Where AI actually helps — and where it's a liability

We added AI consulting to what we offer, so it's fair to ask where we actually think AI belongs. The honest answer is narrower than the hype and wider than the backlash. Most of what gets sold as "AI-powered" is a feature that would work better as a plain function call, and most of the real wins are unglamorous. Here's how we decide.

Where it earns its keep

The tasks AI is genuinely good at share a shape: the input is messy language, the output tolerates being approximately right, and a human — or a cheap check — is in the loop before anything irreversible happens.

• Retrieval over your own content.A support assistant that answers from your real docs, policies, and past tickets — grounded in text you control, with citations back to the source — beats a search box and doesn't make things up nearly as often as an ungrounded chatbot does.
• Structured extraction.Pulling fields out of an invoice, a resume, an inbound email, or a PDF the vendor refuses to send as data. This used to be a regex graveyard. Now it's a schema and a model call — with validation on the way out.
• Triage and routing. Reading an inbound message and deciding which bucket it belongs in — not resolving it, just routing it — is a place where being 95% right saves real hours and the 5% is caught by the human it routes to.
• Drafting with a human editor. First drafts, summaries, and rewrites where a person reads the output before it ships. The model does the blank-page work; the human keeps the judgment.

Where it's a liability

The failure mode is always the same: someone put a language model where they needed a guarantee.

• Anything that must be exactly right, silently.Pricing math, tax, permissions, whether a payment goes through. If a wrong answer is expensive and nobody's checking, that's a job for deterministic code, not a model with a confidence problem.
• Deterministic logic wearing an AI costume."Use AI to sort these by date." No. A lot of "agentic" features are a switchstatement someone didn't want to write, now shipped with latency, a per-token bill, and a new way to fail.
• Costs nobody modeled.A feature that's cheap in a demo can be a runaway bill at real volume. If you can't see the per-request token cost before you ship, you don't know what you're turning on.
• "Agents" for everything.A multi-step autonomous agent is the most expensive, least predictable option on the menu. Sometimes it's right. Usually a single well-prompted call with a validated output does the job for a tenth of the cost and a hundredth of the surprises.

The test we actually apply

Before we build an AI feature — or tell you not to — we run it through four questions:

• Does the task tolerate being wrong sometimes?If a 3% error rate is fine, AI is on the table. If it isn't, we're writing normal code.
• Is there a human or a cheap check downstream? Grounding, validation, a person who reads it — something has to catch the miss. Ungrounded output straight to the user is where trust goes to die.
• Would a plain script be cheaper and more reliable?Often yes. When it is, we'll say so and quote that instead — even though "we added AI" sells better.
• Can we see and cap the cost?Token budgets, alerts at 75% of budget, overage billed at cost. If we can't instrument it, we don't ship it.

What it costs to run

An AI feature isn't "free after the build." It has a metered bill (tokens), it needs evaluations so you know it behaves before andafter a model change, and it needs guardrails so a weird input doesn't become a weird output in front of a customer. We build those in and we make the running cost visible, because a feature you can't see the cost of is a feature you can't make decisions about. The provider accounts are yours from day one — same rule as everything else we ship.

Our rule

We won't bolt AI onto something for the buzzword. We've shipped it in production — CombatScorereads an athlete's last twenty sessions and generates a real game plan — so we know both what it's worth and where it falls over. If a plain automation is cheaper and more reliable, that's what you'll get in the quote. The point was never to use AI. The point was to solve the problem.