The Compounding Cost of Delaying Your Decision

01The clock you cannot rewind

Accounting runs on a calendar. Tax work concentrates heavily in the annual filing season, close and reporting cycles repeat on a yearly rhythm, and for a large share of clients the firm does a deep, document heavy engagement exactly once a year.⁹ That cadence is the defining feature of the business, and it has a consequence that is easy to miss: the richest record a firm can create about a client, a complete, reviewed, start to finish engagement, is produced only once per year, per client.

For a machine learning system, that once a year engagement is a labeled example of the highest possible quality. It is real work, on a real client, checked by a professional, with a known correct outcome. It is exactly the kind of data that alignment depends on. And it is perishable in a specific sense: if this year’s engagement is not captured in a structured, learnable form, the firm does not get another chance at it until next year. You cannot go back and re-run a client’s prior season to recover the signal. The cadence sets a hard limit on how fast a firm can learn from its own book.

This is what makes delay different in accounting than in most industries. Elsewhere, data accumulates continuously, and a system that starts late can catch up quickly by processing a backlog. Here the supply of new, high quality examples is metered by the calendar. A year spent fragmented is not a year you make up later. It is a labeled example, per client, that never existed.

02Data is the moat, and it is time bound

A generation of machine learning research points to the same conclusion: at the frontier, data matters more than cleverness. In a now classic result, a simple learner given far more data outperformed more sophisticated algorithms trained on less, and the authors argued that scale of data, not the elegance of the model, drove the gains.²The point was later generalized under the title “The Unreasonable Effectiveness of Data”: for many hard problems, more and better data beats a better algorithm.¹ Modern scaling work formalized it further, showing that model performance improves as a smooth, predictable function of the data it is trained on, with test error falling as a power law in the volume of training data, and the steepest gains coming at the earliest, lowest data volumes.^3,11

For an accounting firm, the strategically important data is not the open web that every model already has. It is the firm’s own clients: their documents, their prior positions, the judgments a professional made and why. That proprietary history is the moat, and unlike generic web text it cannot be bought or scraped. It can only be earned, one annual engagement at a time. Which means the moat is not just built from data; it is built on a clock.

03Fragmentation is data loss

Most firms do not lose their data in a dramatic way. They lose it to fragmentation. The engagement letter lives in one tool, the source documents in another, the workpapers in a third, the messages with the client in a fourth, and the final positions in the tax software. Each system holds a sliver, none holds the whole, and no two describe the work the same way. The engagement happened, but it was never assembled into a single, structured, standardized record that a model could learn from.

This is a documented failure mode, not a matter of tidiness. Structured, standardized data is the essential infrastructure for trustworthy AI; inconsistent, scattered inputs do not merely lower accuracy, they make error impossible to measure.⁷ The machine learning literature is blunter still: the hardest, most expensive debt in a production AI system is not the model code but its data dependencies, the tangle of undeclared inputs and inconsistent pipelines that quietly erodes the system over time.⁴A firm spread across a dozen disconnected tools is accruing exactly that debt, and paying it as data it can never learn from. And the cost does not stay put. A landmark study of high stakes AI found that upstream data problems tend to trigger “data cascades,” compounding, delayed, downstream failures that are pervasive and largely invisible until they surface far from their source.¹² Fragmentation is the upstream data problem; the cascade is paid later, in the model that never gets good.

Defragmenting onto a single platform is what converts a yearly engagement into a retained asset. When the documents, the workpapers, the communications, and the outcome all live in one normalized estate, the engagement becomes a clean, labeled example the moment it is complete. The same year’s work that a fragmented firm loses, an integrated firm keeps, and compounds.

04Why a lost year never comes back

Put the cadence and the data moat together and the cost of waiting stops being linear. A firm that consolidates this year turns this season’s engagements into structured examples, and its system enters next season already tuned to its clients. A firm that waits does none of that. When it finally adopts, it starts where the first firm started a year earlier, except the first firm is no longer there. It has moved on to its own next cycle.

