The situation
Several of my clients had no AI infrastructure at all. Another had AWS bills that kept climbing.
Those look like opposite problems and they have the same cause. In both cases nobody had looked at the process underneath. The first set of clients were waiting for a clear reason to start. The second had already started, and was paying for the consequences of buying before looking.
The pattern I see most often is a company that has bought licences for everyone, watched about a fifth of the team use them, and concluded that AI does not work here. What did not work was the order.
What I did
I audit the processes first, then find the tool that fits the gap, even if it is one I have never used.
That order is the whole method. A tool bought before the audit has to have a problem found for it afterwards, and the problem it gets given is rarely the expensive one. An audit first means the tool is chosen against a gap somebody can name, which is also what makes it possible to tell later whether it worked.
For the AWS client, I traced the cost back to engineers firefighting the same compliance bugs again and again, then sourced and implemented an AI-based AWS management tool to fix both. The bill was the symptom. The repeated firefighting was the cost, and it was showing up twice: once on the invoice and once in the engineering time nobody was counting.
For clients starting from zero, I deploy Claude. We start with finance-team Excel automation and move up to CEO-level operations like email follow-up and team-activity dashboards. Starting in finance is deliberate. The work there is repetitive and the output is checkable, so the team can see within a fortnight whether it is producing something correct. Trust earned on a checkable task is what buys permission to go near the CEO's week.
Where the tool stack itself is the problem, I map the redundant work and replace it with one connected ops system in monday.com.
Every rollout runs on PACE
Problem, Allies, Concerns, Experiment, with thirty days of proof on activation, return, acceptance, and one business metric the CEO picks.
Allies and Concerns are the two most people skip, and they are the reason rollouts fail. Allies means finding the people who will use the thing without being told, because adoption spreads sideways between colleagues rather than downwards from an announcement. Concerns means asking out loud whether this is how the redundancies start, because the team is already asking it quietly, and an unanswered version of that question is enough to kill any tool.
The CEO picking the business metric matters too. It stops the review turning into a discussion about how impressive the demo was.
The result
I measure every rollout the same way: hours per week saved on manual work, and how many repetitive workflows disappear completely.
- 40% of writing and repetitive work saved with ChatGPT at AdvanceB2B
- 35% productivity lift from Zapier-to-HubSpot automation
- 30 days of proof on every rollout, before anyone commits
If your AI bill is climbing
Work backwards from the invoice to the behaviour. Cost that grows without output behind it is almost always a process being run repeatedly by people who are working around something, and the tool is only making the workaround faster.
Then pick one process, give it thirty days and one number. If it does not move, you have lost a month instead of a budget.
I proved the approach in-house first, before taking it to a client.