Getting ready for AI - Is your business ready for AI, or willing to try it?

The short answer

AI readiness is not an item on a checklist you have to tick-off. It is an artifact of the specific process you are about to hand to AI. A business is ready for AI in a given workflow when it can answer three questions:  

• Does a wrong answer here cost money, trust, or compliance standing?
• Does the data behind this process already sit in a governed, structured system or siloed documents and DRI’s?
• Who is accountable for oversight of AI when it is confidently wrong?  

The answers to the questions need to be sought before the AI-Pilot and not while it is  about to be written off. Most of the businesses inside MIT's often-cited finding that 95 percent of AI pilots fail never asked them.

Three things worth knowing before you spend the next dollar on AI:

1. Readiness gets decided process by process, not company-wide, so the same business can be ready for AI in customer support and unready for it in pricing on the same day.

2. Partnering with a team that has already solved this kind of problem succeeds about twice as often as an internal team building the same system alone, according to MIT's research on enterprise AI deployments.

3. The companies inside MIT's 95 percent failure finding that beat it, many of them mid-sized rather than the largest enterprises, reach full production in about 90 days, close to a third of the timeline the biggest organizations need.

A CFO at a 200-person distributor sat through a flawless demo of an AI pricing tool last spring. It answered every question in the room correctly. Six weeks into production, it quoted a client 40 percent under cost, twice, before anyone noticed. Nothing was wrong with the model, but no one had asked what should happen the day it was wrong.

AI readiness, a few years ago, was about hiring a data scientist of picking the newest data model. Even recently as 2023, a version of the question meant something narrower: could the business afford a data scientist, and did it have enough labeled data to train a model from scratch? It has nothing to do with them today. That shift happened fast.  Large language models erased most of that barrier, where a business can now query a frontier model without hiring anyone. The bottleneck moved from access to judgment - from getting AI in the door to deciding which of the work behind that door AI should be trusted with.

“The 95 percent failure rate is not a bug. It is a structural inevitability,” says Anand Krishnan, thinkbridge's CEO, in a recent essay on his Meaningful Tech blog. “Businesses spend months evaluating whether a model is capable enough, then discover during rollout that capability was never the question. The real question is whether their process can absorb an AI that is occasionally, confidently wrong.”

Readiness, in that light, is a property of the task, not the company, and confusing the two is the biggest reason a business ends up inside MIT's 95 percent. Getting this right means sorting the actual work into two piles before any tool gets touched: work where a wrong answer is negligible in the bottom line, and work where it isn't. It means recognizing that most businesses do better building with a partner who has already solved a similar problem than trying to design and build a system alone, especially for anything that touches how the business competes. And it means proving all of this on a real process, with a real owner, before the AI line item shows up in next year's budget.

Readiness tied to the task leads to outcomes

Most AI readiness checklists ask company-level questions: clean data, executive buy-in, an AI strategy on paper. Those questions matter, but they miss the variable that predicts success, which is the specific process being handed to the AI, not the company handing it over.

Every workflow sits somewhere on two axis. The first is the cost of a wrong answer: negligible and recoverable, or expensive and hard to undo. The second is the kind of data the process runs on: unstructured material like documents, emails, and the knowledge in people's heads, or structured data that already lives in an ERP, a CRM, or a financial system.  

The combination of those two axis decides what AI should be allowed to do:

MIT's NANDA researchers reached a similar conclusion from a different direction. Their GenAI Divide report, built on interviews with more than 150 executives and an analysis of 300 public AI deployments, named the leading cause of stalled pilots as the learning gap: most tools never retain memory of a specific workflow, a specific client, or a specific exception, so they perform the same in month six as they did on day one. That gap barely matters when the mistake has a high cost, as it decides whether the pilot survives contact with a real client.

The practical test is simpler than what it may seem: ask where the answer to a customer's question currently lives. If it lives in a spreadsheet three people maintain by hand, or in the inbox of the longest surviving team member, that is unstructured data wearing a structured company's confidence. AI can help extract and organize it. It should not be trusted to act on it unsupervised until someone has done that extraction and checked it against reality at least once.

