How to implement AI in a company
A seven-phase method to move from the real process to measurable impact. Not a generic “AI adoption” roadmap: it is the sequence Magnum uses to decide whether scale-up is worth it.
Why most AI projects stall
They start with the tool, not the process. A model is chosen, a connector is wired, and magic is expected. Without a business case, a success metric and an operational owner, the pilot becomes an endless demo.
Magnum reverses the order: process and expected ROI first; technical design second. AI is the means, not the goal.
The seven phases of the Magnum method
Each phase has a deliverable and a pass criterion. We do not advance by calendar: we advance when there is enough evidence.
- Diagnosis — map the real process, volumes, exceptions, systems and data risks. Output: a prioritised opportunity map.
- Business case — estimate recoverable hours, pilot cost, payback and risks. Output: a documented go / no-go decision.
- Pilot (3–4 weeks) — closed scope on a measurable flow, with baseline and agreed metrics. Output: a usable system in a bounded perimeter.
- Validation — measure impact against the baseline. Explicit rule: if the pilot does not prove ROI, we do not scale.
- Scale-up — expand channels, document types or business units under the same control model.
- Operations — runbooks, monitoring, exception queues and clear owners on the client side.
- Governance — roles, audit, action limits (read vs write) and human oversight where it matters.
Validation: the filter that protects the budget
Validation is not a cosmetic checklist. We compare times, error rates and operating cost before and after the pilot. If impact does not justify the next investment tranche, we stop or redesign.
That rule aligns incentives: Magnum does not win by scaling projects that do not work, and the client does not carry a rollout nobody uses.
What the company needs to bring
A sponsor with decision authority, access to a representative volume of real cases (anonymised if needed) and an operational counterpart who knows the exceptions. Without that, diagnosis stays theoretical.
- Data or samples from the candidate process
- Controlled access to ERP / Microsoft 365 / repositories
- A measurable success criterion agreed before the pilot