PROVE AI / BLOG

From the Prove AI Team

Engineering depth, product thinking, and field notes from building the debugging layer for GenAI pipelines.

33 articles

Latest

A white architectural model of a house, cut away to show every room: empty, tidy and alike, except one narrow closet upstairs whose door stands open on a small skeleton.
OPINION
AI Agents Have Skeletons in the Closet. But Who’s Checking?

AI agents are getting more responsibility at work. But what happens when they leave problems behind that nobody sees coming?

Kelsi Kruszewski · Oct 30, 2026 · 4 min read
A lone figure in a dark suit, seen from behind, stands on a pale floor in front of a wall-sized spreadsheet: row after row of figures in identical cells, none of them standing out from the rest.
Finance series
FINANCE
What Does a Made-Up Number Look Like in a Workbook?

A wrong number usually looks wrong. A made-up one ties, fits its neighbors, and passes the check you run by eye. So what is that check telling you now?

Trent Fowler · Oct 27, 2026 · 3 min read
One person in a dark suit stands alone in a circle of spotlight on the floor of an empty open-plan office at night, reading a sheet of paper, while rows of unlit desks and monitors recede into the dark around them.
Finance series
FINANCE
What Changed That You Didn’t Ask For?

Reviewing AI-assisted edits to a financial model usually means checking the edit. The changes most likely to slip through are the ones nobody asked for, sitting where nobody thinks to look.

Trent Fowler · Oct 20, 2026 · 3 min read
A printed spreadsheet seen at a low angle under dappled window light — columns of figures run into the distance, and a neat column of teal check marks sits beside every row, each one ticked off.
Finance series
FINANCE
How Were You Taught to Check a Model?

Checking a financial model is something people were trained to do, long before AI arrived. The checks still hold up. What changed is something the discipline once took for granted.

Trent Fowler · Oct 13, 2026 · 3 min read
A wall of identical dark deposit boxes labelled A160, B160, C160 and onward, each with the same steel knob and keyhole, receding out of focus into shadow — nothing on any door says which one is worth opening first.
Finance series
FINANCE
Which of These Hundred Changes Should You Check First?

An AI model can return dozens of edits to a financial model in a single pass, and every one of them shows up at the same size. What runs out first in review is attention, and a flat list of changes gives you no help spending it.

Trent Fowler · Oct 6, 2026 · 4 min read
White sand pouring from a tipped glass jar into an hourglass on a pale blue surface — the upper bulb overflows and the excess spills down the sides into a growing pile around its base.
Finance series
FINANCE
Where Did the Time You Saved Go?

AI models can build a working financial model in the time it used to take to set up the tabs, and the hours saved are real. But they don’t seem to have come back, and I think it’s worth asking where they went.

Trent Fowler · Sep 29, 2026 · 7 min read
Two identical steel rulers, each marked 0 to 12, laid across one another in an X on a dark navy surface — two measuring instruments that agree with each other, which is not the same as either being right.
Finance series
FINANCE
What Is That AI Performance Number Actually Measuring?

Most AI performance numbers in finance are measurements of agreement between a model and a reviewer, presented as measurements of correctness. Four questions separate the two, and they apply as much to your own internal reporting as to anything arriving from outside.

Trent Fowler · Sep 22, 2026 · 6 min read
A creased “Manual Close Procedure” sheet, stamped Rev 01 and Controlled Copy, taped to a grey wall with its last-review date left blank, while the shadow of a person stands beside it, hand at their chin.
Finance series
FINANCE
When Did Anyone Last Run This By Hand?

In this series, every approach to AI in finance assumes somebody could still do the work manually, without the model, if the need should arise. That assumption is about an underlying capability, and finance already has a discipline for maintaining and testing exactly this kind of capability — it just isn’t one most firms have started applying to AI.

Trent Fowler · Sep 15, 2026 · 8 min read
A spiral staircase of worn blue-grey stone steps winding upward between a rough central column and a curved wall, lit from above — the same step repeated over and over before the path reaches anything.
OPINION
Your AI Agents Can Waste Millions of Tokens Before They Write a Single Bad Line

AI agents don’t just make mistakes — they can waste time, money, and context getting there.

Kelsi Kruszewski · Sep 11, 2026 · 5 min read
One handwritten number on a slip of paper beside three progressively larger printed documents that each restate it — a ledger line, a schedule, and a bound report — showing how a single input becomes load-bearing in several places before anyone checks it.
Finance series
FINANCE
Post-Hoc Confirmation: What Does It Cost to Check a Number That Is Already In Use?

Post-hoc confirmation is the AI-based, human-in-the-loop arrangement whose name is most misleading: by the time a human confirms the output of a finance workflow, it has already been posted, aggregated, and reported, and the person confirming it knows what changing it would cost. The check is real, and it happens at the moment it is most expensive to fail.

