Scaling Social Prescribing Without Losing What Made It Work

A pilot with 40 referrals a month and one link worker who knows every venue owner by name looks nothing like a service handling 2,000 referrals a month across a borough. The funder who liked the pilot’s results usually wants the second thing built out of the first, on the same budget per head, on the same timeline. That request is where most of the value gets lost, and it gets lost quietly enough that nobody notices until the outcome data comes back flat two years later.

This note is about the one mechanism that tends to break first: the quality of the match between person and activity. Everything else — referral pathways, data systems, workforce training — is solvable with money and time. Match quality is not, unless you design for it explicitly before you scale.

What the evidence says the mechanism actually is

A 2022 qualitative meta-synthesis in BMC Health Services Research is the most useful single piece of evidence here, because it looks at what participants themselves say the benefit is. People describe gains that go beyond having someone to talk to: restored participation, purpose, a role. The synthesis notes that structured, purposeful group activity appears to work better than contact alone. That is a claim about mechanism, not just outcome — it says the quality of the fit between a person and an activity is doing the work, not the raw fact of a referral being made.

That has a direct implication for scale. If the mechanism is fit, then a service that scales by processing more referrals through the same activities faster is not scaling the intervention. It is diluting it.

The broader trial evidence is thin enough that you should not lean on it to settle arguments with a funder. A 2021 systematic review in the International Journal of Environmental Research and Public Health found consistent gains in self-esteem and confidence across social prescribing studies, but flagged limited trial evidence and heavy heterogeneity between programmes — which makes it hard to say which version of “social prescribing” actually produced the result. A separate 2021 review in Perspectives in Public Health found all nine included studies reported positive impacts, and three found reductions in GP, A&E, or inpatient use. A 2025 systematic review protocol is blunter still: effectiveness for older adults remains unclear, and there is only one peer-reviewed randomised controlled trial in the entire field.

That last point matters for how you scale, not just whether you should. Almost everything you know about your own pilot’s success is observational. You do not have the trial infrastructure to prove which components caused the effect, so you cannot afford to assume the components that don’t look “essential” — the link worker’s local knowledge, the small caseload, the personal relationship with venue staff — are dispensable padding. In the absence of dismantling studies, treat every element of a working pilot as load-bearing until you have evidence otherwise.

The caseload cliff

Name the failure mode so you can watch for it: the caseload cliff. It is the point at which a link worker’s caseload rises past the level where they can hold real knowledge of both the person and the local options, and the work quietly converts from matching into referring. Referring is faster, feels productive, and produces very similar activity data — same number of “social prescriptions” issued per month. But it is not the same service.

You will not see the caseload cliff in your throughput numbers. You will see it, if you look, in match quality: how often a referral results in someone actually turning up more than twice, how often the activity chosen bears any relation to what the person said they wanted, and how often the link worker can name why this activity for this person. None of those show up in a standard commissioning dashboard unless you build them in deliberately.

The pressure toward the cliff is structural, not a failure of individual staff. When a funder asks for 5x referral volume on flat headcount, the arithmetic only balances if time-per-case falls. Something has to give, and the thing that gives first — because it is the least visible in the data you’re required to report — is exactly the fit mechanism the 2022 synthesis identifies as doing the work.

Evidence status of the common scaling claims

Claim	Evidence status
Social prescribing improves self-reported wellbeing and confidence	Reasonably consistent across studies, but heterogeneous programmes and mostly non-randomised designs
It reduces downstream health service use	Suggestive — some included studies found this, not all measured it
Structured, purposeful activity outperforms unstructured contact	Supported by qualitative synthesis; not yet tested in a dismantling trial
Standardised, high-volume delivery preserves the same outcomes as bespoke matching	No direct evidence either way — this is the actual scaling question, and it is unanswered
Volunteering and prosocial engagement reduce loneliness in older adults	Supported by one randomised trial (HEAL-HOA, Hong Kong), a rare RCT in a field mostly running uncontrolled evaluations
Structured befriending reduces loneliness scores in aged care	Supported by a 2025 randomised trial, though outperformed by structured psychological approaches in head-to-head comparison

The two randomised trials in that table — HEAL-HOA and the 2025 aged-care befriending trial — are worth reading together for a reason relevant to scale: both interventions were structured, not just “more contact,” and both were delivered by people trained to run that specific structure. Neither trial tested what happens when the same intervention is delivered by staff carrying triple the caseload. Nobody has run that trial. You are making that decision on judgement, not evidence, and you should say so to your funder rather than implying otherwise.

What this means in practice: before you scale referral volume, decide what ratio of caseload-to-match-quality you are protecting, write it down as a hard constraint, and report it to your funder alongside throughput. If volume targets and that ratio conflict, the volume target should lose. A service that hits its referral numbers and quietly stops matching well is not a bigger version of the pilot — it is a different, weaker programme wearing the same name.

The venue side of the scaling problem

Scaling also assumes the local supply of places to send people to is elastic. It usually isn’t. A 2025 study in Health & Place tracked twelve categories of third place — coffee shops, libraries, museums, recreation centres, restaurants — and found closures across every category between 2019 and 2021, concentrated in socially vulnerable and rural areas. Eric Klinenberg’s argument that libraries, parks, and similar shared spaces are social infrastructure with measurable effects on contact rates cuts the other way when you’re scaling a referral service into an area that has been losing exactly that infrastructure. If the venues your pilot relied on don’t exist in the neighbourhood you’re expanding into, standardising the referral pathway does nothing; there is nowhere to send the referral.

What this does not solve

This note does not tell you how many link workers you need per thousand referrals — nobody has that number, because nobody has run the dismantling trial. It does not solve the venue supply problem in areas where third places have already closed; that is a longer-term infrastructure question, not a programme design fix. And scaling referral capacity does nothing for the people who never get referred in the first place. Social prescribing, at any scale, still depends on someone — a GP, a link worker, a form — noticing a person and initiating contact. The hardest-to-reach lonely people are, by definition, the ones least likely to be in that pathway at all, and no amount of protecting match quality changes that.