Most writing about AI strategy assumes a company with a data team, a platform budget and a transformation office. A freight brokerage with forty people has a dispatcher who is also the IT department. The advice does not translate, and the vendors who arrive with it tend to leave with a pilot that nobody uses.
What follows is the roadmap we actually run with operators of that size. It has four steps and the first one is not technical.
1. Find where the hours go
Sit with the people doing the work for a day. Not a workshop; a day at the desk. Count the interruptions. Watch what gets copied from where to where. Ask what they do when the phone rings while they are mid-email.
The pattern is consistent across industries. The hours go to a small number of repetitive, interrupt-driven workflows that happen to involve reading: the inbox, the phone, the onboarding packet, the quote, the filing. Each is done by a specific person, each has rules that person has never written down, and each is invisible to the owner, who sees only that things are slow.
Write the workflows down with the hours beside them. For a forty-person company this list has eight to twelve items. Three of them will account for most of the hours.
2. Price them
For each workflow: hours per week, who does them, what an error costs, and what would be unlocked if the hours came back. The last is usually the largest number and the hardest to say. A dispatcher who is not filing is quoting; a principal who is not filing is drawing.
Then make a build-buy-wait call on each. Some workflows are solved by software that exists and should be bought. Some are not worth automating because the hours are small or the rules are too fluid. A few are worth building because they are specific to the business and the hours are large. Those few are the roadmap.
3. Build one system
One. Scoped to ship in weeks, not quarters, and scoped to a workflow whose success can be measured in hours returned. The architecture is always the same: a deterministic core for anything that is a function of the business’s numbers, a model at the edges for reading and drafting, a quick-check queue for the ambiguous cases, evidence shown on every decision, the business profile in one configuration file, and a mode that runs with no keys so the demo and the production fallback are the same thing.
The first system should be the one with the clearest hours. For the architecture practice it was email filing. For the brokerage it was carrier onboarding. For the reseller it was the phone. None of these is glamorous and all of them were working within weeks, which is the only measure of a first system that matters.
4. Measure, then widen
After a month, count the hours again. The number is the case for the second system and the budget for it. The second system is usually easier, because the integrations (mail, files, telephony) are already built and the team has learned to trust a dashboard that shows its reasons.
By the third system the company has something that looks like a platform: shared integrations, one business profile, a pattern for review queues. It arrived by building systems that paid for themselves rather than by buying a platform and looking for uses.
What to be sceptical of
Pilots with no hours attached. Platforms that must be adopted before anything works. Assistants that cannot say where an answer came from. Anything that requires the dispatcher to become a prompt engineer. Any vendor who cannot run the demo with the network cable out.
What to insist on
A day at the desk before a proposal. A ranked list with hours. One system first. A dashboard that shows why. A file you can edit yourself. A test suite the vendor leaves behind. And a measured result before anyone says the word transformation.
That is the whole roadmap. It fits on a page because the company fits in a room, and that is the advantage a forty-person company has over the ones the advice is usually written for.
About the authors. Neugent Consulting designs and builds AI systems for operators. The examples in this piece come from client work; clients are anonymised and details are simplified.
