An AI readiness assessment answers one question: where will AI pay in this business, and what has to be true before it does? Done well, it ends in a short written plan: the work worth automating, ranked, with a number on each, an owner for the first one, and what happens in the next 90 days. Done badly, it ends in a maturity score across six pillars and a roadmap with no number on it.
Most of what you will find on this subject is the second kind. This guide is the first kind. It covers the checklist I use, why readiness belongs to a workload and not to a company, how to rank what you find, and what a good assessment should hand you at the end, whether you run it yourself or buy it.
Key takeaways
- Readiness is not a grade for the company. The same business can be ready for invoice matching and nowhere near ready for pricing. Assess work, not the organization.
- Start with the business outcome and the AI you are already running, not with a list of tools.
- Five questions decide each workload: where the data lives and who owns it, whether the process can be modeled, its volume and unit cost, where the exceptions go, and who owns it after go-live.
- Governance comes down to three documents with one owner: which models run on what data, what each may read, and a log of what it decided.
- The output should be a ranked plan with a number on each workload. If an assessment ends in a score, you paid for a description of the problem.
What an AI readiness assessment is, and what it is not
A typical readiness assessment scores the company from one to five on strategy, data, technology, talent, governance and culture, then recommends raising the low scores. It reads well. It rarely changes what anyone does on Monday, because a score does not tell you which work to start with, what it is worth or who will own it.
The useful version works the other way round. It starts with the recurring work in the business, asks what would have to be true for AI to take part of that work, and ranks the candidates by what they are worth and how likely they are to work. Readiness falls out of that as a list of specific gaps attached to specific work: this invoice flow is ready once one person owns the vendor master file; this forecasting idea is not ready because the history lives in three systems and two of them are not trusted.
| A maturity score | A ranked plan |
|---|---|
| Rates the company on five or six pillars | Rates each workload on data, volume, cost and ownership |
| Tells you where you are weak | Tells you where to start and what it is worth |
| “Improve data governance” | “Name an owner for the vendor file before week three” |
| Ends in a roadmap | Ends in one workload with an owner, a baseline and a date |
| Easy to buy, hard to act on | Harder to produce, easy to act on |
I do not start with a maturity assessment, and I would be wary of anyone whose main deliverable is one.
The gap a score misses shows up in the numbers. In McKinsey’s 2026 State of AI survey of 1,719 respondents, 80 percent say AI has improved their own productivity, but only 37 percent report any contribution to EBIT, about the same share as a year earlier. People are getting faster; the business is not yet getting better. Closing that gap happens one workload at a time.
Why readiness belongs to a workload, not a company
“Are we ready for AI?” has no useful answer at the company level. A company is ready for some work and not for other work. Customer email triage might be ready this month. Demand forecasting might need a year of clean data first. A company score averages those two into a number that describes neither.
There is a second reason. The things that block AI are rarely technical in the abstract. They are local: a process that lives in one person’s head, a field that three teams define three ways, an approval step nobody owns. You only find those by looking at a specific piece of work. A useful test before you deploy anything: is this process clean enough for AI to run without a person quietly fixing it along the way? If the honest answer is no, that is the first finding, and it is a process finding, not an AI one.
The companies getting real money out of AI work this way. In the same McKinsey survey, the small group of high performers (about 6 percent of respondents, who attribute at least 5 percent of EBIT to AI) are far more likely to redesign the work itself: nearly three quarters say they fundamentally redesigned workflows because of AI, against one quarter of everyone else.
The AI readiness checklist
Nine sections. Run them in order; the early ones shape the later ones. Each has a few questions and the answer that should worry you.
1. The business outcome comes first
- What would be different in twelve months if this worked? Revenue, cost, cycle time, risk: pick the measure first.
- Name one decision from last quarter that would have gone differently with better data or AI. If no one can, start there.
- Can each executive say what the other departments are working on this month? If not, AI will run into that gap before it runs into any technical one.
Worrying answer: “We need an AI strategy because the board asked.” That is a reason to start, not a goal.
2. What AI is already running, and what it already costs
- List the AI tools, licenses and features in use today, including the ones inside software you already pay for. Pull the list from finance invoices, not from memory.
- Ask whether anyone can launch an AI project today without approval. Often the answer is yes.
- Find out what people use on their own. When the approved tools are slow to arrive, staff use personal accounts, and company data goes with them. In the 2024 Work Trend Index from Microsoft and LinkedIn, a survey of 31,000 people, 78 percent of AI users said they bring their own AI tools to work.
- Note who owns the spend. Spend with no owner is the first number on the plan.
Worrying answer: “We blocked public AI tools, so we don’t have that problem.” You have it; you cannot see it.
3. The workload sweep: five questions per workload
List the recurring processes that carry real volume: invoice handling, intake, reconciliations, quoting, support tickets, weekly reports rebuilt by hand. For each one, answer five questions.
- Where does the data live, and who owns it? One system or four? One named owner or a committee?
- Can the process be modeled? Is it written down, with rules a new hire could follow, or does it depend on someone’s judgment that was never captured?
