What Shadow AI Can Reveal About the Firm

Part one described a researcher with two ChatGPT accounts. One for work, one she pays for herself, where the actual thinking happens. She won’t move it across because handing it over makes her easier to replace.

Part two argued that what sits in shadow AI falls into two tiers. The first is hygiene fixes: the workarounds people build because systems are broken, which most would hand over gladly if admitting them didn’t get them blamed. The second is judgement capital: the read she was hired for, which only moves if something is traded for it.

It also listed three things a firm gets back if it gets this right: fixed plumbing, better hiring and a clearer view of its blind spots.

This piece is about those blind spots.

A firm can see the results of its decisions, but rarely the reasoning that produced them. That’s how the same mistake keeps appearing in different teams without anyone recognising the common cause.

The failure loop

Take a pitch that went nowhere. On its own, a lost pitch tells a firm almost nothing. The client’s stated reason, budget, timing, ‘we went with someone else,’ is usually a softer cover for whatever the real reason was. The deck and the call transcripts capture what got said, but not why the team framed it as they did, or what alternative got dropped.

That reasoning, normally invisible, is often what sits in a personal AI account. Where someone talked themselves out of one angle. Where a colleague tested a framing late on and abandoned it without telling anyone.

None of it arrives on its own. It only shows up if someone chooses to share it, on the terms part two described. But put that reasoning next to the pitch timeline and the firm can see where the answer went vague, or where the prospect’s questions shifted and nobody adjusted.

Do this across enough failures, pitches that stalled, accounts that churned, launches that didn’t land, and a pattern shows up. The firm’s reasoning keeps defaulting to the same framework, whatever the situation called for.

That’s the blind spot. It isn’t a bad hire or a weak deck. It’s a habit of thought the firm shares and can’t see from the inside, because everyone in the room reasons roughly the same way.

Turning a blind spot into something that changes

Naming it is the easy half. A traced blind spot only matters if something happens next.

The response should be training, not a conversation about blame, and it has to come from the failure itself: which framework got used, what the situation needed, where nobody adjusted. A generic course can’t write that brief. It doesn’t know which three pitches this firm lost, or why.

Training like this can also be checked. Run it, then watch what happens in the next comparable cases. If teams default to the old framework again where it doesn’t fit, the training didn’t work, though some of those repeat cases will just mean the old framework was right. A different response in a similar situation is evidence it landed.

Either way, the firm finds out before the next pitch goes wrong rather than after.

One test before any of this works

Everything in these three pieces depends on the firm proving that disclosure is safe. Tier one has to work before anyone attempts tier two, because disclosure only turns honest once it’s safe. The failure can shape the training, but it can’t be used against the person whose work exposed it. If that breaks anywhere, none of the rest works.

So the test any firm can run right now is small. What happened the last time someone admitted a workaround (a manual patch, a broken handoff) they’d been fixing on their own? Did it get built into the system, or held against them?

If it’s the second, the rest of this doesn’t matter yet. No amount of consent, verification, payment, or targeted training convinces someone to hand over judgement capital to a firm that’s already shown what happens to people who disclose the easy stuff.

And the researcher keeps paying for her own account. Which stays the cheap option for her, and the expensive one for the firm.