An AI consultancy is easy to judge during the pitch and harder to judge once the work starts. The useful test is not how many models appear in the deck. It is whether the consultancy turns one important workflow into a system your people can inspect, operate and improve without depending on a permanent demonstration.
Week one should produce a baseline
The team should observe the task as it works now, including exceptions and handovers. Agree representative examples, elapsed time, quality checks and the cost of an error. Without that baseline, every later claim of improvement is only an impression.
The data boundary should be written down
You should know which sources are approved, where prompts and outputs are retained, whether provider terms permit the data, how user permissions are enforced and what the system must refuse. A London meeting is useful for getting the legal, operational and technical owners around one table. The postcode itself proves nothing.
A pilot should fail in public
A credible pilot includes missing documents, conflicting sources, unauthorised users and questions outside scope. Ask the consultancy to show those failures and explain the stop condition. The narrow system that refuses safely is often more valuable than the broad assistant that answers everything.
- —A named senior person remains accountable after discovery.
- —Model choice follows testing on your tasks rather than a provider preference.
- —Acceptance criteria are agreed before the demonstration.
- —Integration and permission work is visible in the plan.
- —Running cost is tested at realistic usage.
- —Staff training and operational ownership are part of handover.
- —The consultancy is prepared to recommend buying an existing tool or stopping the project.
Consulting earns its place when the recommendation survives the workflow, the permissions and the first difficult example.
Integration is where the value or the risk appears
Useful AI rarely lives in a blank chat window. It belongs beside the documents, CRM records, approvals and people already doing the work. That means identity, access, logging, human review and fallback are product requirements, not technical polish for later.
Handover should begin before launch
Your team should receive the prompts, evaluation set, system access, operating guide and named responsibilities. They should know how to review a poor answer, update a source, watch usage and reverse a change. Continued support can be valuable, but dependence should never be the business model.
What M3CA owns
M3CA keeps strategy, engineering, integration and staff enablement together. We work across leading commercial and open models, use our own systems as a proving ground, and offer private UK infrastructure where the data requirement justifies it. Client evidence is published with names withheld until written permission is granted.
If you are comparing AI consultancies in London, bring one workflow to our Mayfair studio. We will help you decide whether it needs a tool, an integration, a specialist model or no AI at all.
