Last updated: October 6, 2026
AI readiness is whether a firm’s important work survives without the colleague everyone asks where the data is. Readiness lives in the operation, in its records, its written rules, and its named owners, and no purchase order supplies it.
Page six of the approval memo had a table with four rows.
| Line | Amount | Who |
|---|---|---|
| Platform license, three years | $210,000 a year | Vendor |
| Implementation, twelve weeks | $95,000 | Vendor-led |
| Training | Two half-days | Vendor |
| Contingency | 10% | Finance |
There was no fifth row. I checked twice, then flipped to the appendix in case it had been moved there. It hadn’t.
The row I wanted would have priced the boring part. The pages the platform couldn’t read, and the analyst who’d clear them. The afternoon somebody would spend deciding which of two borrower files counted. Week thirteen, when the vendor’s implementation team rolled off to its next client and the thing became ours. The committee approved the memo in about forty minutes, which I remember because the meeting before it overran. Eleven months on, the work it was bought for was still done by hand. Same analyst. Same workbook, now on version 14.
I’ve seen that memo, give or take the vendor, at most of the firms I’ve worked with.
When a COO asks me if the firm is ready for AI, the honest translation is usually “Is it safe to sign this?” Fair question. Wrong unit. What I can actually check is whether the operation could carry the thing once it’s signed, and you find that out in the work, weeks before a vendor is in the room. KORE1, whose site this is, places the people who do that work, and its data engineering recruiting team fills most of the seats I get to near the end. The rest of this costs nothing. Some awkward meetings, maybe.
What AI Readiness Means
AI readiness is the condition of an operation in which the work an AI system would touch is already written down, runs off one trusted record, and has a named owner for every output and every exception. When those three hold, a model has something solid to sit on. When they don’t, the model inherits the gaps and hides them better than a person would.
Everything in that definition is about the work. Try it on your own shop. Ask where the inputs for the quarterly pack live and someone will answer quickly, probably correctly. Then ask who decides when the admin’s file and yours disagree. Pause. A first name. In plenty of firms one person is the database, and the whole office could name her. A tool aimed at that setup automates the part she’d already made easy and sends her the rest. She ends up busier. I watched it happen at a fund with three platforms and one analyst copied on every escalation email.
Record, rules, owners. The record means one canonical entry per borrower, policy, or claim, the thing I call a data foundation. Rules means somebody wrote down which file wins and why, somewhere other than their own head. Owners means a name beside every queue, including the unpleasant ones. You can buy around all three for a year or so. Then renewal.
The Purchase Order Is Easier to Approve
A purchase order has a price, a vendor, a start date, and a signature line. Operations work has none of those. It shows up as hours inside other people’s jobs, spread over months, with no invoice to point at when the board asks what the firm is doing about AI. Given a choice between a decision you can minute and one you can’t, a committee will minute the one it can. Every time. I’ve done it myself. A license can be approved, logged, and reported to the board by Friday, while the work of writing down which file wins and who clears the exceptions takes a quarter of somebody’s attention and produces nothing anyone can photograph for a board pack.
The pull to buy is real, and federal data shows it climbing. In the Census Bureau’s biweekly business survey, somewhere between 17% and 20% of U.S. businesses reported using AI in a business function from December 2025 through May 2026, and at the largest firms, 250 employees and up, the share hit 37%. Adoption is a fine thing to measure. It’s also a blunt one. It counts who bought or switched something on, and it can’t see whether anything downstream changed.
Economists measured that gap a long time before anyone said large language model. Erik Brynjolfsson, Lorin Hitt, and Shinkyu Yang studied 1,216 large U.S. firms across eleven years, 1987 to 1997, and reported in Brookings Papers on Economic Activity that each dollar of computer capital was associated with more than $10 of market value, against about a dollar for other tangible assets. Their reading was that most of that value came from intangibles, the business organization and work practices that grew up around the computers, rather than from the machines.
Ten to one. On organization.
Brynjolfsson came back to it with Daniel Rock and Chad Syverson in The Productivity J-Curve, published in 2021 in the American Economic Journal, Macroeconomics, which names AI as a general purpose technology that requires big complementary investments, new processes and human capital among them, mostly intangible and poorly measured. Poorly measured is the polite way of saying nobody puts them in the approval memo. Fifth row again.
