Capability mapping

You can't spot opportunities in a technology you can't describe. Strip the vendor language and today's deployable AI does six things reliably enough to build on. Learn them as verbs, and every process walk-through becomes an opportunity scan:

• Classify — read something messy, assign it a bucket: route the email, triage the claim, flag the risky invoice. The workhorse; most 'AI transformation' is classification wearing a suit.
• Extract — pull structured facts out of unstructured mess: PO numbers from emails, terms from contracts, weights from shipping docs. Alder's ops teams re-type data between documents and systems all day; extraction is that, automated.
• Generate — produce a competent first draft: the customer reply, the job posting, the incident summary. The human moves from author to editor. Value scales with how much your org writes.
• Converse — sustain a useful dialogue over your knowledge: the policy Q&A bot, the onboarding assistant, the customer self-service flow. Needs grounding in your documents to be safe (your teams will hear this as 'RAG').
• Predict — estimate what happens next from history: demand next month, which shipment will be late, which customer will churn. The oldest family (classic ML), still compounding — and entirely dependent on decent historical data.
• Optimize — choose best options under constraints: routes, load plans, schedules, pricing. Deep Alder relevance; often more operations research than 'AI', which matters zero to the P&L.

Matching capability to task: the three fit tests

• The error-tolerance test. AI outputs are usually right. Tasks that tolerate a reviewed draft or a triaged queue fit today; tasks where a single silent error is catastrophic need humans in the loop by design, not aspiration. 'Where does usually right, always reviewed create value?' is the strategist's question.
• The data test. Predict and optimize eat historical data; converse eats well-organized documents; extract and classify are the least demanding starters. If the honest answer to 'do we have the data?' is no, the initiative is a data initiative first (Module 3 returns to this hard).
• The volume test. AI's economics are per-unit: high-volume tasks (Alder: 3,000 delivery-exception emails a month) repay setup; rare tasks don't. Capability without volume is a demo.