The new AI initiative had a budget, a vendor shortlist, and a launch date, and the deputy director who inherited it could not sleep. The board chair had asked for a roadmap by the end of the month, and what existed instead was a wish list assembled from vendor demos: a tool that drafts case notes, a tool that screens intake calls for risk, a tool that automates eligibility determinations, a chatbot that answers client questions about benefits. Each had impressed someone in a meeting. None had been placed in any order, weighed against any harm, or matched against what the agency could actually verify and audit. The director knew that if she sequenced this wrong, if she led with the risk-screening tool because it was the flashiest demo, she could be standing in front of a family's advocate in a year explaining how a screening signal flagged a neighborhood at twice the rate of a comparable one, deployed before anyone audited it. And she knew that if she sequenced it right, if she led with the documentation goldmine that returns hours to home visits while carrying the lowest harm, she could show the board a program that gave caseworkers their afternoons back before it ever touched a consequential decision. A roadmap is not a list of tools. It is the order in which an agency chooses to take on benefit and risk, and the order is the whole strategy.

Why Sequence Is the Strategy

Agencies tend to treat an AI roadmap as a budgeting exercise: a list of tools, a cost for each, a timeline that fits the fiscal year. That framing misses the thing that actually determines whether the program helps families or harms them, which is the order. In human services the decisions that AI can brush against, the decisions to remove a child, substantiate a report, or deny benefits, are bound by due process and equity, and the harm from a wrong move is not a budget overrun, it is a family separated or a household left without food. So the sequence is not a scheduling detail. It is the agency's considered judgment about which benefits to capture first and which risks to defer until the agency has built the capacity to manage them.

The governing logic comes straight from the field's spine. Documentation is the goldmine: AI that drafts a case note from a home visit returns hours to human connection, the most universal and humane use case, and it carries the lowest harm because it informs no decision, it merely drafts a record that a human verifies. Predictive risk-screening is the second well, real but ethically fraught, because a screening signal can quietly encode the inequities in its training data and history proves it has. A roadmap that leads with the goldmine builds workforce trust, verification discipline, and audit capacity on the safest ground, and only then takes on the fraught uses, with that hard-won capacity already in place. A roadmap that leads with risk-screening because it demos well takes on the agency's highest hark before it has built any of the discipline that would catch the harm.

Put it in concrete terms. An agency with one hundred caseworkers, each carrying twenty-five to thirty families and spending half their day documenting, can give back roughly two hours a day per worker with a well-deployed, verified documentation tool. That is real time returned to home visits, and it is returned with almost no due-process exposure, because the worker still writes nothing they did not observe and verifies every claim against the record. The same agency that instead leads with an unaudited risk-screening tool exposes every family the tool touches to a signal that could be biased, before anyone on staff has learned to read a signal as one input rather than a verdict. Same budget, same vendors, opposite risk posture, and the only difference is the order.

A roadmap is not a list of tools and dates. It is the order in which an agency chooses to take on benefit and harm, and the order is the entire strategy.

The Benefit and Harm Grid

The roadmap is built on a deliberate assessment of every candidate use against two axes: the benefit it returns and the harm it could cause if it fails. These are not the same axis, and conflating them is the most common roadmap error. A tool can be high benefit and low harm, which is the goldmine. It can be high benefit and high harm, which is the fraught territory that must be handled with maximum care and taken on late. It can be low benefit and high harm, which is the territory an agency should usually decline entirely. The grid forces the agency to place each candidate honestly rather than ranking by demo impressiveness.

Benefit in this field has a specific texture. It is measured first in hours returned to direct work with families, because the documentation burden is the defining pain and a top driver of burnout and turnover, and turnover raises caseloads for those who remain. A documentation tool that returns two hours a day per worker across a hundred-worker agency is returning the equivalent of dozens of full-time positions to the mission, without hiring. Benefit is also measured in accuracy and consistency, in faster client service where speed genuinely helps a family in crisis, and in reduced backlog. But the anchor metric is time returned to human connection, because that is the benefit that addresses the field's deepest wound.