I don’t need a regression to see it. I need to watch one quarter of the work. Usually one month is plenty.
Technology Question or Operations Question?
Readiness meetings mix two kinds of question, and the technology ones get answered first because they’re answerable by lunch. I started sorting them on a whiteboard a few years ago. This is roughly the board. Read the last column.
| Question asked in the meeting | Kind | What buying something changes |
|---|---|---|
| Which model, or which vendor? | Technology | Everything about the answer, very little about the result |
| Is our data in the cloud? | Technology | Where the mess is stored |
| Which of the two files is the current one? | Operations | Nothing, until someone writes the rule down |
| How often were we wrong last year, on our own deals? | Operations | Nothing. Most firms have never measured it |
| Who clears what the machine can’t read? | Operations | It adds a queue. Someone still has to own it |
| Could the close finish if your senior analyst were out for three weeks? | Operations | Nothing. The tool asks that analyst too |
| Who answers when the board, an LP, or an auditor asks what the system did? | Operations | A log, if someone set one up. At a carrier, a model inventory a committee can read |
Two rows against five. I’d put the real split of effort somewhere near that, maybe worse. On the cloud row, a short aside. Moving the workbook to SharePoint or S3 makes it reachable from more places, which is real progress, and it also means the wrong version opens from more places.
The calibration row gets skipped. Almost always. By calibration I mean the gap between what you underwrote and what actually emerged, measured on your own seasoned deals. If nobody has measured it, nobody knows whether the current process works, and a model trained on that process will copy its habits with great confidence. Bad ones included. Life insurers have run this comparison for decades, every year, as an experience study, and among the credit funds I’ve worked with I can count on one hand the ones that had run the equivalent on their own book before I asked.
Where AI Readiness Work Spends the Calendar
The model is the short part. On every build I’ve run.
Take the longest one, an underwriting engine for a specialty finance investor. Tens of thousands of regulated medical files a year came through that team, and people read every page of them. The engine that replaced the first pass now decomposes each record and drafts a mortality rating, a human approves it, and underwriting cycle time fell 33% at 99.9% decision accuracy. I’d love to credit the model. The months went elsewhere. We spent a long stretch just arguing about what a complete record was, because two senior underwriters had quietly been using different answers for years and neither knew. Then the rules, which took longer than anyone budgeted, since every underwriter had a check they did by feel and had never said out loud. Then the dull question of who signs when the machine and the person disagree, and where the file sits while they sort it out.
The plumbing beat it. Eleven data domains, one medallion build, and a person-matching method that had to find the same insured across hundreds of thousands of records, spelled four or five different ways. Data preparation time fell 84%. Infrastructure spend fell 35%. Nobody trained anything. I still find that a little funny, given how the budget conversations went at the start.
So that’s where the days actually go, into the record. If your plan gives the model eight weeks and the record two, flip it, and expect the vendor’s twelve-week timeline to be the first thing that cracks when you do.
Scorecard now, ML when your data can support it. I say it often enough that people finish the sentence for me. Most credit funds I’ve seen hold maybe five or six years of seasoned deals, some booked under different rules, and that’s thin for training anything I’d let near a credit decision. A plain scorecard with the rules written out holds up better, and every quarter it tells you how close the history is to earning a model.
A Ready Operation on an Ordinary Tuesday
The analyst who used to carry the close in her head is in Lisbon this week, and the close is running anyway, maybe a day behind, because the rule for which file wins sits in a shared document and the exception queue has three names on it instead of hers. A covenant figure comes back from extraction with a low-confidence flag and a reason attached. Somebody picks it up after lunch. They read the record, page 41 of the actual credit agreement, and clear it without sending a single “quick question” message. Last quarter someone lined up twenty seasoned deals, underwritten against emerged, and wrote the gap on one page. It was wider than anyone had guessed. Nobody enjoyed that meeting. That page now gets quoted more than anything the vendor produced.
None of that needed new software. Some firms already have most of it and have never called it AI readiness.
NIST’s AI RMF 1.0, published in January 2023, puts the same idea in a government register. Its GOVERN 2.1 asks that roles, responsibilities, and lines of communication for mapping, measuring, and managing AI risk be documented and clear to individuals and teams throughout the organization. Documented and clear. Most firms I meet could produce a license agreement in a minute. Very few could produce that page. Try it.