Figure 2 shows the shape of it. Capability on a firm’s own clients does not rise smoothly; it rises in steps, one per completed annual cycle, because that is how often the firm gets a new labeled example per client. A one year delay does not shift the curve down by a fixed amount that later effort erases. It shifts the entire staircase to the right by a full step, and since both firms then climb at the same annual rhythm, the gap between them persists. The shaded region is learning the delayed firm never recovers, because the cycles that would have produced it have already passed.

How large is the setback? The floor is one year, and it is structural: because the engagement recurs annually, a skipped cycle cannot be re-run until the next season, so a full year of the firm’s highest quality data simply never comes to exist. The realistic figure is larger than that. A firm that missed the cycle also enters its next engagement with a weaker model, so it captures less clean, well structured signal from that engagement, and it does not resume the leader’s curve one step back; it climbs more slowly. Because the learning curve is steepest at the earliest cycles, a single missed early cycle is worth far more than a late one, and upstream data gaps propagate as compounding downstream failures rather than a one time loss.^11,12 Our estimate, under that power law assumption, is that one skipped year commonly compounds into two to three years of lag before a firm reaches the capability it would otherwise have had. The delay is never shorter than a year, and it is rarely only a year.

05Models decay without consistent learning

There is a second reason consistency matters, and it cuts against standing still. The world a model was trained on does not hold still. Tax law is revised, thresholds and brackets are adjusted, client circumstances change, and the statistical relationship between inputs and correct outputs shifts underneath the system. This is the well studied problem of concept drift: a model trained on last year’s distribution degrades as this year’s diverges from it, and the only durable defense is a stream of fresh, labeled examples to adapt on.⁵ The effect has been measured directly: models trained on a fixed window degrade as time moves past it, even in high stakes domains.¹³In an annual business, that corrective stream is only as reliable as the firm’s ability to capture each season.

Learning systems are also subject to a related fragility: exposed only to old patterns, they lose the ability to handle what they no longer see, a phenomenon studied as catastrophic forgetting.⁶The practical reading for a firm is simple. A model kept current on this year’s standardized engagements stays sharp; a firm that cannot feed it, because its data is fragmented or a cycle went uncaptured, is not merely failing to improve. It is falling behind a moving target.

06Client specific data is the differentiator

Every firm’s AI can start from the same general foundation. What no competitor can copy is a firm’s own history with its own clients. The technique that turns capable models into trusted ones is alignment from expert feedback: real corrections, on real work, become the preference data that steers the system toward what a given firm’s professionals actually want.⁸Applied to a firm’s book, that feedback encodes how this firm handles this client, the positions it has taken, the patterns in its industry niche, the judgment calls its partners have made before.

That is why client specific data is more valuable now than it has ever been. As foundation models commoditize general capability, the durable advantage moves to the proprietary, well structured, and current record of a firm’s own engagements, reviewed by its own licensed professionals.¹⁰The firm that has been capturing that record, one clean annual cycle at a time, owns something its competitors cannot purchase and cannot backfill. The firm that stayed fragmented owns a pile of documents in a dozen tools.

07Position

Most technology decisions can wait a year at the cost of a year. This one cannot. The cadence of accounting means the raw material for an aligned, firm specific model, the complete annual engagement, is produced on a schedule the firm does not control, and a cycle that is not captured in structured form is gone. Fragmentation guarantees it is not captured. Delay guarantees the loss repeats.

Arrive’s view is that consolidating onto a single platform is not only an efficiency decision but a data decision, and a time sensitive one. Defragmenting now converts this year’s work into a retained, learnable asset and starts the compounding a year sooner than waiting does. The firms that begin capturing their cycles today will, a few seasons from now, run on a model tuned to their own clients that no later adopter can catch. Autonomy is earned over time, and in this field time is metered one year at a time.