The pattern: pilot purgatory
Walk into almost any portfolio company with an AI line on the board deck and the story is the same. A pilot ran. Someone can describe a demo. Nobody can name the dollar figure it produced, because it never produced one. That state, not failure exactly and not success either, is pilot purgatory: a program that is technically alive and operationally inert.
The scale of it is not a portfolio company problem alone. More than 80% of AI programs still fail, usually due to misalignment on use cases, weak user adoption and unclear success measures, according to AlixPartners. That is the industry base rate. Portfolio companies do not have a special immunity to it; if anything the hold-period clock makes the consequences more expensive.
What is missing is rarely the model. It is the operating discipline around the model. Accordion's research into the PE AI gap keeps landing on the same finding: firms lack an operational playbook for turning a mandate into a shipped system. The mandate exists. The method to execute it does not. That gap is where all six failure modes below live.
Each mode is specific enough to spot in a single conversation with a function head, and specific enough to fix without waiting for the next platform close. Score them in order. The order matters, and the last one explains why.
1. The License Trap
The License Trap is a purchased seat with no user behind it. The company signed a contract for an AI tool, sometimes at the sponsor's suggestion, and the utilization never followed the purchase order.
How it shows up. A line item in the software budget for an AI tool with a seat count nobody can explain. A renewal conversation where the vendor rep knows more about usage than anyone inside the company. A tool mentioned in the board deck by name, never by a usage number.
What it costs. Direct cost is the wasted license spend, usually five to six figures a year for a mid-market function. The larger cost is reputational: the next AI request from that function gets a harder no, because the last one produced nothing measurable.
The first fix. Pull the usage report before the next renewal, not after. Cancel or downsize any license under 20% weekly active use. Move the freed budget to the function that already has an adoption owner named. Do not buy another seat until the first one is used.
Score it 0 to 3. 0: every purchased seat has a named, active user. 1: minor unused capacity, tracked. 2: a meaningful share of seats unused, no tracking. 3: licenses purchased with no usage report at all.
2. The Strategy Shelf
The Strategy Shelf is a roadmap document that reads well and ships nothing. A consultant or an internal team produced a well-structured plan, everyone nodded, and the document now lives on SharePoint with no owner and no date attached to any line.
How it shows up. A 40-slide AI strategy deck from eighteen months ago that nobody can find without asking three people. A workstream list with department names in the owner column instead of people. A plan that was "approved" but never entered anyone's calendar.
What it costs. The consulting fee is sunk regardless. The compounding cost is time: eighteen months of runway spent producing a document instead of a system, inside a hold period that does not pause for planning cycles.
The first fix. Take the single highest-impact workstream from the document and convert it into a 100-day plan with a named owner this week. Archive the rest. A strategy document earns its cost back only when one line of it becomes a shipped system; the rest is sunk cost, not a backlog.
Score it 0 to 3. 0: no unexecuted strategy document exists. 1: a plan exists, one item is in motion. 2: a plan exists, nothing has started in 90 days. 3: multiple strategy documents, none executed.
3. The Demo Graveyard
The Demo Graveyard is a working prototype that never reaches production. It ran well in a steering-committee meeting, on curated data, in front of an audience that wanted it to succeed, and it has not touched a real workflow since.
How it shows up. A tool with a URL nobody outside the original project team knows. A prototype built on a data export from six months ago. A confident answer to "does it work" and a vague one to "who used it this week."
What it costs. The build cost is sunk. The bigger cost is credibility: a function that watched one demo go nowhere is a harder sell on the next one, and the next request for budget gets read as more of the same.
The first fix. Pick the single demo closest to production and connect it to real data in the real system of record this month. If none is close, the honest fix is to shut the graveyard down publicly and start one new build with a maintainer named before day one.
Score it 0 to 3. 0: every prototype either reached production or was formally retired. 1: one live prototype, path to production defined. 2: multiple prototypes stalled, no path defined. 3: a pattern of demos abandoned after the first presentation.
4. The Workshop Certificate
The Workshop Certificate is training measured by attendance instead of usage. A vendor or consultant ran a session, employees received a certificate, and the operating review never asked what changed in anyone's week afterward.
How it shows up. A completion rate reported to the board instead of a usage rate. Employees who can describe the training but not a task they now do differently because of it. A trainer who left with the fee and no obligation tied to what happened next.
What it costs. The training budget itself, plus the opportunity cost of the hours spent in the room. The real cost is the false signal: a completion certificate looks like progress on a board slide and produces none of the outcome the board is actually tracking.
The first fix. Replace every completion metric with a usage metric: weekly active users over eligible users, tracked from week one after training, not attendance on day one of it. If a vendor cannot report usage after delivery, that is diagnostic information about the vendor.
Score it 0 to 3. 0: every training program is tied to a tracked usage outcome. 1: usage tracked for the newest program only. 2: attendance tracked, usage is not. 3: no training program has a usage metric attached.
