The Translation Problem
Why Justice Technology Implementations Fail and What the Field Can Do About It
Jack Martin
Abstract
Justice technology implementations fail at rates that have remained stubbornly consistent for decades. This paper argues that the primary cause is misclassification: the field consistently treats technology deployments as information technology projects when they are, in fact, organizational change programs that happen to involve technology. Drawing on organizational change theory, implementation science, and practitioner experience spanning more than three decades on both sides of the agency-vendor relationship, this paper identifies the structural gaps that drive implementation failure, examines why the justice context presents unique challenges, introduces the concept of the translation function as a missing organizational capacity, and offers a practitioner-centered framework for what better looks like. The paper concludes with particular attention to artificial intelligence, which threatens to amplify every existing gap if the field does not first address its foundational implementation failures.
Introduction
In 2015, I made every mistake I now warn others about.
I had a legitimate business case. My agency ran on a legacy system that would cost $35 million to update. I was losing a server a year at $250,000 a loss. I had four IT staff who could no longer write the code the old system ran on. Moving to a cloud-based case management system was not a gamble , it was the responsible decision.
A year in, I had a problem.
My team had spent twelve months recreating the old system inside the new one. Same workflows. Same logic. Same limitations. Just hosted differently. I had purchased transformation and was delivering migration.
I had delegated to county IT and a handful of motivated officers, people who knew the old system better than anyone. Which made them exactly the wrong people to reimagine it.
I was not communicating with my frontline staff. I assumed the people I had empowered were doing that. They were not.
The software worked. The vendor delivered. The contract was technically fulfilled.
What failed was translation.
That failure, and why it keeps happening across the justice system , is what this paper is about.
The Misclassification Problem
The justice system is not unique in getting this wrong.
The Standish Group has tracked technology project outcomes for decades. Their CHAOS reports, a widely cited industry survey rather than a peer-reviewed dataset, consistently find that fewer than a third of technology projects succeed as originally defined, on time, on budget, with the intended scope (Standish Group, 2020). Cost overruns, missed timelines, abandoned implementations. The numbers have barely moved in thirty years.
What has moved is our understanding of why.
The research is consistent. Technical failure is rarely the primary cause. The cause is almost always organizational. Unclear requirements. Leadership disengagement. Staff resistance. The gap between what a system can do and what an organization is prepared to do differently because of it. Gartner has estimated that by 2027, more than 70 percent of recently implemented ERP initiatives will fail to fully meet their original business-case goals, not because the software did not work, but because the organizational conditions for adoption were never established (Gartner, 2024).
Kotter (1996) identified this pattern in the private sector decades ago. His eight-stage change model was not written for government technology, but it describes government technology failures with uncomfortable precision. No burning platform to establish urgency. No guiding coalition with enough authority to sustain momentum. Change that gets declared complete before it is actually embedded in practice. Kotter’s research found that organizations that skipped even one stage rarely recovered the initiative.
Prosci’s ADKAR model (Hiatt, 2006) makes the failure mode more specific. Awareness. Desire. Knowledge. Ability. Reinforcement. Five conditions that must exist sequentially for change to stick. Most justice technology implementations address two, knowledge and ability, meaning training , and call it change management. It is not.
In the public sector, Fernandez and Rainey (2006) documented what practitioners already know. Organizational change in government is structurally harder. Incentives are different. Accountability is political rather than market driven. Civil service protections create dynamics that private sector change frameworks do not account for. And the people most invested in the current system, the ones who built their careers around it, are routinely the ones assigned to replace it.
That is precisely what happened in my implementation.
The justice system adds a layer that the broader literature rarely addresses. Probation officers, parole agents, and case managers carry professional discretion that no software can fully encode, and they know it. When a new system feels like it constrains that discretion rather than supports it, resistance is not irrational. Taxman and colleagues (2014) found that organizational readiness to support innovations in juvenile justice offices is tied to practitioner buy-in, and that buy-in cannot be mandated. It has to be earned through genuine engagement with the people doing the work.
