Skip to content
Personal Learning Notes
7 min read

AI for SMEs: Build, Buy SaaS, or Outsource? Don't Forget the Hidden Costs

Facing your first AI project as an SME? Deciding between building in-house, buying SaaS, or outsourcing can be tricky. This article breaks down the hidden costs of data preparation, system integration, and long-term maintenance to help you make a sound decision beyond just model pricing.

When small and medium-sized enterprise (SME) owners first seriously consider AI, it’s often not the technology that captivates them, but the confusing array of quotes. Software vendors suggest annual subscriptions to ready-made systems (SaaS); outsourcing providers quote significantly higher; and some even recommend hiring staff to develop solutions in-house. Each approach has its merits, but sales pitches rarely highlight where the real costs lie. My direct assessment is this: most SMEs don’t need to build their own AI models. What truly determines success or failure are three often-overlooked costs: data preparation, system integration, and long-term maintenance.

This isn’t to say models aren’t important, but rather that model capabilities are no longer a scarce resource. A Gartner survey published in 2026 noted that only 22% of organizations successfully scale AI across multiple business units, yet 85% of functional leaders plan to increase their investment. Increased investment with limited scalability suggests the problem usually isn’t with the model itself, but with the surrounding supporting work. A research report from the China Academy of Information and Communications Technology (CAICT) also characterizes SME AI adoption as being in its “initial stage of scaled application,” with actual investment remaining cautious.

First, Separate “Building Your Own Model” from “Building Your Own Application”

Many discussions conflate these two concepts. “Building your own model” means training an AI large language model from scratch. “Building your own application” means using existing large model APIs to construct an interface according to your business processes and connect it to your own data. The former is like building your own power plant; the latter is like buying electricity and wiring it up.

Buying electricity isn’t expensive. For example, a text generation model listed on Alibaba Cloud’s Bailian Model Pricing page shows an input price of $1.65 per million tokens and an output price of $4.951 per million tokens. A “token” is the unit of measurement large language models use to process text; a piece of text is broken down into many tokens, and the model charges based on the number of tokens. This pricing level means that simply calling the model typically won’t be the largest cost driver for an SME project.

Therefore, owners shouldn’t be intimidated by “building your own model” or misled by the perception that “models are expensive.” The real decisions you need to make are: which business process will AI be integrated into, is your existing data sufficient, can your systems connect, and who will maintain it afterward? The following three cost categories are key.

Hidden Cost One: Data Preparation, Often More Expensive Than the Software Itself

Imagine a small manufacturing plant wanting to use AI to automatically process customer inquiries and quotes. During a demo, the salesperson shows the AI reading a few standardized PDFs, with excellent results. After deployment, however, they discover that historical quotes are scattered across Excel spreadsheets, messaging apps, and scanned paper documents; the same customer might be listed by full name sometimes, and by abbreviation others; the same product might have three different names in the system. AI isn’t incapable of processing this, but faced with inconsistent raw data, the output results are unreliable.

To transform this information into something AI can reliably use, “data preparation” is required. Data preparation involves at least three things: standardizing formats, filling in missing fields, and unifying business rules. For instance, customer names, product numbers, currencies, and dates must be mapped to fixed fields—this is what’s referred to as “structured data.” This work often requires business personnel to make judgments; it cannot be simply handed off to IT, much less solely to an outsourcing vendor.

Many owners assume that once they buy AI, it will be ready to use, effectively deferring the cost of data preparation until mid-project. The workload that emerges at this stage is often higher than the software subscription fee and more taxing on team patience. It’s similar to the problem of unused enterprise knowledge bases: without a long-term maintenance mechanism, data that’s been prepared will eventually become disorganized again. For related mechanisms, see If Your Enterprise Knowledge Base Isn’t Used, Check These Three Mechanisms.

Hidden Cost Two: System Integration, Not Just Installing a Plugin

AI isn’t just a standalone web chat interface. To generate business value, it often needs to connect with existing systems. ERP (Enterprise Resource Planning) systems manage orders, procurement, and finance; MES (Manufacturing Execution Systems) oversee shop floor production; CRM (Customer Relationship Management) systems handle customer interactions and sales tracking.

