How do you know if your company is ready for AI?

You know your company is ready for AI when six things are true at once: your data would survive being asked a direct question, your core processes run the same way regardless of who runs them, your people have both the appetite and the capability to use the tools, your systems can pass data between each other without a human in the middle, someone senior can state in one sentence what AI is supposed to change about the P&L, and a named person owns the outcome. If you cannot demonstrate all six, you are not ready. You may still be able to buy something useful, but you will not be able to prove it worked.

The good news is that every one of those six is testable in an afternoon, using people you already employ, without a consultant, a workshop or a discovery phase. Below are the six checks, one per pillar: Data, Process, People, Technology, Strategy, Governance. Each has a specific test with a specific pass condition. Run them before you sign anything. The reason to run them first is arithmetic. PwC’s 2026 CEO Survey found that 56% of CEOs report neither increased revenue nor decreased costs from AI in the last 12 months, and only 12% report achieving both. Those CEOs did not skip the AI. They skipped the six checks.

Check 1, Data: ask for your best 500 customers

Walk to whoever would produce it, and ask for a clean, accurate list of your 500 best customers. Not next week. Now.

What you are measuring is not the list. It is the answer you get before the list. If the response is “which system?”, you have a fragmentation problem. If it is “depends what you mean by best”, you have a definition problem, which is worse, because it means the business has never agreed what a good customer is. If it is “I’ll need to pull it from three places and clean it up”, you have found the manual export-and-import layer that quietly runs your company.

Pass condition: you have the list within a working day, one person produced it, and a second person independently agrees it is right.

This test is unglamorous and it is the one I would keep if I could only keep one. Every AI system you buy will read your data before it does anything else, and it will read the version that actually exists, not the version described in the tender document. There is more on this in dirty data and why it is the cheapest mistake to fix first.

Check 2, Process: ask three people to do the same task

Pick your most important repeatable process. Onboarding a customer, raising a quote, handling a complaint, closing a month. Ask three people to complete the same task independently, without discussing it, and compare what they did.

You are looking for divergence. Different steps, different tools, different fields filled in, different definitions of finished. Some divergence is normal and some of it is people quietly working around a broken system, which is worth knowing about on its own.

Pass condition: the three approaches are recognisably the same process, and where they differ, the difference is preference rather than substance.

Why this matters before you spend: automation copies the process you have. If three people run it three ways, there is no process to automate, only three habits. You will spend the pilot budget discovering that, and then you will spend the remediation budget fixing it, and by then someone will have written “the model underperformed” in a slide. It did not. There was nothing coherent to model.

Check 3, People: watch what happens when you hand someone a tool

Take one person, not your most enthusiastic one, and ask them to use an AI tool for a real task in their day-to-day work tomorrow. Then watch. Not the output, the behaviour. Do they know where to start? Do they check the answer? Do they know what they are not allowed to paste into it?

Then ask yourself the harder version of this question, which is about you. Are you using it? Talking about it? Have you made any formal investment in helping the team understand it, or has the entire capability plan been “people will pick it up”?

Pass condition: the person completes the task, and can tell you afterwards how they judged whether the answer was any good.

There is a version of this test that runs itself, whether you set it or not. PagerDuty’s 2026 Shadow AI Survey of 1,250 office professionals at companies with $500m or more in annual revenue found that two-thirds had used unauthorised AI tools at work, while only 18% of organisations report formal AI security policies. Your people have already answered the capability question. You just have not seen the results. Whether that is a real risk at 500 people depends less on headcount than on what they are pasting in.

Check 4, Technology: trace one number end to end

Pick a number that appears on a management report. Revenue by region, average handling time, stock position, whatever you look at monthly. Trace it backwards to its source, and count how many times a human being touches it on the way.

Every touch is a manual export, a spreadsheet, a reformat, a copy-paste. Each one is a place where AI cannot reach, because there is no path. A model in production needs a live feed, not a person who does the extract on the last Friday of the month.

Pass condition: fewer than two human touches between source system and report, and you can name the system of record for that number.

This is also the cheapest test to fail well. Finding out that your CRM and your finance system communicate exclusively through a woman called Denise and a shared drive is genuinely useful information, and it costs nothing to discover. It costs a great deal to discover in month four of an implementation. You do not need an engineering team to run this check, which is broadly the argument in AI without a tech team.

Check 5, Strategy: make someone state the baseline in one sentence

If you have anything running already, a pilot, a trial, a proof of concept, put the owner in a room and ask them to state, in one sentence, the baseline number the pilot is beating and where that number came from.

Not “it’s going well”. Not “the team love it”. A number, and a source for that number. If they cannot produce one, it is a demo. Demos are not worthless, they just do not survive a budget cycle, because there is nothing to defend when the CFO asks what it bought.

If you have nothing running yet, run the test forwards. Before you approve spend, require the requester to write the baseline down first. Same sentence, same rule.

Pass condition: a number, a source, and a date it was measured.

The scale of what this filters out is documented. MIT found that 95% of AI pilots deliver zero measurable P&L impact, and S&P Global Market Intelligence found the share of firms scrapping most of their AI initiatives rose to 42% in 2025 from 17% a year earlier, with the average organisation scrapping 46% of its proofs of concept before production. That is not a technology failure curve. It is what happens when nobody wrote down the number to beat. The pattern is set out in more detail in why AI pilots fail to scale.

Check 6, Governance: ask who gets fired

Ask five people in your leadership team, separately, one question: who owns AI outcomes here? Write down the five answers.

If you get five different names, nobody owns it. If you get one name and that person has never been told, nobody owns it. If you get “the exec team”, nobody owns it, because committees do not get performance-managed.

Pass condition: at least four of the five give the same name, and that person, when asked, agrees.

This is not a hypothetical failure mode at senior level. Deloitte’s 2026 Global CSO Survey found that 65% of chief strategy officers do not own the top strategic decisions in their own function. Diffuse ownership is the default state of a mid-size company, and AI does not fix it, it exposes it.

Governance also covers the boring half: who is allowed to use what, on which data, and what happens when someone breaks that. Writer and Workplace Intelligence found that 75% of executives admit their company’s AI strategy is ‘more for show’ than actual internal guidance, and 48% call AI adoption a massive disappointment, up from 34% last year. A strategy that provides no internal guidance is not a strategy. It is a position statement.

What to do with six answers

Run all six and you will have a rough map. Most companies I have worked with fail two or three, usually Data and Process, occasionally Governance in a way that surprises them.

The useful move is not to fix everything. It is to fix the pillars that block the first use case you actually want, and leave the rest until they matter. That prioritisation is the difference between readiness work that takes a quarter and readiness work that becomes its own permanent programme.

If you want the structured version, the free AI readiness audit asks 30 questions across the same six pillars, five each, and takes 7 minutes. It gives you a readiness score out of 120, a band for each pillar, and a short synopsis of where you sit. That is the whole free tier. It does not give you an action plan, it does not inspect your systems, and every answer is self-reported, so it is worth as much as your honesty when answering it. No sales call gets booked unless you ask for one.

What it does give you is a number you can put in front of a board that is asking what your AI position is, and a per-pillar band that tells you which of the six checks above to run first.