Trent Fowler · Sep 8, 2026 · 8 min read
A shattered grey stoneware vase pieced back together with pale seams, standing against a sky-blue backdrop with fragments still scattered beside it — reconstruction as the real work after a failure.
OPINION
We’re Scaling AI Agents Faster Than We Can Debug Them

AI agents are moving into production faster than teams can understand what happens when they go wrong.

Kelsi Kruszewski · Sep 4, 2026 · 4 min read
A grid of identical printed invoices with red hand-marked corrections clustered in a single corner rather than scattered evenly — the shape model errors actually take, and the shape a random sample is not built to find.
Finance series
FINANCE
Spot-Check Sampling: Is Your Sample Sized for the Errors You Actually Have?

Spot-check sampling is the only one of the four human-in-the-loop arrangements that says outright it isn’t looking at everything, which makes it easy to reason about and difficult to defend after the fact. The challenge isn’t that the sample is too small — it’s that sampling math doesn’t always fit well with the kind of error that models make.

Trent Fowler · Sep 1, 2026 · 8 min read
Low-angle view of a pristine white modernist structure curving against a pale blue sky — an elegantly engineered form whose cantilevered deck ends abruptly in mid-air, flawless right up to the break.
OPINION
AI Agents Don’t Have to Be Wrong to Break Production

The most dangerous AI failures aren’t always hallucinations. Sometimes the agent understands the task perfectly and still executes it incorrectly.

Kelsi Kruszewski · Aug 27, 2026 · 4 min read
A hand-drawn flow diagram on paper: an AI node splits a stream of red X and green check marks, routing a few items to “Exceptions (human review)” while the rest land under “Cleared (no human review)” — with red X’s visible among the cleared items nobody looks at.
Finance series
FINANCE
Exception-Only Escalation: When Is a Falling Escalation Rate Worth Worrying About?

The system clears whatever matches the rules, and a person sees only the exceptions. The judgment didn’t disappear — it moved upstream into a definition nobody owns, which is why a falling escalation rate looks like improvement whether the model got better or anomalies started clearing silently.

Trent Fowler · Aug 25, 2026 · 8 min read
A sphere woven from interlocking striped roads resting among looping racetrack curves on a pale blue surface — many separate threads tangled into a single knot.
OPINION
As AI Agents Multiply, We’re Losing the Thread

The challenge of agentic AI isn’t simply controlling what agents can do — it’s understanding how their decisions interact to produce an outcome.

Kelsi Kruszewski · Aug 20, 2026 · 5 min read
A carousel stand crowded with identical blue-grey rubber approval stamps on a dark navy background — every item gets the same stamp, which is exactly what makes an approval hard to read.
Finance series
FINANCE
Approval Before Action: The Arrangement That Is Easiest to Defend and Easiest to Hollow Out

A person handles every item, so coverage is always 1.0 — which is exactly why that number tells you almost nothing. The arrangement that’s easiest to defend to an examiner is also the easiest to hollow out without a single document changing.

Trent Fowler · Aug 18, 2026 · 7 min read
Macro photograph of a mechanical counter mid-roll, digits caught between positions behind a scratched window — a record in the act of being rewritten.
OPINION
AI Agents Can Rewrite the Story. Your Infrastructure Can’t.

The more autonomous AI becomes, the less we can rely on the system itself to tell us what happened.

Kelsi Kruszewski · Aug 13, 2026 · 5 min read
Letterpress print of four squares funneling along a single connector into one square below, captioned “Human in the Loop” — four different control arrangements converging on a single label.
Finance series
FINANCE
Four Arrangements, One Name: What “Human in the Loop” Actually Means

One compliance officer reads every alert; another reads a summary of what closed without her. Both arrangements are documented as human-in-the-loop, and the difference between them is an important question.

Trent Fowler · Aug 11, 2026 · 8 min read
A matte black cube resting on perfectly still, pale blue water beneath an empty sky — a system that gives away nothing about what is happening inside it.
OPINION
The AI Didn’t Surprise Me. The Silence Did.

The biggest takeaway from Anthropic’s latest disclosure wasn’t what the model did. It was how long it went unnoticed.

Kelsi Kruszewski · Aug 6, 2026 · 4 min read
A hand holding a pen, poised to sign a printed agreement on a dark desk — the signature that turns a document into a decision.
Finance series
FINANCE
How Financial Professionals Use AI, Part 3: Who Signs?

The tool produces, a person attests, and the systems remember — or they don’t. Part 3 closes the arc: why the signature, not the model, sets the pace for AI in finance.

Trent Fowler · Aug 4, 2026 · 9 min read
Seen from above, a lone figure snowshoes across unbroken snow, their tracks looping back on themselves — retracing the path is the only way to see where it went wrong.
OPINION
A New AI Benchmark Just Validated What’s Missing From Enterprise AI

AI is getting better at producing answers. The next challenge is helping enterprises understand where those answers came from, and what to do when they’re wrong.

