Ask a room of biotech founders whether adaptive designs save time and money, and most will nod; it is the received wisdom. Adaptive is modern, adaptive is lean, adaptive is what a fast-moving company reaches for instead of the plodding fixed three-phase march the traditional playbook prescribes. So you write "adaptive" into the protocol synopsis, and the board relaxes, the trade-off never priced.
Here is the number that should stop you.
Wilson and colleagues (2021, BMC Medicine) costed real adaptive trials against non-adaptive equivalents across seven UK Clinical Trials Units. Every adaptive design type they studied required more staff resource than its fixed counterpart. Not less. The worst case, sample-size re-estimation, ran a median 26.5% more staff FTE-years [1]. That is a measured, peer-reviewed finding, and it points the opposite way to the folk wisdom.
The problem is the word. "Adaptive design" reads like one lever trading statistical purity for speed, yet under the single label sits a whole family of quite different mechanisms: group-sequential stopping, blinded and unblinded sample-size re-estimation, response-adaptive randomisation, seamless Phase II/III, platform and master protocols. They trade different things in different directions. Some genuinely buy back time. Some charge you more staff, more calendar and more regulatory pre-negotiation than the fixed design they replaced. Lumping them together as "adaptive" is how the charge slips past unnoticed.
Regulators treat them as distinct. The FDA finalised its own adaptive-designs guidance in 2019 (announced in the Federal Register at 84 FR 65983) [2], and ICH E20, "Adaptive Designs for Clinical Trials", reached Step 2b on 25 June 2025 (still draft, still non-binding; FDA docket FDA-2025-D-3023). Section 2 states plainly that planning an adaptive trial "can be more complex and may require more time than for a trial without an adaptive design" [3]. The agencies are not promising you speed. They are warning you about complexity.
Call it the adaptive premium: the extra staff and calendar you pay the moment you make a trial adapt, whether or not the adaptation ever pays back. Every mechanism carries some. The only question is whether the one you chose earns it. So stop asking "should we go adaptive?" and start asking "which mechanism, priced against what a fixed design would have cost us?"
Start with the good news: there is a real one.
Group-sequential designs, pre-planned interim looks that let you stop early for efficacy or futility, are where the savings claim genuinely holds. Zhang and colleagues (2017, Clinical Trials) examined 52 completed Phase III oncology trials and showed well-chosen futility boundaries deliver real, quantified savings in both calendar time and the events you have to sit through [4]. Stop a dead arm at the interim and you spare yourself the back half of a trial that was only ever going to fail.
ACTT-1 (NCT04280705) is the cleanest case of the claim actually holding: remdesivir against placebo, 1,062 patients, a group-sequential design with a pre-registered blinded sample-size reassessment, and a primary readout on time to recovery inside roughly three months during a genuine emergency [5].
That said, read the caveat: it is the whole point. You only bank the saving if a boundary is actually crossed. A middling drug that crosses no efficacy line and trips no futility rule runs close to the full trial anyway, and you still paid for the interim looks and the independent monitoring that made them possible. Wilson's data caught this: group-sequential still ran a median 3.9% more FTE-years than fixed equivalents [1]. The smoke alarm only earns its keep if the building catches fire. You paid for the wiring regardless.
The operational read: reach for group-sequential when there is genuine early efficacy-or-futility uncertainty and enough accruing events to make an interim look mean something. That is the one that pays.
Now for the counterintuitive centre of the whole piece.
Sample-size re-estimation feels like the safe adaptation. You are unsure your effect-size assumption will hold, so you build in a look that lets you top up the sample if the drug turns out weaker than hoped. Prudent, surely.
It was the single most expensive design type in Wilson's data: a median 26.5% increase in staff FTE-years over a fixed equivalent, ranging up to 38.9%, the largest premium of any mechanism they costed. Across the whole study, statistician time rose most, by anywhere from 9.4% to 36.8% [1].
And that is before the statistical trap, which turns on a distinction the word hides: blinded versus unblinded. Blinded SSR re-estimates on nuisance parameters, things like the pooled variance or the control-arm event rate, without anyone seeing the treatment effect; it is cheap in Type I error terms and rarely controversial. Unblinded SSR, where the adjustment responds to the observed effect itself, is another animal. ICH E20 Section 2 puts the number on it: a Type I error rate that "can be more than doubled" against the nominal level if analysed with conventional, non-adaptive-aware methods [3]. A doubled false-positive rate is the sort of finding a regulator will unwind, not a rounding error you can wave through.
Know which SSR you are actually proposing before you write it down. Blinded re-estimation on nuisance parameters is defensible and cheap. Unblinded re-estimation buys the largest staffing premium in the family and a statistical liability you must pre-agree with the agency. Often the better move sits upstream: get the effect-size and standard-of-care assumptions right at the planning stage and you may not need to re-estimate at all.
Free download
The TPP Template
Target and minimally-acceptable profiles side by side, every claim linked to the evidence that has to support it.
Get the template →Response-adaptive randomisation, skewing allocation toward the arm that is winning as data accrue, is where the "modern must be better" instinct does the most damage.
The case against it is not made by traditionalists. It is made by the field's own methodologists. Korn and Freidlin, in JCO (2011) and again in JNCI (2017), concluded RAR "increases trial complexity and duration without offering substantial benefits to the patients in the trial" [6]. Mei and colleagues (2026, Statistical Methods in Medical Research) revisit the same under-explored problems from a different author group [7]. When the people who build these methods for a living tell you a mechanism adds duration without a payoff, listen.