We keep calling resistance a people problem. It is a design problem. A leadership problem. A communication problem. It is not a technology problem.
Why the Justice Context Is Distinctly Hard
Government technology implementations are hard. Justice technology implementations are harder.
The reasons are structural, cultural, and in some ways unique to an industry where the consequences of getting it wrong are not measured in lost revenue, they are measured in public safety, civil liberties, and human lives.
Procurement Protects Against the Wrong Risks
The government procurement process was designed to prevent corruption and ensure fair competition. It does neither particularly well and does almost nothing to ensure a good implementation.
I sat through a vendor meeting during my 2015 implementation where the majority of the discussion, I am not exaggerating, centered on the location and color of a button in the software interface. Forty-five minutes. Stakeholders with genuine operational authority debating something a UX designer should have resolved in an afternoon.
Meanwhile, nobody was asking how the system would change supervision workflows. Nobody was asking how officers would be trained not just to use the software but to work differently because of it. Nobody was asking what happened when the system’s logic conflicted with an officer’s professional judgment.
Those questions do not appear in most RFPs. They do not generate change orders. They do not have a line item. So they do not get asked.
Procurement in the public sector is characterized by specification rigidity, an overemphasis on technical requirements and contractual compliance at the expense of outcome definition. Public procurement operates inside a densely regulated legal and oversight environment (Thai, 2001). Agencies spend enormous energy negotiating liability clauses, indemnification language, and SLA terms that protect against vendor failure while leaving organizational readiness completely unaddressed. The contract becomes the proxy for the plan. It is not.
When the gap between what an agency believes it purchased and what a vendor believes it committed to deliver becomes visible, and it always becomes visible, the contract rarely resolves it. It just determines who pays for the change orders.
Practitioner Discretion Is Not a Variable, It Is the Work
Justice technology has a feature that enterprise software in most other sectors does not have to account for: the people using it carry legally sanctioned discretion that no requirements document can fully capture.
A probation officer deciding whether a technical violation warrants a revocation recommendation is not executing a workflow. They are exercising judgment shaped by years of case experience, knowledge of an individual, assessment of risk, and professional ethics. A case management system can support that judgment. It cannot replace it. And when it feels like it is trying to, when the system’s logic conflicts with what an officer knows to be true about a case , resistance follows.
Taxman and colleagues (2014) found that organizational readiness in juvenile justice settings is tied to frontline practitioner buy-in. Not training completion. Not go-live dates. Actual belief among the people doing the work that the system serves the mission. That belief cannot be mandated. It has to be earned through engagement that most implementation timelines do not budget for.
Political Accountability Changes the Calculus
Private sector technology projects fail with consequences that are largely internal, budget impact, competitive position, executive accountability. Government technology projects fail publicly. A failed justice technology implementation can become a news story, a legislative hearing, a campaign issue.
That reality shapes behavior in ways that make honest implementation harder. Agencies declare success before it is warranted because the political cost of acknowledging failure is higher than the operational cost of pretending. Vendors manage optics alongside deliverables. Go-live becomes the finish line when it should be the starting line.
Fountain (2001) described this as the “enactment” problem in government technology, the gap between technology as designed and technology as it is actually used within specific institutional contexts. Political pressure does not just slow implementation. It distorts it. It creates incentives to close tickets rather than solve problems.
The Translation Function
The hardest problem in justice technology is not the software. It is translation.
Not translation in a technical sense, converting requirements into code, or business rules into system logic. Something more fundamental. The capacity to move meaning accurately across the distance between people who understand the work and people who build the tools, and back again.
That distance is wider than the field acknowledges. And it runs in both directions.
What Agencies Fail to Translate
Agencies struggle to convert operational reality into product requirements. Not because they lack intelligence or commitment, but because the knowledge that makes a probation officer or parole agent effective is largely tacit. It lives in experience, in judgment, in the thousand small decisions that never make it into a policy manual.