Kelsi Kruszewski · Jul 30, 2026 · 3 min read
A stack of receipt-style credit memo review slips impaled on a steel spike — the top slip itemizes a ninety-second AI draft, hours of human verification, and a record marked NOT KEPT.
Finance series
FINANCE
How Financial Professionals Use AI, Part 2: What Does It Cost to Be Sure?

Ninety seconds to produce, hours to verify, and no record that the checking ever happened. Part 2 maps the two costs — checking and documenting the check — that decide what AI actually saves finance teams.

Trent Fowler · Jul 28, 2026 · 9 min read
The Prove AI logomark and wordmark in white over a low-angle view of Manhattan skyscrapers at dusk — the Frontier Lab’s New York City home.
ENGINEERING
Introducing the Frontier Lab

Prove AI’s new research and development team is tackling the problems that sit just beyond current tooling: decision provenance, runtime control for multi-agent systems, and human judgment in the loop.

Prove AI · Jul 27, 2026 · 3 min read
A folded printed bank statement standing open on a pale teal surface, columns of figures running down every panel — a paper record where every number can be traced.
OPINION
Hugging Face Exposed AI’s Next Infrastructure Problem

Financial teams don’t just need AI to produce answers. They need to understand how those answers — and every action behind them — came to be.

Kelsi Kruszewski · Jul 23, 2026 · 4 min read
A red pushpin planted on a dark financial data table, marking one glowing blue figure among thousands — the single number that needs a closer look.
Finance series
FINANCE
How Financial Professionals Use AI, Part 1: How It’s Actually Used Today (And How to Avoid the Most Common Pitfalls)

AI already earns its keep in finance on extraction, summarization, and first drafts. But the two dominant failure modes both produce output that looks correct — which is why everyone still checks every number.

Trent Fowler · Jul 21, 2026 · 10 min read
Rows of pale grey filing cabinets stretching in every direction, with a single drawer pulled open to reveal that it is empty — total recall, no answers.
OPINION
Two Headlines, One Trend: AI Doesn’t Need More Data, It Needs More Answers

Two very different stories this week — a security audit and a debugging technique — reveal the same shift.

Kelsi Kruszewski · Jul 17, 2026 · 3 min read
A lone climber silhouetted at dusk, clinging to the underside of a vast overhanging rock ceiling above the sea — agentic AI pressing against a reliability ceiling, not an intelligence ceiling.
OPINION
Agentic AI Isn’t Hitting an Intelligence Ceiling. It’s Hitting a Reliability Ceiling.

Meta’s slowdown says more about production than intelligence.

Kelsi Kruszewski · Jul 7, 2026 · 2 min read
A side-on view of an iceberg in open ocean: only a small white peak shows above the waterline while a vastly larger mass extends deep underwater — a metaphor for enterprise LLM bills, where the visible usage finance sees is just the tip and re-sent context is the hidden bulk below.
ENGINEERING
Why are enterprise LLM bills exploding?

Two mechanisms are driving enterprise LLM bills sky-high: unbounded usage (the loud one finance sees) and context bloat (the quiet one nobody’s measuring). One audit attributes 62% of agentic AI spend to re-sent context — and the model never even needs most of it.

Trent Fowler · Jun 10, 2026 · 7 min read
Diagram contrasting an open observability stack (AI agents → OpenTelemetry → Prove AI surfacing latency, cost, and safety signals) with a closed proprietary tool rendered as an opaque grid of question marks.
ENGINEERING
Is Open Source a Secure Foundation for Multi-Agent Systems?

The myth that proprietary code is inherently more secure than open source falls apart on inspection. Hidden blueprints aren’t a security posture — they’re a deferred bill, payable when the vendor changes terms, ships late, or leaves you with no path to a patch.

Trent Fowler · May 29, 2026 · 6 min read
Diagram: an orchestrator dispatches three parallel agents (A, B, C), each making an identical db.query() call. Two agents return record 42; one returns record 66. A downstream node asks, 'which is canonical?' — illustrating the coordination tax in a multi-agent workflow.
ENGINEERING
The coordination tax: why your multi-agent system is doing the same work twice

When agents in a multi-agent workflow quietly duplicate work or produce contradictory outputs — and nothing throws an exception — you’re paying the coordination tax.

Trent Fowler · May 27, 2026 · 8 min read
Cover image for article: The compound reliability problem: why your 95% agent is failing 40% of the time
ENGINEERING
The compound reliability problem: why your 95% agent is failing 40% of the time

Reliability multiplies instead of adding — so a chain of individually solid steps can still fail most of the time in production. The decay is a function of architecture, not model quality, and the right composition patterns bend it back.

Trent Fowler · May 5, 2026 · 6 min read
Cover image for article: Clean trace, wrong output: the visibility gap nobody talks about
ENGINEERING
Clean trace, wrong output: the visibility gap nobody talks about

Traditional observability catches execution failures — exceptions, timeouts, 5xxs. But an agent can execute flawlessly and still return the wrong answer, and that’s the gap standard tracing never sees.

Trent Fowler · May 5, 2026 · 4 min read