More capacity than they could hire and more capability than the market would sell them, on hardware they own, built one layer at a time.
Diagnostic → Systems Design → Capability Install
Ballpark, conservative end. On recursive recoding, the finding the team ranked first: a full corpus re-index in under fifteen minutes against days by hand. Extraction runs about thirty. These are engineering estimates against SPR’s own hours, not measured results.
Net over three years. What is left after those ballparks are cut by how good the evidence is, cut again for adoption, and zeroed entirely in the two phases where the human judgment lives. Payback in under three months.
The margin of safety. How much of that projected efficiency has to actually land for the build to pay for itself. Four fifths of it can fail to arrive and this still works.
Their hardware, in their building. No per seat charge. Every answer cites its source file, page and section. Settled before a line of the build was quoted.
Where the year goes today
Today, mechanical work takes 37% of analyst hours, the analytical middle 44%, and the expert end 19%. With the system, mechanical falls to 22%, the analytical middle to 25%, and the expert end rises to 53%.
Today the mechanical work takes more of the year than the expertise does, and the ninth phase, turning analysis into something a client can use, gets nothing at all. The system does not make anyone work faster. It moves the hours to the end of the chain, where SPR is actually being paid to think. Roughly a thousand of them a year, on this one evaluation.
SPR evaluates workforce and social programs. The research governments and foundations use to decide what actually worked. Economists, social scientists, political scientists. What they sell is judgment: reading evidence and saying what it means.
Then the ground moved under them. Federal cuts pulled money out of the sector, so everyone is bidding thinner. Clients started expecting lower prices, because everyone is using AI now, and wanting the bottom line faster with less appetite for nuance. Budgets cut, timelines compressed, and the same standard of evidence expected at the end of it.
Meanwhile the work still moves the way it has always moved. Material arrives in every format there is and somebody has to clean it before anything else can happen. One analyst budgeted forty hours to pull data out of a set of files and it took roughly a hundred and twenty. Their software is licensed by the seat, so sharing a seat does not share the work: the second person redoes the analysis and the effort doubles.
So the judgment was worth more than ever and had less room to happen in.
That is the shape of the ask when it arrived. A ten year evaluation was landing, four pillars, up to three hundred interviews, a decade of archived documents. The platform they had vetted for it had just lost its federal funding and pivoted out of research, and SPR had already told the client in writing that interview data would be handled on a private, secure, AI assisted platform. Could we help the team get up to speed on AI.
Nobody can tell you which tool to buy until somebody has described how the work actually moves. That is the part firms almost never have time to do for themselves, and it is what they hired us for.
Two working sessions with the people who do the work. Not a demo and not a vendor comparison. Nine phases came out of it. The first four are mechanical: cleaning, extraction, summarization, sorting. Necessary, and not what SPR is hired for. The last five are where the expertise turns into something a client pays for.
The team confirmed, without being prompted, that the same nine phases describe most of the research SPR does. Not just this project. That single sentence is what turned a project purchase into a firm level investment.
We evaluated about a dozen products against those nine phases. The pricing floor is real and it is structural: capability that will handle this data starts at $25K a year and is sold by the seat, sized for a company several times their size. Nothing we looked at covered more than four of the nine.
The better news was already inside the building. The tools SPR pays for today carry a configuration layer on top of them, skills and custom instructions and projects, and none of it is switched on. That is capability they have already bought and nobody has been given the hours to turn on. Standardising how four separate pillars summarise their material was the cheapest win in the whole engagement. No procurement, no approval, no incremental spend.
So a bigger licence would not have bought more capability. It would have bought more ceiling above a floor nobody was standing on.
We built the business case the boring way. The low end of every range, then a haircut by how good the evidence is, then an assumption that only seventy percent of the team actually changes how it works, then zero claimed in the two phases where the human judgment lives.
What survives all of that is about $379K net over three years, payback in under three months, and a break even at 19%.
We are not claiming the five to eight times faster that got said out loud in the room. After the discounts it is closer to one and a half. That is the number to use, and it is enough.
This is the part with no number on it, and we are not going to invent one.
The ninth phase, turning analysis into something a client can use, is not underfunded. It is unstaffed, because there is nothing left by the time you reach it. Emergent discovery, letting a corpus say something nobody thought to ask, has no tool support anywhere in the category, and it is the thing that separates a good evaluation from a competent one. There is work they decline because the capacity is not there, which never appears in any model of hours saved.
And the same system, given new instructions, drafts proposals from their own past proposals, writes apprenticeship standards, checks local policy against state and federal code. Each of those came up unprompted in the room. Each is instructions on a system they are already buying rather than another purchase.
We can prove the floor. What we are actually asking them to approve is the ceiling, and the honest reason it is unquantified is that it is wide open.
Hardware SPR owns, standing up in two to three weeks. No per seat charge and no usage metering, so five people and twenty five people cost the same, and the licence scarcity that currently doubles analytical work simply stops existing. Every answer traces back to its source file, page and section, which for a federally funded evaluation whose methodology will be examined is the difference between a finding the firm can stand behind and one it cannot. The data does not leave a machine in their building.
And it gets built in the order the pain is in. Ingestion first, because that is where forty hours became a hundred and twenty. Then summarization. Then the analytical phases. Each layer is working and in use before the next one starts, so capacity arrives along the way instead of all at once at the end.
Training runs alongside the build, not after it. As each layer is configured, the people who will use that layer learn it, so the team gains skill at the same rate the system gains capability. Not how to use somebody's software. How to instruct their own. The goal we agreed on is that SPR ends up writing its own instructions without calling us.
Nobody needed them to work faster. They needed room to do the part only they can do.
The diagnostic is closed and Phase One is agreed. The build is underway and nothing has run yet, so every efficiency figure here is a projection we discounted hard, not a result we measured. When there is a measured one, it goes on this page.
What is already true is not small. They have the map, which is the thing the market cannot sell them. They know what the answer costs and what it is worth. And the next thing they build on it is theirs, which is how a firm ends up able to take on work it has been turning down.
We were never hired to make a research firm faster. We were hired to give the judgment somewhere to happen.
Amoebaworks.
That’s what the Outside Read is for: a senior outside diagnosis of why growth stalled. One focused week, a clear diagnosis, and the first move.
$1K · One focused week See If the Outside Read FitsIdeas, teardowns, and field notes. The stuff we use to help companies make sharper decisions.
Occasional offers. Mostly stuff worth stealing.
SPR is an Install. Not a project handed back at the end, but a capability the company owns and runs after we leave.
See the InstallThe Bottleneck Wasn’t Marketing. It Was the Tax Code.
Sales Enablement → Capability Install → Content SystemHe Asked for Social Media. He Left With His Own Brokerage.
Positioning → Content System → PipelineA Calendar Full of People Who Weren’t Ready.
Category Design → Brand → Content SystemWhat it looked like, and what it was.
Twelve case studies