- What is the volume, and what does one unit cost today? Hours per week times loaded cost. No volume, no business case.
- Where do the exceptions go? Processes have cases that do not fit. If there is no named person to send them to, the automation fails the first time one arrives.
- Who owns it after go-live? Not the vendor, not the project team. A person inside the business whose job includes it.
Worrying answer: a workload where the data owner, the process owner and the post-launch owner are three different people, or nobody.
4. Data: ready for this workload, not ready in general
- For each field the workload depends on, can someone say where it comes from and when it was last refreshed?
- Of the documents and data you would feed it, how much is still true today? Old price lists, renamed products and retired processes are worse than missing data, because people trust them.
- Do not wait for all of your data to be clean. Clean the data this workload needs. Getting data in order is not the step before the work; it is the work.
Worrying answer: “We’ll start once the data warehouse project is finished.” That project will not finish on your timeline.
5. Systems: can it plug in safely?
- Does the AI see only what the person using it is allowed to see? If it reads around your existing permissions, you have created a new security hole.
- What happens when a connected system changes a field or a format? The usual failure is not an error message; it is a confident wrong answer.
- Can your own people maintain the connection, or does it only work because the vendor’s engineer remembers a workaround?
- Was it tested on your real, messy data, or on a clean sample? A demo on a sample export proves little.
- Treat the AI as a dependency, like any other system you rely on. How would you know an answer was wrong? And what can the model reach? Your existing monitoring answers neither.
6. Governance: three documents and one owner
Before a model touches customer, financial or employee data, one person owns three things:
- Which models are running, and on what data.
- What each one is allowed to read and do. Which actions it can only recommend, which it can take, and which need a person’s approval first.
- A log of what it decided. So when something goes wrong, you can see why.
Add one rule: any decision that could end up in front of a court or a regulator, such as hiring, firing or credit, has a named human decision-maker. “The AI did it” is not a governance model.
Worrying answer: a forty-page policy with no owner. A one-page policy with an owner beats it.
Few companies are there yet, especially for agents that act on their own. In Deloitte’s 2026 State of AI in the Enterprise survey of 3,235 leaders, only 21 percent said their organization has a mature governance model for agentic AI.
7. People: owners, managers and the time that comes back
- Does each workload have an owner who will be measured on it?
- Do the managers of the affected teams use the tools themselves? Teams follow what their manager does, not the memo.
- What happens to the time AI frees up? Decide before you start: more volume, better service, harder work the team never had time for, or fewer hires. People can tell which one you picked, and adoption follows. This is the step that gets skipped. In BCG’s 2026 AI at Work survey of 11,749 workers, 42 percent of regular frontline users said AI saves them at least a full workday a week, and 66 percent said they get limited or no guidance on what to do with that time.
- Have you told people how it changes their jobs? If you do not say it, they will assume the worst version.
8. From pilot to production: what proof looks like
AI projects rarely fail in the pilot. They stall between the pilot and production. A pilot runs on curated data with someone watching; production has neither. Before you call anything ready:
- Build a case file of twenty or thirty real requests from last month, with the right answer for each and a written list of what the system must never do without a person signing off.
- Run it in shadow mode: the system and a person do the same work for a couple of weeks, and you compare.
- Go live on one workload at a set share of volume, with an exception path to a named person and a baseline measured before the switch is thrown.
- Measure the share of volume that leaves the human queue, times the unit cost, against the baseline. Logins are not adoption.
A workload in production at a measured share of volume, with an owner, beats a pilot that handled a demo. Measurement is one of the habits that separates the companies getting a return: McKinsey’s high performers are twice as likely as others to have defined processes to measure the impact of their AI work.
9. What it costs to run
- What does one completed task cost, all in? Pilots are priced on demo volume. Agentic work, where the system reasons over documents, calls tools and checks itself, can burn many times the tokens the demo did.
- If this workload runs at ten times today’s volume, what is the bill, and who owns it?
- Are simple tasks going to the most expensive model by default? Routing routine work to a smaller model is often the fastest saving available.
Cost is already biting. One in five respondents in the McKinsey survey say their organization is limiting AI use because of operating costs, token costs included. And Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls: the three things sections 3, 6 and 9 of this checklist are built to catch.
One enterprise I worked with went all in with a frontier model vendor, and the tokens burned with no proof of value. When we mapped the workloads honestly, open models were plenty smart for a large share of the tasks, which cut that spend by 80 percent, and the speed-critical paths moved to faster inference and came back roughly 18 times faster. Same outcomes, a fraction of the bill.
How to rank what you find
Once the sweep is done, rank the workloads on two things: what each is worth, and how confident you are it will work. Worth is volume times unit cost, plus any revenue or risk effect you can state. Confidence comes from sections 3 to 6: data with an owner, a process that can be modeled, an exception path, a system it can plug into.
