Custom AI vs off-the-shelf tools: how to decide

Buy off-the-shelf AI when the work is standard and a tool fits without workarounds. Build custom AI when the workflow is distinctive, central to how you compete, or deeply tied to your own data and systems. Most companies end up doing both, for different jobs.

The harder part is that "build or buy" hides more than two choices, and the option that looks cheapest is not always cheapest over the life of the system. This guide sets out the four real options, the questions that decide between them, and the hybrid approach that often wins. We write it from both sides of the table: DeepVelocity sells its own AI products and builds custom systems for clients, so we see when each one is the right call.

The four options, not two

• Buy an off-the-shelf tool. A ready-made application, such as a meeting note-taker, a writing assistant, a transcription service or a general chat assistant. You sign up and use it. Little setup, little control.
• Configure a platform. Business editions of general AI assistants, or the AI features now built into software you already run, such as your CRM, help desk or office suite. You set them up around your process, connect your data and set permissions, but write little or no code.
• Buy a specialist product. Software built for one job in one industry, with the domain rules already encoded. Two examples from our own portfolio: VelocityBids finds public bids and fills out the paperwork, and QuarterHour audits outside counsel legal bills line by line.
• Build custom. A system designed around your workflow, data and systems: agents that take over steps of a process, AI tools built into your existing software, or a machine learning model trained on your own data. You own it, and you are responsible for it.

The options sit on a spectrum. Moving from buying toward building, you gain fit, control and ownership, and you take on more cost, time and responsibility. The right point on that spectrum depends on the workflow, not on a company-wide policy.

The questions that decide build vs buy

Judge each workflow on its own. The same company can rightly buy a note-taker, configure the AI in its help desk, and build a custom quoting system. These seven questions settle most cases.

How standard is the workflow?

If most companies do a task in roughly the same way, such as summarizing meetings, drafting routine emails or transcribing calls, someone has already built a good tool for it, and it will keep improving without any effort from you. If your process has its own rules, exceptions and judgment calls, a generic tool will either ignore them or push your team into workarounds. Workarounds are a hidden cost: they show up as rework, missed steps, and people quietly going back to the old way.

Is it how you compete?

Buy for parity, build for advantage. If a workflow is a back-office chore, being as good as everyone else is enough, so buy. If it is how you win business, such as how fast you quote, how accurately you price or how well you find leads, then a tool your competitors can buy too cannot be your edge. That is where custom work earns its cost.

How sensitive is the data, and where may it go?

Check where a vendor stores and processes your data, who can see it, how long it is kept, and whether it is used to train the vendor's models. Business and enterprise plans usually offer stronger terms than consumer ones, and they are often enough. Where contracts, regulation or client confidentiality demand tighter control, a configured platform with the right agreements, or a custom system running in your own cloud, may be the only acceptable choice.

How deep are the integrations?

A tool that works on its own is easy to buy. But most of the value of business AI comes from reading and writing your real systems: pulling an order from the ERP, updating the CRM, filing a document in the right place. Off-the-shelf tools connect to popular systems in standard ways. If your work spans several systems, older software or your own databases, integration can become most of the effort, and a build is easier to control.

What volume will it handle?

Per-seat and per-use pricing is cheap at small scale and can become the largest cost at high volume. When a workflow runs thousands of times a day, the cost of each task matters. A custom system can use a smaller, cheaper model for routine cases, avoid repeating work it has already done, and send only the hard cases to the most capable models. At low volume, none of that is worth engineering.

How accurate does it need to be?

A draft that a person reviews can tolerate mistakes. A decision the system makes on its own, or a number that moves money, cannot. General tools are tuned for the average user. When accuracy on your own cases is what matters, you need to measure it on your own data, and often to tune or build for it.

How fast do you need results?

Buying is faster: days or weeks, against the weeks or months a well-run custom project takes. If the need is urgent and the work is standard, buy now. If the work is strategic but also urgent, start on a platform and build the distinctive parts in stages, which is the hybrid approach described below.

Try the chart below. Rate a workflow on each question, or pick an example, and see where it lands between buying and building.

Compare the total cost, not the price tag

A monthly subscription is easy to set against a build quote, and the comparison is usually misleading. Over three years, count:

• For buying: subscriptions for every person who needs access, usage fees above plan limits, setup and integration work, the time your team spends on workarounds, and the cost of switching if the vendor changes its price or direction.
• For building: the build itself, ongoing maintenance and model updates, AI model usage and hosting, and the time of whoever owns the system on your side.

Neither side is cheap by default. Our guide to AI consulting pricing models explains how custom builds are usually priced, including structures where the firm that builds the system shares the risk.

Lock-in and ownership

Every option creates some dependence. With a bought tool, your process, data and team habits sit inside someone else's product, and its prices, features and terms can change. With a custom build, you depend on whoever maintains it. Protect yourself either way:

• Make sure you can export your data in a usable format, whatever you buy.
• For custom work, own the code, prompts, trained models and documentation, and put that in the contract.
• Keep the parts that are specific to your business, such as rules, prompts and data pipelines, separate from any one AI model provider, so you can switch models as better or cheaper ones arrive.

The hybrid: buy the base, build the last mile

The best answer is often neither pure option. Buy or configure the common layer, such as an AI platform, a model provider, or the AI already in your software, and build the last mile that makes it fit: the connections to your systems, your business rules, your review steps, and the checks on quality.

This keeps you on general-purpose technology that improves every year, while putting your money where your business is different. It is also how we build most custom AI solutions: we use proven models and services as components, and spend the effort on the workflow around them. Our article on AI agents for business covers how to decide which steps of a workflow an agent should own.

When a custom machine learning model beats general AI

General-purpose AI models, the large language models behind chat assistants, are strong at reading, writing, summarizing and reasoning over text. They are not designed to predict a number from your history. When the task is a prediction from your own structured data, a machine learning model trained on that data is usually more accurate, cheaper to run, and easier to test. Typical cases:

• Forecasting: demand, inventory, staffing or cash, from your own history and seasonality.
• Scoring: which leads are likely to buy, which accounts are likely to leave, which invoices need a closer look.
• Pricing and estimating: quotes and estimates based on the jobs you have actually done.
• Classifying images, signals or sensor data into categories that are specific to your business.

Our machine learning lead engine for commercial roofing is an example: a custom model reads roofs from aerial imagery and scores which ones need coating, something no general assistant does out of the box. A custom model needs enough good, labeled data, and a fair test against a simple baseline before anyone relies on it. If that data is not there yet, a general model with careful instructions may be the right place to start while you collect it.

Signs you made the wrong call

Revisit the decision when you see:

• You bought, but the team works around the tool, the data you need is stuck inside it, costs keep rising with usage, or your competitors have exactly the same capability.
• You built, but the system does something a mature product now does better, nobody owns its upkeep, or most of the effort goes into parts that are no different from anyone else's.

Neither is a failure. Needs change and products improve. What matters is reviewing the choice on purpose rather than by drift. If a custom pilot has stalled, our guide to why AI pilots fail covers the usual causes and how to fix them.

How DeepVelocity approaches build vs buy

We work on both sides of this decision. We build and run our own products, VelocityBids, QuarterHour and Editorial AI, for jobs where many companies share the same need. We build custom AI solutions and machine learning models for work that is specific to one business, such as the AI-assisted quote workflow that turns customer drawings into quote inputs using a shop's own machines, labor and material rates. When a client asks us to build something they could buy, we tell them. Every custom engagement starts by mapping the workflow and sizing the value, so the decision to build rests on evidence rather than enthusiasm.