introduced later, once the client has developed monitoring, approval and escalation procedures.
The objective of an early project is not to display the maximum autonomy the technology can achieve. It is to establish that one process can be improved reliably enough for the client to expand the relationship.
The automation audit is a product in its own right
One of the stronger elements in Herk’s progression from freelancer to consultant is the use of a paid automation audit.
Most companies know that some of their work is inefficient, but they do not necessarily know which processes should be automated first. Interviews frequently produce wish lists shaped by the loudest department or the latest AI demonstration rather than by commercial impact.
A proper audit maps how work currently moves through the organisation. It identifies the systems involved, the people making decisions, the data required and the exceptions that interrupt the standard process.
The agency can then rank opportunities according to volume, labour cost, technical feasibility, risk and expected return. A repetitive task performed by ten employees may appear attractive, but not if the input data are inconsistent or every fifth case requires subjective judgement.
Charging for this analysis changes the relationship. The agency is paid for diagnosing the problem rather than depending entirely on a later implementation contract. The client receives a usable roadmap even if it selects another provider for part of the work.
It also filters prospects. A company unwilling to provide data, name a process owner or define a success metric is unlikely to support a successful production deployment.
Recurring revenue requires recurring responsibility
Herk argues for retainers rather than relying entirely on one-off project fees. The economic logic is clear: monthly contracts create more predictable revenue and increase the value of the agency.
The client needs an equally strong reason to keep paying.
An automation is not a static website that can be handed over and forgotten. External applications change their interfaces, language-model providers update their products and client employees alter the processes around the system. A workflow that operated reliably at launch may fail after a vendor changes an authentication method or the client modifies a field in its CRM.
A defensible retainer can include monitoring, incident response, model evaluation, cost control, security updates and the development of additional workflows. The agency should report uptime, errors, usage and realised savings rather than sending an invoice for undefined “AI support”.
This service model is more demanding than its recurring revenue suggests. It creates obligations outside normal project delivery and requires the agency to maintain knowledge of each client’s architecture. A small firm accumulating customised systems across many customers can quickly develop an unmanageable support burden.
Standardisation is therefore essential. Agencies need reusable logging, approval, testing and deployment components even when each client’s business process differs. The more every implementation is built from scratch, the less attractive the recurring economics become.
No-code does not mean no engineering
Herk’s educational material frequently demonstrates systems built through platforms such as n8n, which allow applications, databases and AI models to be connected through visual workflows.
These tools have lowered the barrier to entry. A consultant can produce a working prototype without developing an entire software application and can show a prospective client a result relatively quickly.
The ease of building a demonstration can disguise the difference between a prototype and production infrastructure.
A reliable workflow needs authentication, permissions, logging, error handling and a method for recovering when one service becomes unavailable. Data must be stored and transmitted appropriately. Model outputs may require validation before another system acts on them.
The agency must also determine who owns the workflow, accounts and credentials. A customer should not discover at the end of an engagement that essential infrastructure is tied to the founder’s personal account or that the agency alone can maintain it.
No-code and low-code platforms reduce the amount of custom programming required. They do not remove the need for systems architecture, testing or security discipline.
This creates a natural division within the agency model. The founder may lead discovery, sales and client strategy, while experienced automation engineers handle implementation and quality assurance. Trying to preserve the image of a one-person agency while quietly depending on unstructured subcontracting can create delivery and confidentiality risks.
Case studies must withstand financial scrutiny
Herk’s growth model relies heavily on content: build an automation, explain it publicly and allow the demonstration to attract prospective clients. This is an effective acquisition strategy because it turns technical work into evidence of capability.
A credible case study needs more than the time taken to build the workflow and the fee charged.
The client should be able to establish the previous cost, the implementation expense and the ongoing operating cost. Savings need to account for human review, software subscriptions, model usage and maintenance. Revenue gains should be separated from opportunities merely processed by the system.
A claim that an automation “saved 20 hours a week” is useful only when the business explains what happened to those hours. If employees simply redirected them towards other low-value work, the financial return is different from eliminating outsourced expenditure or allowing the company to grow without hiring.
This is also where the agency must avoid using theoretical wage savings as though they were realised cash. Saving half of one employee’s time does not reduce payroll unless the company can redeploy that capacity productively.
The best case studies connect process improvements with a financial mechanism: more qualified leads contacted, fewer customer cancellations, lower contractor expenditure or shorter invoice-collection cycles.
Content can become a second business
Herk’s own trajectory illustrates a feature of the modern AI-agency market: education and audience building may become as commercially important as client services.
His website, YouTube channel and online communities teach users to build and sell automations. The audience then supports paid memberships, training products, sponsorships and consulting leads.
This is not necessarily a weakness in the model, but it complicates the evidence. Revenue attributed broadly to “AI” may come from agency retainers, media, education, affiliate partnerships or advisory work. Prospective agency founders should not assume that client delivery alone produced every publicly described result.
Content distribution can nevertheless be a powerful competitive advantage. AI services are difficult for buyers to evaluate, and trust often develops before a formal sales conversation. A consultant who explains real workflows, limitations and implementation decisions can demonstrate expertise at scale.
The risk is that content incentives reward speed and novelty rather than long-term reliability. A video can show an agent working once under controlled conditions. A client needs it to continue working when data are incomplete, a supplier changes its software or the model produces an unexpected response.
The agency’s delivery standards must therefore be more conservative than its demonstrations.
The $100 million problem
The headline claim that Herk built a $100 million AI agency cannot be supported by the available evidence. His own account is that the business reached $100,000 in monthly recurring revenue, equivalent to $1.2 million annually if maintained for a full year.
Even this figure requires context. Monthly recurring revenue is not profit, enterprise value or cash received from a sale. The agency would incur delivery, software, contractor, sales and support costs. Revenue may also be concentrated among a small number of clients whose departure would materially affect the business.
A services company reaching that run rate in nine months would still represent notable commercial progress. Exaggerating it into a $100 million business makes the underlying achievement less credible rather than more impressive.
The more transferable insight is not the final number. It is the progression from small implementations to paid diagnosis, larger projects, recurring partnerships and eventually an audience-based education business.
Each stage solves a different constraint. Freelance projects create evidence. Consulting establishes strategic credibility. Retainers improve revenue visibility. Content lowers customer-acquisition costs and opens additional sources of income.
Skipping those stages can leave an agency selling enterprise transformation without the delivery record, systems or team required to provide it.
What businesses should buy from an AI agency
A client should begin by asking for a defined process improvement rather than a broad AI strategy presentation.
The provider should be able to map the existing workflow, quantify the opportunity and explain why the proposed architecture uses AI where it does. It should identify the data required, the decisions that remain with people and the consequences of an incorrect output.
The commercial terms should distinguish implementation from ongoing support. Ownership, hosting, credentials and exit arrangements need to be agreed before work begins. The client should also have access to monitoring data and be able to suspend the automation without depending entirely on the agency.
Most importantly, success should be tied to a business metric established before the system is built.
Herk’s playbook is persuasive where it treats AI automation as operational consulting rather than a catalogue of impressive agents. It is less useful when founder claims are repeated without financial context or when rapid prototyping is mistaken for durable implementation.
The market does not need more agencies promising to automate everything. It needs providers prepared to understand one process deeply enough to improve it, measure the result and remain accountable after the demonstration is over.