When that knowledge has to become a requirements document, a list of fields, workflows, and system behaviors that a vendor can price and build , most of it gets lost. What survives is the surface. The screens. The reports. The buttons.
Which is how you end up paying for a new system that works exactly like the old one.
Polanyi (1966) described this as the problem of tacit knowledge, the things practitioners know but cannot fully articulate. “We know more than we can tell,” he wrote. Justice technology implementations collide with this reality constantly. The people who know the work most deeply are often the least equipped to translate it into procurement language. And the people who write the RFPs are often the least connected to the work.
The result is a requirements document that describes the current state, what the agency does now , rather than the future state , what the agency needs to become. Vendors build what they are asked to build. Agencies get what they specified. And neither party gets what the mission actually required.
What Vendors Fail to Translate
I spent years on the agency side watching vendors miss this. Then I went inside one of the largest justice technology companies in the country and understood why.
Vendors in this space are not primarily driven by product quality. They are driven by perception.
Quarterly earnings. Contract renewals. Demo environments that look nothing like production deployments. Sales cycles that overpromise and implementation teams that inherit the gap. The pressure is not to build the best product for public safety outcomes. The pressure is to manage optics for shareholders, for procurement evaluators, for the conference booth.
That is not a cynical observation. It is a structural one. Publicly traded companies answer to shareholders before they answer to agencies. And shareholders do not measure success in supervision fidelity or recidivism outcomes. They measure it in revenue retention and contract growth.
The consequence is an industry that innovates at the margin. Features that photograph well in RFP responses. AI capabilities announced before they are ready. Integrations that work in demonstrations and struggle in production. A product roadmap driven by what wins contracts rather than what moves the field forward.
Building the Translation Capacity
What the field needs, and almost never builds for is a translation function. Not a role, necessarily, though that helps. A capacity. The organizational ability to move operational knowledge into product requirements, contract language into realistic delivery plans, and technology capabilities into practitioner workflows.
That capacity requires people who understand both sides. Who have worked cases and worked product. Who can sit in an agency meeting and recognize what the RFP is missing. Who can sit in a vendor meeting and recognize what the demo is not showing.
Those people exist. The field does not systematically develop them, deploy them, or listen to them.
Fixsen and colleagues (2005), in their foundational work on implementation science, identified what they called “implementation drivers”, the competency, organizational, and leadership conditions that determine whether an evidence-based practice actually gets used. Technical installation is not implementation. Implementation is the sustained change in practice that a technology is supposed to enable. Most justice technology projects fund the installation. They never fund the implementation.
AI as the Emerging Stress Test
Everything this paper has described, the misclassification problem, the procurement failure, the translation gap, the distance between what vendors sell and what agencies need , artificial intelligence is about to make all of it worse.
Not because AI is bad technology. Because we are about to introduce powerful technology into a system that has not yet learned to implement ordinary technology well.
What AI Actually Is in This Context
The justice system is not being asked to evaluate AI abstractly. It is being asked to purchase it. Vendors are already selling risk assessment tools, predictive analytics platforms, automated case noting, and decision-support systems under the AI banner. Some of these tools are genuinely useful. Some are not ready. Most agencies cannot tell the difference, and the procurement process, for all the reasons described in this paper, is not equipped to help them.
The pattern is familiar. A capability gets announced before it is production ready. It photographs well in a demo. It wins contracts. Implementation teams arrive and discover that the operational reality is more complicated than the sales cycle suggested. And by the time the gap is visible, the contract is signed and the change orders are coming.
The Rubber Stamp Problem
Here is what actually worries me about AI in the justice system. Not that it will make bad decisions. That officers will stop making decisions at all.
AI-generated risk assessments, case notes, and supervision recommendations create a powerful cognitive shortcut. The system says low risk. The officer checks the box. The system generates a case note. The officer approves it without reading it. The system flags a violation. The officer acts on it without investigating further.
