Ask ten consultants how AI can help your business and you'll get eleven strategy decks on ai for business, each with a slide reading "AI-first" in a font that costs more than the insight beneath it. None of them are exactly lying. They're just drawing the map for the wrong country. Most of what ranks for that phrase was written by software companies, about software companies, for other software companies. If what you run involves hotel rooms, insurance policies, pallets, or anything else you could plausibly drop on your foot, that map doesn't cover your terrain.
This one does. It's organized the way an engineer organizes anything worth trusting: not by buzzword, but by function, meaning what your people actually do all day, and which slice of it a machine can take over cleanly. No "transformation journey," no roadmap slide with a rocket ship on it. Just the territory, and where to plant your flag first.
What is AI actually good for in a company that isn't software?
AI earns its keep on high-volume work that needs judgment a rulebook can't fully write down: reading documents that never arrive the same way twice, forecasting from your own operating data, flagging the transaction that merely looks wrong, and making institutional know-how searchable instead of tribal. None of that requires becoming a tech company. It attaches to the operation you already run.
McKinsey's latest global survey found that 88% of organizations now use AI in at least one function, up from 78% a year earlier, and yet only about 6% qualify as "AI high performers," seeing more than 5% EBIT impact from it. The gap between those two numbers is companies that adopted broadly and shallowly instead of picking one territory and committing to it, exactly the mistake this map is built to prevent. The value sits in a small number of well-chosen territories, not spread evenly across the org chart like fertilizer.
Where's the highest-value territory on the map?
Two axes decide it: how often the work happens, and how much judgment it requires.
- High volume, low judgment. Already solved. A rule fully describes the work, so it's automation's job, not AI's.
- High volume, high judgment. Start here. This is AI's actual home turf: work that repeats constantly and still needs a judgment call.
- Low volume, either kind of judgment. Not worth automating. Keep the human; rare cases are exactly why you pay for experience.
If a process owner says "well, it depends" more than twice describing one workflow, and that workflow runs hundreds of times a month, you've found your first project. Optimizing anything else first is the business equivalent of tuning a function that runs once a year: technically satisfying, financially irrelevant.
What does this look like in practice?
The shape repeats everywhere: a retailer's demand forecast, a clinic's referral queue, a distributor's routing plan, a hotel group's month-end close. Same territory, different paperwork. Here's what it looks like in the two industries we know best, insurance and hospitality:
- Turning messy documents into usable data. In insurance, that's broker loss-run PDFs standardized into structured underwriting data. In hospitality, it's PMS, RMS, POS and CRM exports reconciled into one portfolio view.
- Predictive triage on your own data. Insurers use it for underwriting risk scoring that routes files to the right desk. Hotel groups use it for multi-property reporting that replaces twenty spreadsheet formats with one number.
- Searchable institutional knowledge. For carriers, that's underwriting guidelines and policy wording, answerable in seconds. For hotel groups, it's brand standards and SOPs held consistent across every property.
Notice what's missing: no chatbot greeting your guests, no pricing engine. The map says the highest-value territory in both industries right now is internal: get the data and the judgment-heavy back-office work right before anything customer-facing gets near a model.
Do you need pristine data before you start?
No, and waiting for it is the single most expensive mistake on this map. A company-wide "data readiness" initiative before any AI project starts is Zeno's paradox with a budget line: you halve the distance to "ready" forever and arrive nowhere, on schedule.
What you actually need is enough history in one workflow: a year of loss runs, two years of portfolio data, a folder of contracts. The honest caveat: if a workflow's data lives only on paper or in one person's head, it isn't your first project. Pick one where the data already exists digitally, even if it's messy.
Should you buy a tool, bolt one on, or build custom?
Buy commodity, build the differentiator: the same rule as every other software decision you've made. Gartner predicts at least 30% of generative AI projects will be abandoned after proof-of-concept by the end of 2025, and the reasons it cites, poor data quality, inadequate risk controls, escalating costs, unclear business value, are planning failures, not evidence that AI doesn't work. Skip the coin flip on buy versus build (the map above already answers it) and you're simply not in that group.
Running an off-the-shelf model against a workflow it wasn't built for is a lot like running someone else's regex against your data: occasionally it matches. Mostly you're catching the exception by hand and calling it "AI-assisted."
What does a first project actually cost?
Directional, not a quote: anyone pricing this before scoping the workflow is quoting the invoice, not the project. Three things move the number, and "the AI" is rarely the biggest one:
- Integration. Wiring into the policy admin, PMS, ERP or EHR system you already run is usually the harder half of the build.
- Data readiness. Messy-but-digital is workable; paper archives and tribal knowledge add a digitization step.
- Human review. Anywhere a mistake is expensive, whether that's underwriting, medical, or financial, you're engineering audit trails and review steps. That's what makes the system trustworthy, and it's not optional.
MIT's Project NANDA found that 95% of enterprise generative AI pilots show no measurable P&L impact, and pinned the cause on the unglamorous stuff: workflows nobody scoped and outcomes nobody defined before the build started, not model quality. That's a map problem, not a technology problem, which is a strange kind of good news, since a map problem is the cheaper one to fix.
Where the work gets built matters too. A cross-border team spanning San Diego and Tijuana, working inside your business hours end to end, typically brings total project cost down 30 to 50% versus a build staffed entirely out of the US. That gap is often what separates a pilot that clears the bar from one that quietly doesn't.
How do you start without turning this into a science project?
- Pick one workflow with a number stapled to it. "Loss runs take three weeks to standardize" beats "we should have an AI strategy" every time.
- Write the definition of done before a line of code exists. Hours saved, days cut, error rate down: whatever the number is. Software engineers have called this "acceptance criteria" for decades; it works just as well on a pilot.
- Timebox it. Eight to twelve weeks, one workflow, real data. A pilot that needs a year isn't a pilot; it's a program with worse reporting.
- Ship it, measure it, let the result pick the next territory. Kill what didn't work. It's the cheapest lesson you'll buy all year.
We build these systems with teams across San Diego and Tijuana. Our AI development work covers the general approach, and if you're in insurance or hospitality specifically, our insurance and hospitality pages show what the map looks like once it's built. Or skip the reading and bring us the workflow. If the honest answer is "you don't need AI for this," we'll say so: it's the smaller invoice, and we'd rather earn the bigger one when you actually do.
Frequently asked questions
Will AI replace jobs on our team?
Mostly it absorbs the backlog work people already complain about: retyping, triaging, searching for the one document that matters, and moves them up to the exceptions and the customers. Headcount rarely drops at the pilot stage; capacity grows first. The teams that struggle are the ones that automated nothing and kept burning experts on data entry.
Is my company too small or too traditional for this?
Workflow volume matters more than company size or industry. A 40-person agency processing 500 files a month has a better first AI project than a 5,000-person company with nothing repetitive to point it at. If your team makes the same judgment call often enough to complain about it, you have AI-shaped work.
How is this different from the automation we already run?
Automation follows rules someone wrote; AI makes the call on inputs it hasn't seen before. We drew out the whole distinction, costs included, in AI vs. Automation, worth reading before you sign anything with "AI" in the pricing tier.
How long until we see results?
A well-scoped pilot shows measurable results in 8 to 12 weeks: one workflow, real data, a success bar you set before you started. Department-wide rollout typically follows the next quarter. If a proposal can't show value inside a quarter, the scope is too big; cut it down before you approve it.
