Build, Buy or Commission? How to Decide on Your Agentic System
Build, buy or commission your agentic system? Compare time to production, fit, memory ownership, control and risk so you can make the call with confidence.
The problemYou can't tell whether to build your own AI team, license a platform or commission a custom system.
Thursday, 6 p.m. The CEO walks out of the month's third AI meeting with three proposals on the table. The CTO wants to hire four engineers and build in-house. A vendor is offering a ready-to-use platform with per-seat licensing. And a software factory is proposing to build a custom system. The underlying question is build, buy or commission, and every option sounds reasonable when it's presented by the person selling it.
The decision matters more than it seems. It doesn't just set what you'll spend this year: it determines who owns your business memory three years from now.
Why this decision is different from traditional software
With conventional software, buying was almost always the sensible choice: a CRM or ERP handles similar processes at every company. Agentic systems change that logic for one reason: their value depends on how well they know your business.
A useful agent needs your rules, your customers, the way you make decisions and your exceptions. That accumulated memory, what we call your company brain, becomes an asset. And the question "who controls that asset?" never came up with traditional software.
On top of that, an agent in production is more than a model. It needs a harness: tools, memory, orchestration, human approval, observability, evals and cost control. Building that harness from scratch is the most underestimated part. We break it down in the anatomy of an enterprise harness.
The three options, in plain terms
Build in-house
You hire or train a team that designs, codes and maintains the system. Maximum control and full ownership. In exchange, you take on the learning curve, the difficulty of hiring scarce talent and the responsibility for building the entire harness, not just the agents.
Buy a platform
You pay for a license to a product that already exists. It's the fastest way to get started. In exchange, you adapt to what the platform allows, your data and memory live in the vendor's system, and the cost grows with every user or every task.
Commission a software factory
A specialized team builds a system tailored to your processes, on a harness it has already proven. You reach production faster than building from scratch, with a better fit than a platform. The key condition: the system, the memory and the data must be yours.
Comparison: build, buy or commission
The table summarizes how the three options perform on the criteria leadership cares about most. These are general tendencies, not rules: any specific case can vary.
| Criterion | Build in-house | Buy a platform | Commission a factory |
|---|---|---|---|
| Time to production | Long: you have to build a team and the harness | Short for standard use cases | Medium to short: the harness already exists |
| Fit with your processes | Total, if the team pulls it off | Limited to what the product offers | High: designed around your processes |
| Ownership of data and memory | Yours | Usually on the vendor's side | Yours, if the contract guarantees it |
| Control over evolution | Total | Depends on the vendor's roadmap | High, with ongoing support |
| Model choice | Open | Whatever the platform decides | Open: the best model for each task |
| Main risk | The project never matures | Lock-in and higher costs as you grow | Choosing a vendor without a proven method |
| Upfront cost | High, in headcount | Low | Medium |
A decision framework in four questions
- Does this process set you apart? If it's the same at any company, buy. If the way you do it is your competitive advantage, don't leave it on a generic platform.
- How urgent is it? If you need results this quarter, building from scratch rarely gets there in time.
- Who will maintain it? An agentic system evolves: new processes, new models, new rules. Someone has to own that.
- Who owns the memory? Ask explicitly what happens to your data, your configurations and your business memory if you switch vendors tomorrow.
Illustrative example: a mid-sized insurance broker decides
Imagine an insurance brokerage with an overloaded claims department. They review all three options.
Build. Their tech team is small and busy with the core system. They figure they'd first have to hire specialists, then build error handling, memory, approvals and monitoring. The timeline doesn't match the pressure from clients.
Buy. They try an AI customer service platform. It works for FAQs, but it doesn't understand their coverage rules or integrate well with their systems. And every adjustment depends on the vendor.
Commission. They decide to commission a claims system: one agent that receives the report, another that checks the policy against coverage rules, another that prepares the claim file, and a human adjuster who approves every decision. They require by contract that the memory and data be theirs and exportable. In parallel, they train two internal staff to operate the system.
In the end, the brokerage combines approaches: it buys a generic tool for FAQs and commissions what sets it apart. Before signing, it estimates the return using the method in how to calculate the ROI of an agentic system.
Common mistakes when deciding
- Comparing only the entry price. Today's cheap license can become the most expensive option once you scale.
- Underestimating the harness. Building agents is the visible part; handling errors, memory and approvals is most of the work.
- Not negotiating ownership. If the contract doesn't say the memory is yours, assume it isn't.
- Locking into one model. The best model for each task changes fast; your system needs to be able to change with it.
- Deciding without a specific process. Choosing an "AI strategy" in the abstract leads to endless debate; choosing for a specific process grounds the decision.
What to do on Monday
Pick a specific process and answer the framework's four questions with your leadership team. Then ask each option to answer three things in writing: a realistic time to production, who owns the memory and what happens if you want out.
Addabra is a software factory: we don't sell licenses or a platform. We build each company's system on our own proven harness, use the best model for each task, and the business memory belongs to the client. If you're facing this decision, we'll help you weigh it with your numbers, not ours.
Frequently asked questions
When does it make sense to buy an AI SaaS platform?
When the process is standard and doesn't set you apart: the platform already handles it the same way for everyone, and you don't need it to know your business in depth. For core processes, where the way you work is your advantage, it usually falls short.
Does commissioning a custom system mean depending on the vendor?
It shouldn't. Before signing, make sure the business memory, data and configurations are yours and exportable, and that the system uses interchangeable models. That is the difference between commissioning and getting locked in.
Can I combine all three options?
Yes, and it is often the smartest approach. Buy the generic pieces, commission the core to get to production quickly, and train internal people to operate and evolve the system over time.
