Four rails, ordered by how hard each is to reverse. The differences matter more than they look, because the rail decides both how much friction protects the instruction and which consumer rules apply to it.
A payroll deferral into a workplace retirement plan is the strongest version. The money is withheld before it reaches any account you can spend from, changing it usually requires a plan-portal election that takes a payroll cycle to land, and it is the only rail on which an employer may add money alongside yours. A split direct deposit is the same idea one step later: the employer's payroll system sends a stated share or dollar amount of each paycheck to a second account, so the balance you see is already net of saving. Most payroll systems support a split and most employees have never asked. A scheduled transfer at your own bank or credit union is the most common and the easiest to cancel, since it lives in the same app you use to check the balance. An app-based rule, such as rounding each card purchase up to the next dollar or sweeping a percentage of every deposit, is the weakest as a savings engine and the most useful as a starting point, because the amounts are small by design.
The federal rules that attach depend on who initiates the transfer, and one common arrangement falls outside them entirely. Regulation E, which governs consumer electronic fund transfers, expressly excludes at 12 CFR 1005.3(c)(5) any transfer the account-holding institution initiates without a specific request from the consumer between that consumer's own accounts at the same institution. So the standing checking-to-savings transfer inside one bank is not an electronic fund transfer at all, and none of Regulation E's error-resolution or stop-payment machinery reaches it. Where an outside institution or app pulls the money instead, the transfer is a preauthorized electronic fund transfer, and three protections attach: it may be authorized only by a writing signed or similarly authenticated by the consumer, with a copy provided to you (1005.10(b)); you may stop payment by notifying your own bank orally or in writing at least three business days before the scheduled date (1005.10(c)(1)), with an oral order lapsing after fourteen days if the bank required written confirmation and did not get it; and an error is subject to the investigation timetable in 1005.11. A narrow exemption at 1005.3(c)(7) removes preauthorized transfers at institutions holding $100 million or less in assets, so the protections are close to universal rather than universal. The practical reading is not that one arrangement is better, but that a transfer you can stop through your own bank and a transfer you can only stop by logging into someone else's app are different things.
Percentage or fixed dollars is the design choice with the largest long-run consequence. A fixed-dollar instruction is easier to reason about and goes stale by construction: every raise flows entirely to spendable income unless someone remembers to revisit the number. A percentage instruction absorbs part of each raise automatically, which is the same reason it feels worse in a month when income drops. Neither is correct in general, and the useful question is which failure you would rather have, an amount that quietly shrinks in real terms or one that follows income down as well as up.
What the evidence actually shows, stated carefully, because the famous numbers overstate it. Choi, Laibson, Cammarota, Lombardo and Beshears estimated in December 2024 that automatic enrollment raises steady-state saving by roughly 0.6% of income, and automatic escalation by roughly 0.2%, once employee turnover, unvested employer contributions, cash-outs at job change, and opt-outs are all counted. They also found only about 43% of participants accept the first scheduled escalation. Those are genuine effects from a very cheap intervention, and they are far below the figures usually quoted from earlier work, which measured participants who had opted in rather than a treatment effect across everyone exposed. The honest summary is that defaults reliably move behavior in the intended direction and that leakage gives a large share of the gain back, which argues for pairing automation with the boring question of what happens to the balance when you change jobs.
Where the money lands is part of the design, and labeling it has a documented cost as well as a benefit. Soman and Cheema found in the Journal of Marketing Research in 2011 that partitioning money into separate pots raises the amount saved. Sussman and O'Brien, in the same journal in 2016, found the other half: people will preserve a labeled savings balance while taking on high-interest debt for an expense the label was meant to cover, which is a strictly worse outcome than spending the labeled money. So named accounts are a good default and a bad rule, and the moment to override the label is when the alternative to using it is borrowing at a card rate.
Two mechanical failures worth designing around. Automating a transfer out of a savings account, rather than into one, can collide with the institution's own transaction limit. The Federal Reserve deleted Regulation D's six-transfer cap in 2020, but it permits rather than requires institutions to stop enforcing their own, so many account agreements still limit outbound transfers and charge an excess-withdrawal fee. And an automation that carries a subscription fee has to clear a hurdle the free rails do not: a monthly charge on a small round-up balance can exceed anything the balance earns, which makes the fee the first thing to check rather than the last.
