There is a sign and banner print shop near my home. Orders come in by phone, email, and from walk-ins. Every job has a different material type, size, and installation method. Design drafts go through multiple rounds of revisions. Approvals need to clear several stages. Production slots have to be squeezed into an already packed schedule. The owner has heard about AI, and knows software could be useful. But his biggest concern is: will a new system get the quotes wrong, mess up the production schedule, and force his employees to spend weeks learning something that turns out to be unfit for their needs?
There are over 40,000 sign and banner shops in the United States. IBISWorld puts the count at 41,805, and Mordor Intelligence estimates the market size at approximately $8.86 billion. But the market is highly fragmented: small and medium-sized shops range from $200,000 to $3 million in annual revenue. The group includes shops generating millions with a dozen employees, as well as a large number of micro workshops that do not have a single full-time IT person. Shops of different sizes vary significantly in their need for software and ability to pay, making it unrealistic to assume a uniform budget across the board.
This is the dilemma of general-purpose software. It assumes customers can describe their own processes, clean their data, configure the system, and evaluate the results — conditions that low-digitization small shops precisely lack. A Capterra survey cited by CompassApp found that 61% of US small businesses regretted a technology purchase in the past year or more, primarily due to costs exceeding expectations or incompatibility with existing systems. Avoma also observed in practice that even when a product offers a full self-serve experience, many small businesses still proactively request demos — because they need someone to help them confirm whether the solution is the right fit.
FDE stands for Forward Deployed Engineer, a role first systematized by Palantir. A traditional product engineer builds one feature to serve many customers; an FDE deploys multiple capabilities to serve one customer. The difference can be judged by a simple scenario: a pre-sales engineer’s job ends when the customer says “this should work”; an FDE’s job starts when the customer says “let’s make it run.” If code has entered the customer’s production environment and the engineer needs to handle issues at 2 AM, that is FDE work.
Former Palantir FDE Barry wrote in a personal reflection that customer deployments serve as testing grounds for new technologies — what proves effective gets migrated back to the product team. In a traditional SaaS company, a customer’s improvement suggestion has to pass through pre-sales to a product manager, get scheduled into the iteration plan, and might become a feature months later. An FDE hears it in the morning and can try it out the same day.
The core mechanism here is product discovery: the customer site continuously surfaces real needs. When similar needs recur across multiple customers, the product team can abstract them into reusable capabilities within the platform. This is also why Palantir can simultaneously achieve high per-customer revenue and platform reuse.
AI companies are now hiring for this role as well. OpenAI, Anthropic, Databricks, and ServiceNow all have FDE or similar positions open. Anthropic’s FDEs embed with strategic customers, understand their workflows, and develop AI applications that solve real problems. OpenAI’s Forward Deployed Software Engineers design custom software alongside customers and write best practices back into the internal knowledge base.
The premise holds: low-digitization businesses need services. But a critical half is missing: the FDE model as practiced by Palantir and OpenAI cannot cost-effectively reach small shops.
Colin Jarvis, head of OpenAI’s FDE team, mentioned in an interview that their projects range in value from tens of millions to the low billions of dollars. This figure comes from interview reporting and has not been independently verified. The compensation for FDEs themselves is also high: public job postings show annual total compensation for AI FDEs ranging from $135,000 to $600,000. This cost is acceptable when serving a client like Morgan Stanley. For a print shop with annual profits in the hundreds of thousands, it simply does not pencil out.
Two companies are worth examining here. ServiceTitan builds software for the trades — putting dispatch, quoting, scheduling, payments, and customer management for plumbers, HVAC, and electrical contractors into a single system. ScaleFactor automated bookkeeping for small businesses and once marketed itself as AI-powered accounting software. One went public; the other shut down. The difference lies precisely in whether service can be productized.
ServiceTitan’s significance is that it proves traditional industries can accept a hands-on rollout: someone helps the customer map processes, configure the system, train employees, and get the software into daily operations. But this only works under specific conditions: the customer has to be large enough, and the product has to cover enough of the business. According to its S-1 filing, gross dollar retention exceeded 95% and net dollar retention exceeded 110% for the past ten quarters (note: these are quarterly figures as used in the S-1; SaaS Capital has noted that a quarterly 95% figure may correspond to an annualized rate of approximately 81.5%). Meritech’s analysis notes that approximately 8,000 active customers pay an average of about $78,000 per year, and over 1,000 customers have annualized billings exceeding $100,000. SaaStr further notes that 96% of ServiceTitan’s revenue comes from customers paying over $10,000 annually. These customers typically have multiple technicians and are closer to the mid-market, with a clear gap from the smallest trades SMBs.
ScaleFactor’s problem was the opposite. It promised to let small businesses hand over bookkeeping to AI, but in practice relied heavily on human bookkeepers in Austin and the Philippines. It raised over $100 million in funding and shut down in 2020. Failory’s postmortem found that the company called its bookkeepers “customer service officers” and did not count them as costs, in order to make the reality less obvious. The books customers received were full of errors, forcing them to rehire accountants or clean things up themselves. The lesson: if service has not been productized into an asset, scaling only amplifies errors and costs.
AI is also changing the cost structure of FDE work. Standardizable tasks — data migration, field mapping, document processing, initial configuration — are being compressed. Flatfile claims its data import gets customer data into systems 70% faster and speeds up onboarding by 2.17x. Glean’s enterprise AI case studies show that Koch Industries indexed over 1 billion objects in seven weeks, saving an average of 36 hours per new employee during onboarding.
But the harder layer — judgment — still depends on people. Moxo, in its summary of AI onboarding, noted: if the team itself cannot agree on the current process, the process needs to be clarified before considering automation. AI amplifies whatever you build. The same applies to small shop digitization. AI can import data and configure systems faster, but how a shop should quote, schedule production, and define a successful delivery — these judgments still require human involvement.
Beyond using AI to reduce implementation costs, another approach sidesteps the problem of “selling software to small shops” entirely. The idea is to acquire service companies directly, embed AI into their operations, and capture the efficiency gains as profit.
General Catalyst notes that Long Lake acquired 18 service companies and used AI to improve workflow efficiency by 25–30% in HOA management, while growing the new customer pipeline by 10x. Crescendo acquired PartnerHero, merging a customer service BPO with an AI customer service platform; AI now handles 70–80% of routine tickets, reportedly achieving gross margins of 60–65% — four times that of traditional BPO. Eudia acquired Johnson Hana, a firm with over 300 legal professionals, backed by $105 million in funding.
These figures come from early-stage narratives by the companies and investors and have not been independently validated across a full cycle. But they point to a shift: AI companies do not necessarily have to sell software — they can also own service delivery. FDE transitions from an externally billable role to an internal transformation capability deployed post-acquisition.
Back to that sign print shop. When the owner sees a SaaS page that says “auto-quoting,” he usually does not buy in right away. He needs someone who understands his materials, his customers, rework, installation, cash flow, and employee habits — and then translates these into data, processes, and criteria for judgment. The FDE model matters because it names this translation work: converting real-world business logic into digital systems.
But the cost of translation cannot stay this high forever. Three paths could make the equation work. First, lightweight FDE with vertical templates and AI agents, where humans handle only the most critical business judgments while AI handles migration and configuration. Second, following ServiceTitan’s path — building a complete system within a sufficiently large vertical, absorbing implementation costs through higher per-customer revenue. Third, the AI roll-up path — owning service delivery directly and transforming operations with AI.
AI can bring software to businesses it could not reach before, but this will not happen automatically. The most valuable capability is translating real-world business into computable systems. The real question of the AI era is: who can reduce the cost of translation, and ensure that each translation makes the next one cheaper?