← Knowledge base
Category 05

Implementation

What the installation actually involves: timelines, your team's time, the data required, and what happens after go-live.

What does an AI implementation actually involve?

Four stages. Map the process and the numbers as they stand. Build against real historical cases rather than invented examples. Run supervised, with a person reviewing decisions while thresholds are tuned. Then hand over with documentation, reporting and the accounts in your name. Most of the risk sits in stage one, which is why the audit comes first.

Your team's time is concentrated early: access to systems, and a few hours with the person who currently does the work. After go-live the commitment drops to reviewing escalations.

What data do you need from us?

Examples of the work being done well. For intake, historical forms and transcripts; for triage, past filings and how they were classified; for proposals, previous documents and outcomes. We need enough real cases to test against and to define what correct looks like — not a data warehouse, and not anything the system will not use in production.

Where the historical record disagrees with itself, that disagreement is the finding. Resolving it usually improves the human process as much as the system.

Will this disrupt the systems we already run?

It should not. The infrastructure sits alongside your CRM, inbox, calendar and document store and communicates with them through their own interfaces. Nothing is replaced on day one, and the first role runs in parallel with the existing process until its output is trusted. Rollback means switching the routing back, not unwinding a migration.

We insist on the parallel period even when clients are impatient. It is what turns the confidence threshold from a guess into a measured setting.

Who maintains the system afterwards?

You own it, and you can maintain it. Everything runs in your accounts on your data, documented so an internal team or another provider can take it on. Most clients keep us on a light advisory retainer for evaluation review and iteration, but that is a choice rather than a dependency — there is no licence that switches the system off.

Ask any provider this question early. If the answer involves their proprietary platform, the cost of leaving is part of the price you are agreeing to.

What is the smallest sensible place to start?

One role, one measurable outcome, one department. For most high-ticket businesses that is automated first response with human-quality qualification on inbound enquiries — small enough to build in weeks, visible enough that the whole business notices, and directly tied to revenue rather than to an efficiency abstraction.

Small does not mean unserious. It means the proof arrives before the budget conversation does.

Other categories

Next step

Get the answer for your own numbers.

The AI Revenue Audit maps where your business is leaking revenue and hours, and ranks what to install first.