The challenge
The group's growth equity owner engaged Midici to work with the company on bringing AI into the business. The group had been assembled through acquisition, and each company had its own systems, support model and engineering structure: a sixteen-person in-house team in the UK, a three-person founder-led team in Australia, and outsourced development in the US. Leadership at all three companies was enthusiastic about AI. The owner wanted to know where AI could expand the group's capability within 90 days, and where it should not be used at all.
Our approach
We ran a structured interview program over six weeks. It began with the group CEO, continued with the functional leaders at each company, and finished with individual contributors whom those leaders identified as doing notable work. We scored every session against a five-dimension rubric: process maturity, data and systems readiness, current AI adoption, team change appetite, and opportunity density. The scores describe the state of each workflow and the systems around it, so the assessment rates the work rather than the people doing it.
Interviews with leaders describe their intentions accurately but their teams' behavior less reliably. To close that gap, we sent an anonymous survey to the in-house engineering teams, and nearly every engineer responded (17 in total).
What we found
Engineering adoption was at an earlier stage than it appeared
The survey showed that 41% of developers had no AI tool license, 88% had no configuration specific to their codebase, and every pull request was reviewed entirely by people. Confidence in AI output scored a net promoter score of −56. Reviewing the workflows showed why. Most developers were pasting code into a chat window that could not see their codebase, and judging AI by the results. The developers wanted clear guidance, and one wrote:
“It would also be useful to know which tools we are allowed to use.”
Support knowledge was scattered across individual inboxes
We found at least six separate support channels across the three companies, three of them individual people's inboxes. None had ticketing, response-time tracking or satisfaction measurement. Each answer to a customer lived only in one person's sent folder, which created key-person risk and hid the company's best source of product feedback.
Senior operators spent much of their time gathering context
Senior people at all three companies gathered information by hand from email, calls, documents and calendars before they could start the work that needed their judgment. As one of them put it:
“Everyone has big ideas, everyone wants to be doing more, and everyone is already doing too much.”
The team's own figures sized the opportunities
Renewals took about 40% of sales representatives' time in peak season, and the sales lead estimated that 75% of that work was routine. About 145 recent trial prospects had never been followed up systematically, and new trial users waited one to three days for their login details.
What we recommended
The report prioritized 30 projects and built three flagship programs from them:
- Consolidating support onto a platform the group already licensed, followed by AI-drafted replies that a person always reviews and sends.
- Tools for operators to gather context: short, practical training that connects existing tools to AI agents, starting with the people already asking for it.
- An engineering AI program that puts guardrails in place first (test integrity, stronger continuous integration and shared skills), following the UK engineering lead's own principle of “go slow to go fast.”
Three of the most valuable steps cost almost nothing: a published policy on approved AI use, a license for every developer, and protected time for training.
The report was equally specific about where to hold back. It kept AI away from the group's government-level enterprise partnerships, ruled out automated responses to customers, and deferred autonomous code merges until guardrails were in place. It also recommended an investment unrelated to AI, a business process analyst, because applying AI to an undocumented process automates its inconsistencies along with the work.