Imagine a small factory implementing AI for quality inspection: after the model identifies a defect, it needs to read real-time equipment parameters, write the results into the MES, and then call different judgment rules based on the product model. If the equipment doesn’t have open APIs, or the MES vendor doesn’t cooperate in granting access, even the most accurate model won’t make it to the production line.

Public data also shows integration as a major hurdle. A Gartner survey from 2025 found that 48% of infrastructure and operations leaders cited integration difficulty as a major challenge for AI adoption, and 50% mentioned insufficient budget. Integration is expensive because it involves permission synchronization, data synchronization, error handling, and security boundaries—it’s not as simple as installing a plugin.

This aligns with my assessment of digitalization for small factories: system integration costs must be calculated upfront, not after deployment. For related calculations, see SMEs: Don’t Implement MES Yet – Calculate These Three Costs First.

Hidden Cost Three: Long-Term Maintenance, Running in Year One Doesn’t Mean Running Forever

An AI project doesn’t end after delivery. Models change versions, business processes adjust, employees come and go, and knowledge becomes outdated. A system that appears to run well in its first year might cease to function in subsequent years due to an interface change or a data format modification.

The same Gartner survey from 2026 revealed a notable contrast: high-performing organizations consistently track the ROI of their AI projects, with 81% reporting positive returns. In contrast, 29% of AI projects in low-performing organizations had an unclear ROI. This disparity indicates that continuous management is more crucial than whether the initial version worked.

The maintenance aspects differ across the three approaches. SaaS involves annual renewals, version upgrades, and data export limitations. For outsourced projects, if no one internally takes ownership after delivery, subsequent modifications can be very troublesome. In-house built applications have no subscription fees but rely on individual employees, making the system a “black box” that no one dares touch if those key personnel leave.

Therefore, when comparing quotes, don’t just look at the first year. Factor in subsequent maintenance, redevelopment, and internal coordination time to get a complete picture of these three hidden costs.

My Assessment: How to Choose

My judgment is: most SMEs should follow the principle of “buy off-the-shelf for general use, customize for core functions later, and only build in-house if you have strong data capabilities.”

For general, non-core scenarios, start by buying SaaS. Examples include meeting minute summarization, initial customer service screening, or product description translation and refinement. The key isn’t how powerful its features are, but whether data can be easily imported and exported, preventing vendor lock-in.

For core scenarios that are highly standardized, prioritize integrable SaaS. Make sure open APIs, clear permissions, and data exportability are written into the procurement terms, rather than just taking a salesperson’s word that “we support everything.”

For core, differentiated scenarios that require deep integration with your business data, then consider custom outsourcing. In this case, don’t just specify “accuracy” requirements; include verifiable business metrics in the contract. You can refer to AI Contract Acceptance: Don’t Just Write Accuracy, Write These 4 Types of Metrics.

Completely building an application in-house should only be considered if the company already has internal IT development and data leads, and is willing to commit long-term resources. Even then, building your own model is unnecessary; using existing model APIs is sufficient.

When faced with any quote, ask just three questions: Does it include data preparation? Does it include system integration? How are maintenance and modifications priced for subsequent years?

What I Haven’t Fully Figured Out

This assessment might not hold true in certain situations. If an enterprise has a well-standardized data foundation, open system interfaces, and a dedicated digitalization lead, these three hidden costs will be significantly reduced, making SaaS purchases or even in-house builds much smoother. If the budget is very small and the goal is just a one-off small task, any systematic solution might be overkill. If an industry has extremely strict data compliance requirements, private deployment might become a hard constraint, forcing a change in the choice of SaaS and public cloud models.

What I haven’t fully figured out is whether future advancements in model capabilities will significantly lower the barriers to data preparation and system integration. Based on currently available public materials, models are progressing rapidly in processing text and code, but internal enterprise permission boundaries, business rules, and data reliability still require human definition. It’s possible this assessment will need revision in a few years, but based on publicly available information in 2026, SMEs should not bet their investment on the idea that “a stronger model will eliminate the need for data preparation.”

Reference Sources

Write to me

Thoughts or questions about a note or a project? Leave me a message here.

Messages are never published. They go straight to my inbox and are used only to reply to you.

Or just email me: pierceliu@live.com