Harm has an equally specific texture, and it is where this field diverges sharply from a corporate AI roadmap. Harm is measured by proximity to a consequential, due-process-bound decision. A tool that drafts a record a human verifies is far from the decision. A tool that produces an eligibility determination a worker signs is closer. A tool that scores a family's risk of future harm is closer still, because the signal can substitute for judgment under caseload pressure. And a tool that automates any part of a decision to remove, substantiate, or deny crosses the line entirely and should never be on the roadmap, because it violates the cardinal rule that AI informs and humans decide. Harm is also measured by equity exposure: whether the use could produce disparate outcomes across the populations the agency serves, and whether an audit could detect that disparity before it harms a family.

When the four common use cases are placed on the grid, the pattern is clear. Documentation support sits high benefit, low harm: enormous time returned, far from any decision, the goldmine. Resource navigation and client support sits moderate benefit, low-to-moderate harm: it helps connect people to services, and its harm is bounded as long as client-facing content is verified before it reaches a person in crisis. Eligibility support sits high benefit, moderate-to-high harm: real time returned and faster determinations, but every output applies policy that, misapplied, denies a family food or shelter, so it demands verification against the current policy source and the determination stays human. Risk-screening sits moderate benefit, high harm: it can surface early indicators, but it carries the field's deepest equity peril and must be treated as an audited input under mandatory human review, never a verdict, and taken on only after the agency has built equity-auditing capacity.

The Three Waves

The grid produces a sequence, and the cleanest way to express that sequence is in three waves, each of which builds the capacity the next wave requires. The waves are not arbitrary phases on a calendar; each one is defined by what it lets the agency do that it could not do before.

Wave One: The Documentation Goldmine

The first wave is documentation support, because it is the highest benefit at the lowest harm and because it builds the disciplines every later wave depends on. In this wave the agency deploys AI to draft case notes, court reports, and intake summaries, grounded in the record, with every factual claim verified against the source before filing. The benefit is immediate and humane: hours returned to home visits. But the strategic value is what the wave teaches the workforce. Caseworkers learn to verify an AI draft claim by claim, to catch a fabricated observation before it reaches a court report, and to treat the model as a capable but unreliable colleague. Supervisors learn to review AI-assisted documentation as a draft, not a finished product. The agency learns to log AI use for an audit trail. None of these capacities can be skipped before the fraught waves, and Wave One builds all of them on ground where a mistake is recoverable, because a fabricated observation caught in verification harms no one, while the same fabrication in a deployed screening tool harms a family.

A concrete Wave One target: across a unit of forty caseworkers each carrying twenty-five families, deploy a grounded documentation tool with a mandatory verification log, measure the time returned and the verification completion rate, and do not advance to a fraught wave until verification discipline holds steady, meaning logs are complete on the high-stakes documents at near one hundred percent. The verification completion rate is the readiness signal: an agency where verification erodes under caseload pressure in Wave One is an agency that is not ready to put a screening signal in front of a worker who will treat it as a verdict.

Wave Two: Eligibility Support and Navigation

The second wave takes on eligibility support and resource navigation, uses that carry more harm than documentation because they are closer to a determination or reach a client directly, but that are still bounded by a clear verification discipline. Eligibility support drafts a determination that a benefits worker verifies against the current policy source, because a misapplied rule can deny a family SNAP, Medicaid, or TANF (Temporary Assistance for Needy Families, the federal cash-assistance program) that they were entitled to. The wave is taken on second because it relies on the verification muscle built in Wave One, now applied to policy rather than observations, and because the determination must stay human, which the agency has already learned to enforce. Resource navigation, helping connect a family to services, is bounded by verifying client-facing content before it reaches a person in crisis, since a wrong benefit deadline sent to a household in acute need does real harm.

Wave Two is where the roadmap's discipline about staying human pays off. The agency does not deploy a tool that auto-approves or auto-denies; it deploys a tool that drafts a determination a worker owns and signs, with the policy verified against the manual rather than against the model that proposed it. The harm is real but managed, because the verification discipline and the human-decides boundary are now established practice rather than aspiration.

Wave Three: Screening Support, Handled Carefully

The third wave, and only the third, takes on risk-screening support, the second well, and it is taken on last because it carries the field's deepest harm and requires every capacity the prior waves built plus one more the agency must stand up explicitly: equity auditing as a continuous practice. A screening signal is treated as one audited input under mandatory human review, never a verdict, and it is not deployed until the agency has an equity-auditing program that tests for disparate outcomes before harm and a governance body that can suspend the tool if an audit surfaces disparity. The history that makes this non-negotiable, the Allegheny Family Screening Tool debate, the Dutch childcare-benefits scandal, Michigan's MiDAS, is a history of tools deployed before the auditing and oversight existed. The roadmap's entire purpose is to reverse that order: build the capacity first, take on the fraught use last.