Across the research behind this guide, the same workloads tend to land in the same places. Your own sweep decides; treat this as the usual pattern, not a verdict.
| Workload | Why it tends to rank where it does | Usual call |
|---|---|---|
| Invoice processing and matching | High volume, standard formats, a clear right answer | Often first |
| Reconciliations and the close | Rules-based and recurring; the exceptions are known | Early |
| Document intake and data entry | Forms, onboarding, the same fields pulled every time; easy to check | Early |
| Support ticket sorting and routing | High volume, defined categories; leaves people with the hardest cases | Early, with a plan for the hard cases |
| Weekly reports rebuilt by hand | Recurring, low risk, quick to prove | A good second |
| Collections prioritization, contract review | Less visible, often sitting on cleaner data than people expect | Worth checking |
| Demand forecasting, pricing | High value, but only as good as current, trusted data | Not until the data has an owner |
| Customer-facing chatbot, proposal-writing pilot | Impressive in a demo; the return is usually marginal | Rarely first |
One shortcut when two candidates look equal: the workflow that sends the most escalations to leadership is often the one worth most, because escalations mean the manual process cannot keep up.
The most visible idea is usually not the best one. The chatbot for the board slide gets funded while the boring, high-volume flows that carry the margin sit untouched. Two warnings on the ranking. First, an easy win can move the bottleneck instead of removing it: if AI takes the simple tickets, people are left with only the hardest ones, so ask where the work will pile up next. Second, some inefficient-looking steps are doing a job, such as a junior learning the business by doing the routine work. Know which ones before you remove them.
What a good assessment hands you at the end
Whether you run it yourself or hire someone, the output should include:
- The workloads, ranked, with a number on each.
- Build, buy or skip for each, and which vendors, if any.
- The first workload with an owner, a baseline and a date.
- What AI already costs you, and who owns that spend.
- The three governance documents, or the plan to write them, with an owner.
- A 90-day plan, and what the board hears.
If what you receive is a score, a heat map and a recommendation to “build an AI culture”, ask for the list above.
Should you run it yourself?
You can run most of this checklist yourself. Sections 1, 2 and 3 need nothing but honest answers and a few days of the right people’s time. Bring in outside help when the leadership team cannot agree what the outcome is, when no one internal has taken a workload to production before, or when the board needs an answer it will trust from someone without a stake in the result.
For reference, here is how mine runs. A small business, under 50 people, takes two weeks, remote, from $3,500. A mid-market company, 50 people up to $1B in revenue, or any PE-backed company, takes four weeks at $15,000 per company. An enterprise is quoted after a scoping call. The fee credits in full toward an Adoption Sprint that starts within 90 days. The detail is on the AI consulting page.
The 90 days after the assessment
- Weeks 1 to 2: confirm the first workload. Owner named, baseline measured, case file built.
- Weeks 3 to 10: take it to production, not to a pilot. Shadow mode, then live at a set share of volume with an exception path.
- Weeks 11 to 12: report two numbers. The share of volume off the human queue and what it saved against the baseline. Then pick the second workload.
Boards forgive “not yet” when it comes with a plan. The board-facing version of this is in the AI answer your board actually wants. If an AI program has already stalled, start with restarting a stalled digital transformation.
AI readiness is a people and process question
Most companies treat AI as a technology project. It is a people and process project with new tools in it. Chief executives are starting to say so: in IBM’s 2026 CEO study of more than 2,000 chief executives, 83 percent said AI success depends more on people’s adoption than on the technology itself. The same study found only 25 percent of workers use AI regularly in their jobs. The checklist above is mostly about owners, decisions and habits, because that is where the work stalls. The model is the smallest part.
Frequently asked questions
What is an AI readiness assessment?
A structured look at where AI will pay in your business and what has to be true first. A good one ends in a ranked list of workloads with a number on each, an owner for the first one and a 90-day plan.
What does an AI readiness assessment cover?
The business outcome, the AI you already run and pay for, a sweep of recurring workloads, data and systems for each, governance, people and ownership, the proof needed for production, and running cost.
What is the difference between AI readiness and AI maturity?
A maturity model places the whole company on a ladder of stages. Readiness, done properly, is narrower and more useful: is this specific work ready for AI, and if not, what is missing?
How long does an AI readiness assessment take?
Two to four weeks for most companies. Mine takes two weeks for a small business and four for a mid-market or PE-backed company. Longer than that, and it is usually a strategy project wearing another name.
How much does an AI readiness assessment cost?
It depends on the size of the business and who runs it. Mine starts at $3,500 for a small business and is $15,000 per company for mid-market and PE-backed companies.
Can we do an AI readiness assessment ourselves?
Yes, for most of it. The outcome, the inventory and the workload sweep need honest answers more than outside expertise. Outside help earns its fee when the team cannot agree on the outcome, nobody has taken AI to production before, or the board wants an independent read.
Is our data ready for AI?
Ask it per workload, not for the company. For the work you want to automate: do you know where each field comes from, who owns it and when it was last updated, and is what you would feed the system still true today?
Who should be involved?
The CEO or a sponsor who owns the outcome, the leaders of the teams whose work is in scope, whoever owns the data, and someone from finance to put a cost on each unit of work.
What should we do after the assessment?
Take one workload to production within a quarter, with an owner and a baseline, and report the result in two numbers. Then the second.