The board has asked what our AI strategy is
Usually the honest answer is that there is no strategy yet, only tools. We start from the work instead: which processes are repeatable enough to hand over safely, and which are not.
AI
Everyone has assistants. Almost nothing has changed.
The gap between using AI and getting work out of it comes down to whether the system can reach your data, act inside your permissions, be measured when it is wrong, and stop for a person when the decision matters.
Where we are called in
Four situations that reach us.
Licences are paid and nothing got faster
Assistants are everywhere, adoption is reported, and no process is shorter. We put models inside a defined process, with real data and named approvers, so the work itself gets shorter.
It cannot go near our data
Regulated, confidential or client information, where the real questions are what the system may see, what it may keep and what it may say. That is an architecture problem before it is a model problem.
We tried and it was wrong too often
Wrong how often, on which cases, measured against what? We build the evaluation set first, so quality becomes a number you can track.
Where models fail
A confident answer and a correct answer look identical.
The failure that matters is fluent, plausible and wrong, with nothing on the surface to say so. One model cannot warn you, because whatever produced the error also produced the confidence.
Make them disagree on purpose
Several models, given the same work from different vantage points, fail in different ways. Where they diverge is where the answer is fragile, and that divergence is something you can act on.
Keep the disagreement visible
Averaging opinions buries the one that was right. A split panel is a signal, and it belongs in front of a person.
Ask reviewers to refute
A reviewer asked to confirm will confirm. Reviewers are instructed to break the claim, and it stands only when the attempt fails.
Blind spots are structural
Models trained on the same material share the same errors, so a second opinion from a near-identical system is worth very little. Diversity of model, instruction and vantage point is what makes review real.
Measure the failures
Evaluation sets are built from the cases that went wrong before, so the number tracks the thing you are actually afraid of.
What it actually takes
Underneath the demo, this is the work.
A demonstration proves a model can do something once. Putting it into an operation is a different discipline, and it is where most programmes quietly stop.
Working from your own data
Retrieval over defined sources, with citations back to the source, so an answer can be checked.
Inside your permissions
The system acts as the person asking. What that person cannot see, the system cannot use.
Measured against real cases
Evaluation sets built from your real cases, so changing a prompt or a model becomes a measurable regression.
Tools in a sandbox
An agent that can act runs inside a bounded environment, with an audit trail of everything it touched.
A person on the consequential path
Named approvers on the decisions that carry material consequence, and a record of who approved what.
The honest part
Most of the value sits around the model.
The work that makes AI useful is unglamorous: naming the process, finding where the data actually lives, deciding who is accountable for a wrong answer, and being willing to leave a step human.
Rules or ordinary software are often the better answer. We choose the smallest intervention that holds.
Discuss an AI programme