Last week, a CEO in the capital markets sector asked us whether he should build or buy. There is a product on the market that offers the functions he wants. It has AI embedded, of course.

As our Head of Automation explained why he should build, I understood why software companies are in trouble. The hard part of AI is not generating output. It is connecting systems, structuring data, and deciding how much autonomy software should have.

These are the major obstacles to using AI. Given that work has to be done either way, building starts to look far more attractive than buying.

The value of hardware companies supporting AI workflows has risen in public markets. Often this has been at the expense of software suppliers. The read-through to private markets has hit the valuations of venture capital and private equity funds with heavy software exposure.

These investments are illiquid at the best of times. They received a lot of support from Middle Eastern investors in the past. Now, with tensions in the Strait of Hormuz putting pressure on dollar liquidity, the temptation is to sell.

Exiting now could turn into a fire sale, which is why the US Treasury opened swap lines with the UAE. Silicon Valley cannot afford for cornerstone investors to exit at this time of uncertainty. The US government stepped in to keep important investors solvent.

I wrote in The Hidden Cost of Vibe Coding that small businesses were not going to rebuild their CRMs and ERPs using AI. There is no business case for changing systems that are working well. As a result, the large incumbent software companies should be fine, although they may have to change their business model from per seat subscription to payments for usage.

Growth is another matter. Both incumbents and startups are building AI products that handle coordination inside businesses. In practice, that means linking systems that were never designed to work together. This could be matching sales systems to the production schedule, finance with monitoring of warehouse inventory, or connecting customer service directly to the product team. AI is excellent for this.

The coordination work is simple. A small library of tested prompts and templates is enough to let an LLM handle much of this. The challenge is setting up the system, and you have to do that if you buy software, so it makes sense to build the exact workflow you need.

That said, it is not quite that simple. There is the challenge of creating a data lake. This is sometimes referred to as a vault or a single source of truth. The client I was speaking with has seven different folders where legal files are stored. Data across the business needs to be accessible in one place. Yet you must do this work whether you are building or buying.

Infrastructure is another issue. Our client runs projects from a laptop and they stop working when it sleeps. We need to move those workflows into a cost-efficient cloud environment.

Efficiency also requires the agents to know when to pause. If humans are involved and responses take time, you do not want agents burning tokens while they wait. When building agentic workflows, it is as important to know when to stop as when to go.

All of this may sound like hard work. But the off-the-shelf solution does not address these challenges for you. We run coaching sessions at MSBC to enable you and your team to build AI workflows that behave exactly as you expect.

Meetings, emails and chat platforms are all ways businesses coordinate work. Long before AI, I was recommending ways to make meetings more efficient, write effective emails, and use a single communication tool. AI that does the coordination for you is a better outcome.

A lot of coordination is done by management. If this is all they do, their roles are at risk. But managers also have the experience to be the ones who design, set up and run the AI processes. Those that adapt will survive.

One reason I remain sceptical about AI replacing large numbers of jobs is that companies will run more of their own AI workflows. These will require in-house expertise. It will be easier to teach a domain expert to use AI than to provide an AI specialist with a decade of business experience.

In this case, the losers will be the startup software companies backed by venture capital firms. The growth prospects for more mature software businesses may also be revised. Public and private market investors may be right to mark valuations lower. After all, the role of markets is to place a value on the future cash flows of companies.