The Glue Work

Vas on X, writing about AI agents in enterprise finance:

Your software handles big ticket items. It automated the core ledger, moved invoices into a digital workflow, gave everyone a dashboard. But between those systems, there are still hundreds of small tasks and workflows that require a human to copy a value from one screen, paste it into another, check whether two numbers match, send a follow-up email when they don't, escalate when nobody responds. Humans are still the glue holding these systems together.

This is the right framing. Most AI-in-finance talk fixates on the glamorous end - predictive analytics, autonomous FP&A, the mythical "finance brain." Meanwhile, someone on the AP team is spending Tuesday afternoon emailing a vendor about a missing W-9 for the third time.

The piece is full of sharp observations. The distinction between a dashboard that shows you 47 unresolved exceptions and an agent that resolves 40 of them is the kind of thing that should be obvious but apparently is not to most vendors. The "Brittany" example - where the real process involves tribal knowledge that no SOP captures - is worth the read alone. Anyone who has sat through an automation project that faithfully reproduced the documented workflow while ignoring how the work actually gets done will wince in recognition.

But. This is a sales piece. A good one, well-disguised behind genuine insight, but a sales piece nonetheless. "Don't buy a platform. Don't hire a data science team. Find someone who can sit with your team and build an agent inside your existing systems." The someone, naturally, is the author's company.

The claims are confident - "we brought month-end close from 12 days to 5 days" - and unsourced. No company named, no baseline maturity described, no mention of what else changed. The failure modes discussed are all conveniently the other approaches: chatbots, platform purchases, IT-led projects. The author's own model - embedded consultants building custom agents - gets no critical examination. What happens when those agents break after a system update? What's the ongoing maintenance cost? What if you're a GBP 30M company, not a $100M+ one? Silence.

The 87% pilot failure stat from Gartner is doing a lot of heavy lifting here. It's real, and it's damning, but it's also being used as a punchline that lands on everyone else's approach while exempting the author's.

None of this means the advice is wrong. "Map the real process, not the documented one" is good counsel. "Build agents that do the work, not dashboards that display it" is a useful filter. The five use cases are concrete and grounded. If you're a CFO, there's real value here - provided you read it as one perspective from a vendor with skin in the game, not as disinterested analysis.

That's the gap worth noting. There's plenty of AI content that hypes. There's plenty that sells. There's not nearly enough that helps a senior finance leader evaluate these claims independently - before they're on a sales call.