Many businesses start with AI in the same way. Someone finds a tool that writes emails, summarises calls or answers questions, and the team starts using it. Output goes up. A few weeks later, the same problems are still there: leads are still followed up late, clients still wait for answers, and leaders still do not trust the numbers.
The reason is simple. AI makes a step faster. It does not decide whether the step should exist, who owns it, or what happens next. If the workflow around the tool is broken, AI helps you make the same mistakes more quickly.
Three signs the workflow is the real problem
- Nobody owns the outcome. Several people touch a lead or a request, but nobody is accountable for the result.
- The hand-offs are invisible. Work moves between tools by copy and paste, email or memory, so things fall through the gaps.
- There is no baseline. Nobody can say how long the process takes today, so nobody can say if it improved.
If any of these are true, a new AI tool will add activity, not results.
Fix the workflow, then add intelligence
We use a simple order of work. It is the same order behind our method: discover, design, build, test, deploy and optimise.
- Map what really happens. Follow one real lead, order or request from start to finish. Write down every step, tool and person.
- Agree the outcome and the owner. Decide what good looks like, for example a reply within one hour, and who is accountable.
- Remove steps before you automate them. Many steps exist only because two tools are not connected.
- Decide where people must approve. Pricing, commitments and sensitive messages stay with a person.
- Only then add AI and automation. Give AI the repetitive, low-risk steps: sorting, summarising, drafting and routing.
- Measure against the baseline. Compare the new process with the old one, using the same measure.
- DATAMap the real processFollow one real case.
- HUMAN APPROVALAgree owner and outcomeWho is accountable?
- AUTOMATIONRemove and connect stepsFewer hand-offs.
- AI ACTIONAdd AI to low-risk stepsSort, draft, route.
- KPIMeasure against baselineSame measure, before and after.
The question is not “Which AI tool should we use?” It is “Which workflow, if it ran reliably, would change the business?”
What this looks like in practice
Take website enquiries. In a broken version, enquiries land in a shared inbox, someone forwards them, a reply goes out a day later, and the CRM is updated when someone remembers. Adding an AI email writer only produces faster drafts that still wait in the same inbox.
In a designed version, every enquiry creates a CRM record automatically. AI sorts the need and suggests a fit score with its reasons. Low-confidence enquiries go to a review queue. A person reviews high-value leads and sends the first real reply. The response time is measured every week.
The AI in both versions might be the same. The result is not, because the second version is a system.
Where to start
Pick one workflow that happens often, matters commercially and has a clear owner. Measure it as it is today. Then redesign it before you add any new tool. One well-designed system will teach you more than ten disconnected experiments.