5. Vendor Lock-in
Vendor Lock-in is a system the company cannot extend, cannot export data from and cannot renegotiate at renewal without starting over. It was signed for speed and now controls the roadmap.
How it shows up. A renewal quote that increased sharply with no corresponding increase in usage or capability. A request to export historical data that the vendor cannot or will not fulfill cleanly. A build-versus-buy decision that was never actually made, because a vendor got there first.
What it costs. Pricing power shifts to the vendor at exactly the moment the company has the least leverage: renewal, often timed near a board meeting or close to exit diligence. At exit, an un-auditable, non-portable system is a diligence flag, not an asset.
The first fix. Before the next renewal, get a written data-export path and a realistic estimate of switching cost. Bring both to the next build-versus-buy decision instead of accepting the renewal as the only option. Write exit terms into every new AI vendor contract from the start.
Score it 0 to 3. 0: every vendor contract has a clear export path and reviewed exit terms. 1: most do, one legacy contract does not. 2: several contracts lack export terms, unreviewed. 3: the company cannot describe its own data-export rights for its primary AI vendor.
6. No Adoption Owner
No Adoption Owner is the absence of one named person inside the function, with protected time and a line in their goals, accountable for whether the system gets used. Everything else on this list either causes this or is caused by it.
How it shows up. A steering committee instead of an individual. A system that reports to the CIO, who owns infrastructure, not the function head, who owns the workflow it is meant to change. A rollout with no answer to "who loses their job if adoption stays at zero."
What it costs. Everything above, plus the option value of every future rollout in that function. A function that has watched one AI initiative die from no ownership treats the next one with justified suspicion.
The first fix. Name one person before any system design work starts, not after. Write the usage number into their goals and their manager's operating review. If nobody in the function will take the role, that is the diagnostic finding: the function is not ready, and the honest plan says so.
Score it 0 to 3. 0: a named owner with protected time and the usage number in their goals. 1: an owner named, not yet in their goals. 2: a committee, no individual. 3: nobody accountable for adoption in the function.
| Failure mode | 0 | 1 | 2 | 3 |
|---|---|---|---|---|
| The License Trap | All seats used | Minor unused, tracked | Meaningful unused, untracked | No usage report exists |
| The Strategy Shelf | No stalled plan | One item in motion | Nothing started in 90 days | Multiple unexecuted plans |
| The Demo Graveyard | All shipped or retired | One live, path defined | Stalled, no path | Pattern of abandonment |
| The Workshop Certificate | Usage tracked always | Newest program only | Attendance only | No usage metric anywhere |
| Vendor Lock-in | Export path, reviewed | One legacy gap | Several untracked gaps | No known export rights |
| No Adoption Owner | Owner, in their goals | Owner, not yet in goals | Committee only | Nobody accountable |
Full interactive version with your total and band: Pilot-to-P&L Scorecard.
Why they compound
None of the six modes stays isolated. The Strategy Shelf produces the workstream that becomes a Demo Graveyard project when nobody assigns an owner to execute it. The Demo Graveyard produces the disillusioned function head who signs the next License Trap out of frustration, buying a tool instead of finishing the build. An unused license invites a renewal negotiated from a position of weakness, which is how Vendor Lock-in gets written into the next contract. A rollout with no Workshop Certificate discipline trains people on a system that then sits idle for lack of an owner, which is No Adoption Owner again, closing the loop.
A portfolio company scoring high on two or more modes is not twice as far from the P&L. It is closer to the top of the Acute band, because each active mode removes a check that would otherwise have caught one of the others. This is why the scorecard sums all six rather than reporting the worst one: the total tells you how much of the system's self-correction is still intact.
The one that decides the others: the CHRO argument
Score No Adoption Owner first, before the other five. A named owner with the usage number in their goals will not let a license sit idle, because it is their number that suffers. They will not accept a strategy document as the finish line, because their goals require a shipped system, not a deck. They will push a stalled demo into production because their review depends on it. They will treat a training session as a means to an adoption number, not an attendance sheet. They will negotiate exit terms into the next vendor contract because they are the one who will live with the renewal. This is workforce design, not technology, which is why the fix runs through a CHRO's discipline as much as a builder's.
How to use the score in a value-creation review
Score every candidate function before it enters the 100-day plan, not after the fact. Bring the total, out of 18, into the same review where the rest of the value-creation plan is discussed, and report it in the same cadence: monthly, alongside the other numbers the board already tracks. Re-score at day 100 against the same six modes and put both numbers on the same slide. A function that moves from Acute, 12 to 18, into On Track, 0 to 5, in one quarter is evidence worth including in an exit narrative. A function that has not moved in two quarters is a signal to change the owner, the scope or the system, not to renew the plan as written and hope the third quarter is different.
Run the score across the whole portfolio, not one company at a time, and the pattern usually repeats: the same two or three modes drive most of the total across every portfolio company, because they trace back to the same organizational gap. That repeatable pattern is what the 100-day AI value-creation plan is built to close, one function at a time.