Should you build your own AI strategy agent?

Start with the strategic process

The first question is not which model to use. It is what needs to happen before what. If competitive analysis is supposed to influence positioning, what has to come out of the competitive work so the positioning step can actually use it? If the audience work changes the diagnosis, which downstream work needs to be reconsidered?

If a client rejects one assumption but approves the rest, what becomes authoritative? A collection of impressive AI tasks is not yet a strategic system. The moment I became convinced that our own work was becoming a system was not when one module produced a good answer in isolation.

It was when the outcome of one piece of work visibly changed the outcome of the next. That connection is the job.

Define what the strategy platform needs to connect before choosing the implementation.

Your best strategist becomes part of the build

This is where the work gets uncomfortable. The agency needs somebody who can explain how they think. Not only which framework they use. “Look at the 3Cs and the answer will present itself” is not a system description. What does the strategist notice first? What makes them distrust an apparently sensible answer?

How do they decide whether an insight is actually useful? When do they stop analysing and choose? What does a bad positioning look like even when all the words are technically correct? A great deal of senior expertise is tacit. People know when something is wrong before they can explain why.

If you want to build a serious strategic system, some of that judgement has to become explicit enough for the system to work with. Many very good strategists find this surprisingly difficult.

Do not begin with the shiny parts

Teams often start by debating models, vector databases, data ingestion and whether the newest release is 8 per cent better on some benchmark. Those things matter eventually. They are not the first problem. Decide which strategic modules you actually need. Define the process. Define the dependencies. Decide what should persist. Work out the agency's strategic philosophy and which frameworks fit it.

Then architecture starts to become meaningful. And yes, think properly about architecture.

Do not assume a no-code workflow builder is automatically the right foundation because the first demo takes an afternoon. “No code” has a habit of becoming a lot of code plus a visual interface once the system becomes complicated.

The hardest handover may be preserving the right client context, including accepted and superseded decisions.

Budget serious time

Before Mirel became a product, we spent roughly 800 hours on the system thinking alone. That is not 800 hours of writing prompts. It is not a claim that another team must spend exactly the same amount. It is a useful warning about the category of problem. You are trying to formalise a strategic process, connect dependent work, preserve context, encode quality criteria and create something that remains coherent when real agency mess enters the system. That takes time from people who understand strategy and people who understand systems.

When building may make sense

There are agencies where I can see the case. An execution-heavy agency with deep internal AI capability may already have strong proprietary workflows downstream. Strategy may only need to feed those systems in a particular way. A very specific internal process can justify a very specific internal build.

The decision becomes less attractive when the agency mainly wants to avoid paying for software while underestimating the internal cost of product development. That cost includes maintenance too. Models change. APIs change. Requirements change. The strategic process evolves. Somebody has to own the thing after the interesting build phase is over.

Compare that requirement with the existing tools available to strategists.

The honest answer

Can you build your own AI strategy agent? Absolutely. If you want to do it properly, treat it like building a strategic product for your agency. Because that is what you are doing.

Make the cost model explicit

Strategy expertise is part of the build budget. RedSofa’s Berlin 2025 report lists freelance senior strategy at €400–€600 a day. Ten days of that input would cost €4,000–€6,000 in fees, before engineering, testing, model usage or maintenance. Ten days is an illustrative allowance, not an estimate of what your build will need.

For an illustrative October 2026 scenario, assume 30 development days at €700 a day (€21,000), ten strategy days at €500 (€5,000), and two maintenance days a month at €700 for a year (€16,800). That is €42,800 for those labour inputs alone. Development rates and effort are assumptions, not a supplier quote or a forecast from our own build.

Add separately priced hosting and model usage, evaluation work, security review and any internal operating time. Obtain those figures from the intended architecture and contracts before comparing against a product subscription. A prototype budget that excludes ongoing ownership cannot answer the build-versus-buy question.