That is not supervision. That is rubber stamping with extra steps.
The officer takes a longer lunch. The caseload feels more manageable. And nobody notices until something goes wrong and the question becomes: who actually made this decision? The answer, in too many cases, will be nobody. The system did. And the officer signed off.
Dressel and Farid (2018) found that human participants with no criminal justice training performed comparably to COMPAS, one of the most widely used algorithmic risk assessment tools in the country when given the same information. The algorithm was not delivering superior judgment. It was delivering the appearance of superior judgment. And that appearance can change how practitioners engage with the decision.
Stevenson and Doleac (2024) found similarly complicated results when examining the real-world impact of risk assessment tools on sentencing. The tools changed behavior, but not always in the ways intended, and the changes did not produce detectable gains in public safety or reduced incarceration.
The problem is not that AI gets things wrong. The problem is that when AI gets things wrong inside a justice system, the consequences fall on people who are already among the most vulnerable in that system. A bad software implementation costs money. A bad AI decision can cost someone their liberty.
What the Field Is Not Asking
The questions justice agencies are asking about AI tend to be technical. What is the accuracy rate. What data does it train on. How does it integrate with our case management system.
The questions they are not asking are organizational. How will this change what my officers actually do. How will I know if they are using it as a tool or as a substitute for judgment. How will I maintain accountability when the decision trail runs through an algorithm. What happens to officer skill development when AI handles the cognitive work that used to build professional expertise.
Those are translation questions. And they will not appear in the RFP.
Angwin and colleagues (2016), in their landmark investigation of algorithmic bias in criminal justice, demonstrated that the gap between what these tools claim to measure and what they actually measure can have profound disparate impacts on communities of color. That finding did not slow AI adoption in the justice system. It should have demanded a more rigorous translation process between what the technology does and what the agency believes it is doing. That translation process does not yet exist at scale. And without it, AI will not be the next generation of justice technology. It will be the next generation of justice technology failure.
What Better Looks Like
The field does not lack for failed implementations to study. It lacks for honest prescriptions. This section offers one. Not a framework with an acronym. Not a maturity model with five levels and a consulting engagement attached. A practitioner’s answer to a practitioner’s problem.
Start With Preparation, Not Procurement
Better starts before the RFP. It starts with an agency asking an honest question: are we actually ready to change? Not ready to purchase. Not ready to go live. Ready to work differently.
Agencies treat procurement as the starting line when it should be the finish line of an internal readiness process. Before a vendor is selected, an agency should understand which workflows serve the mission and which serve habit. Which processes are grounded in evidence, and which exist because that is how it has always been done.
Kotter (1996) called this establishing urgency. In the justice context it means something more specific, distinguishing between workflows worth preserving and workflows worth replacing. Technology built on best practice cannot deliver its value if an agency insists on bending it back toward legacy process. The willingness to adapt small workflows to meet a well-designed system is not compromise. It is organizational maturity.
Measure What Actually Matters
Most justice technology implementations track the wrong metrics. Ticket closure rates. Training completion percentages. Go-live dates. These are implementation metrics. They tell you whether the project is on schedule. They tell you almost nothing about whether the technology is doing what the mission requires.
Better means tracking meaningful outcomes from the start. Disproportionality in supervision decisions are certain populations being flagged, violated, or escalated at rates that demand examination. Crime change and displacement is the supervision strategy actually affecting public safety or moving the problem. Individual officer metrics: caseload equity, contact quality, decision consistency. Early indicators of burnout and attrition.
If an agency cannot articulate what success looks like in outcome terms before implementation begins, it cannot evaluate whether it succeeded after go-live. And it will have no basis for the conversation with a vendor about whether the system delivered what the mission required.
Let Technology Do What Technology Should Do
The goal of justice technology is not to replace officer judgment. It is to protect it.