There is one honest counter-figure, shown for exactly what it is. KENDO's registered protocol (NCT03227328) reports a simulated statistical power of 0.911 for its group-sequential response-adaptive design against 0.717 for complete randomisation at the same sample size [8]. That is a real advantage on paper, but it is a simulation from a protocol, not an observed result, and it rides on a group-sequential component doing much of the work.
Reach for RAR only when you can articulate a specific allocation or ethical rationale for putting more patients on the better arm within the trial. If the best justification you can muster is that it looks sophisticated, you are buying complexity and duration for nothing.
Seamless Phase II/III is sold as the elimination of dead time: collapse the white space between phases, run one continuous protocol, save the months lost to a database lock, a readout and a fresh start-up. Often it just relocates the cost upstream, into regulatory pre-negotiation and error control agreed before a single patient is dosed.
The EMA has watched this play out. Elsässer and colleagues (2014, Trials) surveyed 59 adaptive-design proposals brought for scientific advice: 20% were not accepted, 80% of the 41 more recent proposals examined in detail involved unblinded interim analyses, and CHMP itself cautioned that "the improvement in efficiency may be small when taking into account the time needed for the conduct of interim analyses and decision making" [9]. The survey is a decade old; newer sources point the same way, so treat the direction as current and the percentages as dated.
Many years ago I sat with a team that chose a seamless Phase II/III largely because it read well to a board that wanted something modern on the slide. The between-phase white space they were collapsing was maybe four months. The design negotiation with the agency, the pre-agreed interim, the multiplicity control, the firewall to keep the Phase II decision away from the sponsor team, swallowed the better part of a year before the first patient was dosed. They had paid the saving back, with interest, up front.
Before committing to seamless, put two numbers side by side: the calendar you will spend negotiating the design and its error control, and the white space you hope to save between phases. The second is frequently the smaller of the two.
Every unblinded adaptation needs an independent statistical and data-monitoring firewall, the machinery that lets someone look at accumulating data and act on it without that knowledge leaking back to the people running the trial. That machinery is not free, and it is not a one-off design decision.
The EMA's own Guideline on Data Monitoring Committees (EMEA/CHMP/EWP/5872/03 Corr, final, effective January 2006) is blunt about it. Section 2 notes that setting up a DMC and preparing its meetings "take some time (up to a few weeks)", and that "the use of a DMC might not be beneficial for the study but might even delay the finalisation of such a trial" [11]. The integrity infrastructure itself eats part of the time the adaptation was meant to save. You erect the scaffolding, and pay rent on it, before the building can save you anything.
And the firewall is a live task, never a solved one. The RECOVERY trial's data monitoring committee (Sandercock et al., 2022, Trials) recounts an independent DMC holding its line under sustained pressure from the UK regulator to unblind, and a genuine near-miss when a different platform trial, REMAP-CAP, publicly announced a 99.75% posterior probability of tocilizumab benefit while RECOVERY's own tocilizumab comparison was still recruiting [12]. The firewall held. But it held because a well-resourced, independent committee actively managed pressure coming partly from outside the trial entirely, the sort of pressure no single sponsor controls.
So the practical step is blunt: price the independent statistical and DMC infrastructure, calendar and headcount both, before you commit to any unblinded adaptation. For many resource-lean sponsors the plain finding is that they cannot stand up the firewall an unblinded design demands. That, and not the statistics, is the binding constraint.
You are counting FTEs and missing the prize, the argument goes. The real payoff was never staffing efficiency; it is killing dead assets early and lifting the probability a Phase III succeeds at all. Mahlich and colleagues (2021, Health Economics Review) modelled up to a 14.4% reduction in per-drug development cost, roughly $370 million, if adaptive designs lifted Phase III success rates from 63% to 80% [15].
That said, the Mahlich figure is a modelled scenario, conditioned on a success-rate jump nobody has observed and covering only Phase III attrition; the authors flag it themselves. It is an upper bound on a conditional, not a saving anyone has banked. (The other number that circulates, a Tufts CSDD estimate of $100–200 million a year, is over a decade old and I could not trace it to its primary source, so I will not lean on it [16].) And the platforms that delivered are national infrastructure answering pandemic- or population-scale questions, not a protocol a single-asset company clips on. The one figure that is both peer-reviewed and empirically measured for a resource-lean sponsor is Wilson's staffing premium, and it points at cost.
None of this is an argument that adaptive never pays. The payoff is mechanism-specific, and for most small sponsors it concentrates in one mechanism: group-sequential stopping. Like most things in clinical research, there is no such thing as a free lunch.
No rally. Just the checklist this whole piece reduces to. Run it before the word goes near the protocol.
Get that far honestly and the answer is usually one of two things: a group-sequential design that earns its premium, or a clean fixed design that was quietly the faster route all along.
Pricing that trade-off mechanism by mechanism, before it is locked into a protocol, is exactly what Inovia is built to pressure-test. The word "adaptive" will not tell you which one you have chosen. The mechanism will.
Get the monthly digest
The 5 things evidence leads need to know each month: regulatory moves, RWE developments and what they mean in practice. No pitch, one email a month.