Graeme Crawford

CEO at Crawford McMillan | Helping PE firms protect and grow company valuations by telling the uncomfortable truth with clean, reliable data.

[Graeme Crawford @pedataguy]

@exec_sum Guess the headlights got a little brighter this week!

[Graeme Crawford @pedataguy]

@rohitdotmittal Merging operational systems takes forever if not planned well on a good foundations, and most roll-up plans assume it's done in a few months with marginal incremental effort from the teams.

[Graeme Crawford @pedataguy]

@FarmCreek Usually nobody checks whether the systems can handle it either. Third acquisition, third ERP, and finance is still reconciling the first two manually every month.

[Graeme Crawford @pedataguy]

$31M in missed billing. Found in 60 days.

Carson Marr reviewed every contract at a $1.2B company. Leakage usually runs 3 to 5% of revenue, before you touch a single fancy value creation play.

From the Private Equity Data Guy podcast 👇

[Graeme Crawford @pedataguy]

Everyone's talking about Nvidia's succession plan this week. Or the lack of one.

Portfolio companies have the same problem on a smaller scale. Usually it's the one person who knows how EBITDA actually gets calculated.

If your finance lead left Friday, could anyone rebuild Monday's numbers?

[Graeme Crawford @pedataguy]

Roll-up with three different board decks. The board wants AI to make one.

Bad idea.

AI on data that isn't aligned just compounds your mistakes. And it makes them look more credible, because the dashboards are so good.

Fix the definitions first. Then use AI.

[Graeme Crawford @pedataguy]

Every light on the dashboard is green. The CFO still doesn't trust it.

Vinay Raman (CEO, CAARMO) on what happens when your leadership team's numbers sit a standard deviation apart: decisions stall.

From the Private Equity Data Guy podcast 👇

[Graeme Crawford @pedataguy]

@sahilpatwa Agree. Pick one EBITDA definition for the platform and stick to it. Otherwise every add-on shows up with its own add-backs and the buyer at exit gets to pick them all apart.

[Graeme Crawford @pedataguy]

@BigJohn043 Yep. And the next fund mostly comes down to DPI, which means selling companies to buyers who believe the numbers. Harder than it sounds in the lower middle market.

[Graeme Crawford @pedataguy]

@gregisenberg I loved that episode. Got me the grok and have done a bunch of cool stuff already.

I never got the "why" with Grok @bot until watching @Rasmic rock it!!

[Graeme Crawford @pedataguy]

@IlliquidInsight Agree. A lot of it is just reconciling the board deck to the ERP. Same number, two answers, and an hour of the meeting spent on why.

[Graeme Crawford @pedataguy]

@jjacobs22 Good story. I'd love to know how they got 15 add-ons onto one set of books. That's usually where these fall apart, and where the controller's hair goes.

[Graeme Crawford @pedataguy]

@mardehaym The bit nobody budgets for is getting every add-on to agree on what revenue means. Skip that and the AI just gives you the wrong number quicker. Very confidently.

[Graeme Crawford @pedataguy]

Every KPI on the value creation plan is green. Has been for two years.

Enterprise value hasn't moved.

That's what happens when people set and mark their own homework.

Define the KPI. Govern the data. Then trust the dashboard.

[Graeme Crawford @pedataguy]

@WSJPE A $3.54B credit-data asset can still find a strategic in a tough exit market. Most mid-market portcos don't get that bid. The ones that do can rebuild every CIM number from the source systems. The rest sit in the backlog.

[Graeme Crawford @pedataguy]

@SMBDealGuy The $5k AI QoE will chew through the ledger in an afternoon. It will not notice that revenue, bookings, and bank deposits are three different numbers until a human has to defend one of them.

[Graeme Crawford @pedataguy]

@exec_sum Continuation funds reset the clock. They do not reset the data. The new LPs still inherit the same source systems, and they have gotten better at asking where the KPI actually lives.

[Graeme Crawford @pedataguy]

@rohitdotmittal The second close went in four weeks because the first process left a data room behind. The part that usually kills week 24 is not missing folders. It's when revenue in the CIM can't be rebuilt from the billing system.

[Graeme Crawford @pedataguy]

@mardehaym The deck described a model. The repo had a prompt. I'd open the customer and billing tables next. If those don't agree with the ARR on slide 3, the multiple was attached to a story, not a system.

[Mark Ajzenstadt @mardehaym]

A lot of PE-backed companies want to build an in-house AI team. They usually start by hiring one or two strong people.

A partner at a large advisory firm told us how that often plays out. Those hires end up carrying everything. They burn out, leave, and the company calls the outside firm back. The partner expects AI teams to follow the same pattern.

I think they're right. One person can't set the strategy, build the systems, review the work and cover the on-call rotation.

That's key-person risk. Hiring someone excellent can hide it for a while because the work keeps getting done.

If you're about to hire an AI lead, make sure the work can survive their departure:
1. Give them a team to build with from day one, whether it's internal or external.
2. Document what each agent does, what it can access, which decisions need approval and who owns it.
3. Keep the code in your repository and access under your company's accounts. Include prompts, workflow definitions and deployment configuration.
4. Plan the handover before the first build, so the work stays with the business when people move on.

Then cover the responsibilities around that person:

- Business ownership: someone decides which workflow matters and whether the result is useful.
- Technical review: someone other than the builder can challenge the design and inspect changes.
- Evaluation: representative test cases show whether the agent still behaves as expected after a change.
- Production support: someone watches failures, handles incidents and knows when to disable or roll back the workflow.

These don't all need to be separate hires. They do need named owners and enough time to do the work.

A runbook should let another person understand how to deploy the agent, investigate a failure and recover the service. If the only workable instruction is "ask the AI lead," the handover isn't ready.

For a portfolio company, I'd ask what happens if that person is unavailable during month-end or a customer incident. The answer tells you whether you've built a capability the business can rely on.

Make the hire. Just don't make one person responsible for holding the entire AI programme together.