Who Owns AI Readiness
Operations. IT gets a vote.
Put it with whoever takes the call when month-end slips, the COO or CFO or head of portfolio operations, since that’s the person who would do the work if the tool fell over on a Friday. IT runs the systems well, in my experience. What it can’t do is rule on which borrower file is current, because nobody ever told it.
Below that owner I usually end up recommending the same handful of seats. Somebody has to build the canonical record and keep it honest, and that’s a data engineer’s job. Somebody has to keep the written rules from rotting, which is where a data governance analyst earns their keep, along with reporting on the exception queue so the owner can see it filling. And the first reviewer of what the machine produces? Nearly always a credit or operations analyst you already employ, someone who knows the work cold. Promote them into it.
Plenty of firms also need someone to set the order of work for a year or two, without hiring a full-time executive for it. My view, one I’ll defend, is that bringing in a fractional head of data gets further in a year than nine months of hunting for a permanent one. The searches I’ve seen stall for a familiar reason. Nobody inside can describe the job yet, so nobody can judge the candidates, and the firm ends the year exactly where it started with a recruiting bill attached. The ones that land usually had someone spend a quarter first writing down which records, queues, and decisions the role would own.
On timing, an IT seat takes KORE1 17 days to fill on average, and much of the build work fits a contract engagement while the record takes shape. Staffing is the easy half, though. Funding the fifth row is the hard one.
Two Tools for Counting It
Two tools on this site do the counting. The AI readiness scorecard walks one real decision through your operation and finds where it stops. The pilot-to-production checklist is for the moment before you sign, when a vendor contract is on the table and the questions above haven’t been answered in writing.
If month-end is where the pain sits, replay a closed one first. The close replay diagnostic sorts a finished month by what held each day up. Hand work shows up fast. Usually on day two.
Things COOs Raise When the Memo Comes Back
Is AI Readiness Just Data Readiness With a New Name?
Data readiness is one third of it, the record, and AI readiness adds the written rules and the named owners around that record.
A clean record with nobody accountable for it decays within a few quarters. I’ve watched it. A named owner with no record ends up as the database, which is the problem we started with.
Our Board Wants a Yes or No on AI Readiness. What Do I Give Them?
Give them one process, one number, and one name instead of a yes or no.
Something like this. Month-end close, 19 business days last quarter, owned by the head of portfolio operations, and it couldn’t finish without one analyst. That sentence tells a board more than a maturity score, and it’s a sentence you can improve on by the next meeting. Boards and LPs tend to come back to the same question each quarter, so give them something that moves.
Should IT or Operations Run the Readiness Assessment?
Operations runs it, with IT in the room.
Ask IT about connectors and permissions. Ask operations about the $340,000, the gap between your NAV and the administrator’s this month, because somebody in operations already knows roughly why and has been meaning to write it up since spring.
Realistically, How Long Before an Operation Is Ready?
For one process, about a quarter, and for the whole firm it’s ongoing work that never really finishes.
A quarter is enough to write the rules for one workflow, build its record, put names on its queues, and run it once without the key person. The firm-wide version is a habit more than a project. I’d rather see one process truly ready by spring than a roadmap covering twelve by next year. The firms that get there fastest are seldom the biggest spenders. They’re the ones where the owner of the process actually has a few hours a week protected for it and the authority to write a rule other teams have to follow.
We’re 40 People. Does Any of This Apply?
More than at a big firm, because at 40 people the one-person dependency is usually literal.
Small firms have less to rebuild, too. Much less. The record might be a handful of tables. The written rules might fit in ten pages. Start there. This month.
Will a Better Platform Get Us There Faster?
Rarely.
A better platform makes good operations faster and weak ones more expensive. If the tool you already bought isn’t being used, find out why before replacing it, because the reason is almost always one of the three things in the definition above and a new contract won’t touch any of them.
Add the Fifth Row
Get the approval memo for the last system your firm bought, the one with the cost table in it.
Under the last line, write a fifth one. Who runs the work in week thirteen, how many hours a week it costs them, and what they quietly stopped doing to make room. If the cell stays empty, you have your readiness answer, and it’s worth more than a score out of 100.
Send me the row once it’s written. I’m curious what goes in it. Message me on LinkedIn with it. If the name in that cell belongs to somebody you haven’t hired yet, KORE1’s recruiters can go and find them.