Partnership, not procurement: The real lesson in MIT's data

The MIT research’s numbers get interpreted as an argument for buying software off the shelf, which is far from what the underlying research measured. The finding groups two different paths into a single winning category: purchasing a tool from a specialist vendor who has solved for the specific task, and building a custom system in partnership with an outside team that has solved a similar problem before. This is the primary and best solution rather that the obvious: an internal team designing and building an AI system alone, encountering the problem for the first time. Solving a hard issue from zero costs more than most businesses budget for. But solving it alongside someone who has the technical know how does.
The clarity arrives when a business decides what is worth building versus buying outright. Here is a useful filter: would it bother you to learn a direct competitor is running the exact same system? For operations like HR, payroll, tax filing, or benefits administration, the honest answer is no. But for the process that are unique to the operational workflow of the business: whatever decides pricing, service levels, or how a customer gets treated, the answer is usually yes. That is the process worth owning outright, built around how the business operates rather than retrofitted to fit a vendor's template.  

Speed, scale, and governance

Speed follows the same pattern. The companies in MIT's research that move from pilot to full production fastest, about 90 days on average, are not working in silos. They pair an internal team that knows the business with technology-partners who have the engineering expertise and can skip the trial and error an internal team would otherwise cost in time and resources. Large enterprises average nine months or longer because of their scale with internal coordination and signoff. Having a strong partnership shrinks that problem, but doesn't remove it.

The 90-day production advantage disappears the moment a non-enterprise business borrows an enterprise buying process for it. Enterprise buying processes include a committee, year-long vendor shopping and pilotsjudged by a steering group that meets monthly. The advantage belongs to the businesses that pick the partner for the process, test the system against real work within weeks, and kill or rebuild it fast if it doesn't hold up. Owning the resulting system, is what turns that early speed into a lasting edge: the business controls the roadmap, keeps the data, and compounds every improvement instead of paying for the next version of the off-the shelf software.

None of this is happening in a vacuum of enthusiasm. 90 percent of employees already use personal AI tools for work even when their employer has no official subscription, which sits at closer to 40 percent of companies. The tools employees use every day are running ahead of the tools the business has sanctioned, which also indicates how the gap in readiness in most companies is tied to a governance gap.
The governance gap gets worse under agentic AI. Gartner estimates that of the thousands of vendors marketing agentic AI products, only around 130 are building systems with real autonomous capability. The rest are rebadged chatbots and automation scripts, what Gartner calls agent washing. And it also predicts that 40 percent of agentic AI projects will be canceled by the end of 2027 over cost, unclear value, or weak governance, not because the underlying models fail. If you couple that data with the MIT Sloan and BCG's late-2025 survey of more than 2,100 executives found agentic AI adoption climbing faster than any prior wave of business technology, already at 35 percent with another 44 percent planning to follow. But 47 percent of the same respondents admitted they have no strategy for what the technology should do. When the speed of adoption and clarity of purpose are moving in opposite directions, governance becomes an obvious issue. And the businesses that end up in Gartner's cancellation pile are the ones that let adoption run faster than clarity and governance.  

Proof before practice

Business should start, where the work is variance-tolerant and low-stakes. The real productivity gains from AI over the last has been automating the mundane: drafting, summarizing, retrieving documents, triaging inboxes. But when you use the same logic and confidence for the higher-stakes half of the business the risk gets higher for AI mistakes. The mistakes get billed, shipped, or filed.

Industry estimates put the global cost of AI hallucinations, confident but false outputs, at more than $67 billion in 2024. And a separate Deloitte survey found that 47 percent of enterprise AI users had made at least one major business decision based on content the AI had fabricated. Both numbers point to the same failure: a business skipped the step where a human, or a rule, checks the output before it reaches a customer, a regulator, or a P&L line.

This cost is what is known as the “verification tax”.  For productivity improvements, that overhead amounts to a quick scan.  But in high-stakes, business workflows that contribute to the bottom line the verification is the actual work. And if nobody accounted for it, the AI project will be a win on a slide but certainly a loss on the P&L six months later. The fix is a specific decision, made process by process, about who signs off before the AI's output leaves the building.

That decision is the actual readiness test: being ready for what happens when AI is confidently wrong, and having a human-in-the-loop whose job it is to catch it.

When you are ready

AI readiness measures how well a business knows itself: which of its processes can absorb a mistake and which can't. The 95 percent of pilots that stall inside MIT's research didn't fail because the underlying models were bad. Someone skipped that question and found out the hard way, on a live process, with a real customer watching.

The way forward is to sort the work first. Buy the narrow tool before building the general one. Ensure the human-in-the-loop process checks the output before it leaves the building. Businesses that do these three things reach full production in about 90 days, the same pace MIT found among the mid-market companies that cross the divide.

Readiness, in the end, comes down to awareness: know which of your processes can survive AI being wrong, before the AI finds out for you. If the answer isn't clear yet for the process that is about to be automated, that'sthe assessment worth running before the next AI budget line.