From a small startup to a leading global agency in 10 years

Ten years ago Vertexa was four people and one client who had taken a risk on us. This month we passed five hundred completed engagements across twenty-six countries. That is the version that goes in a press release. The useful version is what happened in between.

The first thing we got wrong

For the first three years we sold capacity. Clients asked for engineers, we supplied engineers, everyone was reasonably happy and almost nothing improved. The projects that went badly went badly for organisational reasons — unclear ownership, a decision nobody would make, a process the software was expected to fix by implication — and supplying more engineers made those worse rather than better.

The change was deciding that we would say so. Out loud, in the room, early. We lost work over it. We also stopped losing engagements at month seven, which turned out to be the better trade.

What discovery actually buys you

Every engagement now starts with two to four weeks of discovery, priced separately and delivered as artefacts the client owns whether or not they continue with us. Roughly one in eight discoveries ends with us recommending the client does not build the thing.

That sounds like a terrible commercial policy. In practice it is the single strongest thing we do for referrals, because the recommendation is credible precisely because it cost us the follow-on work.

Distributed before it was fashionable

We have been distributed since 2014, initially because the best data engineer we could find lived in Bengaluru and had no interest in moving to London. That forced a set of habits — written decisions, asynchronous review, documentation as a first-class deliverable — that turned out to be good engineering practice independent of geography.

The decisions we would repeat

• Hiring for judgement over stack familiarity. Frameworks change; the instinct for what will hurt in eighteen months does not.
• Refusing fixed-price contracts on genuinely exploratory work. Fixing a price on something nobody understands yet just moves the argument to month four.
• Building our own products. Operating Orbit and Lumen at scale keeps us honest about what we recommend to clients.

The decisions we would change

We waited far too long to build an internal platform. For six years every project rebuilt its own CI pipeline, its own environment provisioning, its own observability defaults. The engineering team had been asking for it since year two. We should have listened in year two.

What the next ten years look like

More product, more data, and considerably more work on the operational side of AI — the governance, evaluation and monitoring that most organisations are currently skipping. The technology will change. The reason programmes fail almost certainly will not.