Officers in community supervision carry caseloads that would overwhelm any professional. Documentation requirements, court deadlines, collateral contacts, case notes, risk reassessments; the administrative burden of the work has grown faster than the workforce trained to carry it. The result is burnout, attrition, and the quiet erosion of the quality time that actually drives supervision outcomes.
Technology should absorb that burden. The mundane, repetitive, administrative chores that consume officer time without requiring officer judgment, are appropriate targets for automation, AI assistance, and workflow efficiency. Free the officer from the documentation so the officer can do the work that only an officer can do.
The Officer Is Still the Change Agent
Every section of this paper has described a gap. Between executives and frontline staff. Between procurement language and implementation reality. Between what AI promises and what practitioners will actually do with it.
The way across every one of those gaps runs through the same person.
The officer sitting across from someone on supervision. Making a judgment call. Building a relationship. Deciding what the data means in the context of what they know about this individual, this family, this neighborhood, this moment.
Technology can inform that judgment. It can support it. It can document it, analyze it, and surface patterns that no individual officer could see across a full caseload. But it cannot replace it. And any implementation that tries through automation that removes discretion, AI that substitutes for assessment, or efficiency metrics that reward speed over quality will fail the mission regardless of whether it succeeds as a project.
The next generation of justice technology will not be defined by its algorithms or its architecture. It will be defined by whether it makes that officer more effective, less burned out, and better equipped to do work that is genuinely hard.
That requires agencies brave enough to change. Vendors honest enough to build for outcomes rather than optics. And practitioners ethical enough to use the efficiencies technology creates for the benefit of the people they supervise , not just for longer lunches.
The opportunity is real. The technology is ready. The question is whether the field is.
References
Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016, May 23). Machine bias. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing
Dressel, J., & Farid, H. (2018). The accuracy, fairness, and limits of predicting recidivism. Science Advances, 4(1), Article eaao5580. https://doi.org/10.1126/sciadv.aao5580
Fernandez, S., & Rainey, H. G. (2006). Managing successful organizational change in the public sector. Public Administration Review, 66(2), 168–176. https://doi.org/10.1111/j.1540-6210.2006.00570.x
Fixsen, D. L., Naoom, S. F., Blase, K. A., Friedman, R. M., & Wallace, F. (2005). Implementation research: A synthesis of the literature (FMHI Publication No. 231). University of South Florida, Louis de la Parte Florida Mental Health Institute, National Implementation Research Network. https://nirn.fpg.unc.edu/resources/implementation-research-synthesis-literature
Fountain, J. E. (2001). Building the virtual state: Information technology and institutional change. Brookings Institution Press. https://www.brookings.edu/books/building-the-virtual-state/
Gartner. (2024). What IT leaders must do to avoid disappointing ERP initiatives. https://www.gartner.com/en/information-technology/insights/what-it-leaders-must-do-to-avoid-disappointing-erp-initiatives
Hiatt, J. (2006). ADKAR: A model for change in business, government and our community. Prosci Learning Center Publications.
Kotter, J. P. (1996). Leading change. Harvard Business School Press.
Polanyi, M. (1966). The tacit dimension. Doubleday.
Standish Group. (2020). CHAOS report 2020: Beyond Infinity. The Standish Group International.
Stevenson, M. T., & Doleac, J. L. (2024). Algorithmic risk assessment in the hands of humans. American Economic Journal: Economic Policy, 16(4), 382–414. https://doi.org/10.1257/pol.20220620
Taxman, F. S., Henderson, C., Young, D., & Farrell, J. (2014). The impact of training interventions on organizational readiness to support innovations in juvenile justice offices. Administration and Policy in Mental Health and Mental Health Services Research, 41(2), 177–188. https://doi.org/10.1007/s10488-012-0445-5
Thai, K. V. (2001). Public procurement re-examined. Journal of Public Procurement, 1(1), 9–50. https://doi.org/10.1108/jopp-01-01-2001-b001