The readiness gate for Wave Three is the strictest. The agency does not advance until it can audit a screening tool for disparate outcomes across the populations it serves, until a governance board with real suspend authority exists, until the workforce reliably reads a signal as one input rather than a verdict, and until disclosure and due-process safeguards are in place so a family can understand and challenge how a signal was used. An agency that cannot meet that gate does not get to Wave Three. Deferring it is not a failure of the roadmap; it is the roadmap working as designed, because a screening tool an agency cannot audit is a tool it should not deploy.

Readiness Gates Between the Waves

The waves are connected by gates, and the gates are what make the roadmap a strategy rather than a timeline. A gate is a set of conditions the agency must meet before it advances, and the discipline of the roadmap is the willingness to not advance when the gate is not met, even when a budget cycle or a vendor relationship is pushing forward. An agency that treats the waves as calendar phases, advancing on schedule regardless of readiness, has thrown away the protection the sequence was designed to provide.

The gate from Wave One to Wave Two is verification discipline that holds. Concretely: verification logs complete at near one hundred percent on high-stakes documents, supervisors reviewing AI-assisted documentation as drafts, an audit trail that a court or advocate could read, and measured time returned that the agency can defend. If verification is eroding under caseload pressure in Wave One, advancing to eligibility support, where a misapplied policy denies a family benefits, takes on more harm with a discipline the agency has already shown it cannot sustain. The gate says stop and fix the verification before adding harm.

The gate from Wave Two to Wave Three is the strictest and most consequential: a standing equity-auditing program, a governance board with suspend authority, demonstrated workforce capacity to treat a signal as an input rather than a verdict, and due-process safeguards including disclosure and the right to challenge. This gate exists because Wave Three is where the field's history of harm lives, and the entire roadmap is built to ensure the agency arrives at that gate with the capacity the prior agencies lacked. The willingness to hold at this gate indefinitely, to run a strong documentation and eligibility program and simply never take on screening because the agency cannot audit it responsibly, is a legitimate and often correct outcome. The roadmap does not require reaching Wave Three. It requires never reaching it unprepared.

A worked example shows the gate doing its job. An agency completes a strong Wave One, deploys eligibility support in Wave Two, and faces budget pressure to add a vendor's risk-screening tool because the grant year is ending. The roadmap's gate forces the question: does the agency have an equity-auditing program and a governance board with suspend authority? It does not. Under a real roadmap, the answer is that the agency does not advance, and the grant either funds standing up the auditing and governance capacity first or funds deepening the documentation program instead. The gate converts budget pressure into a decision that protects families rather than a deployment that exposes them.

Building the Roadmap on a Page

A roadmap that no one can read is a roadmap no one will follow. The deliverable that comes out of all this analysis should fit on a page that a board, a director, a union, and an advocate can each understand, because the roadmap has to be defensible to all of them. The structure that works is simple and follows directly from the grid and the waves.

The page states the sequencing principle first, in one sentence: the agency pursues the documentation goldmine first because it returns the most hours at the least harm, takes on eligibility and navigation second under an established verification discipline, and takes on risk-screening last and only behind an equity-auditing and governance gate, because it carries the field's deepest harm. That sentence is the strategy, and everything else on the page supports it.

The page then lists each wave with three things: what it deploys, what benefit it captures expressed in hours returned and families served, and what capacity it builds that the next wave requires. It states the readiness gate between each wave as explicit, measurable conditions, not vague aspirations, so that the decision to advance or hold is a judgment against criteria rather than a matter of schedule. It names what the agency declines entirely: any use that automates a consequential decision, because that violates the cardinal rule, and any use that is high harm and low benefit, because the trade is not worth the exposure. Naming the declined uses is as important as naming the pursued ones, because it shows the board and the advocate that the agency considered and rejected the dangerous options deliberately.

Finally the page ties to governance and equity explicitly: every wave is subject to the governance board's approval, every fraught use carries an equity-audit obligation, and every consequential decision stays human and documented for an audit trail. The roadmap is not a separate document from the agency's equity and governance commitments; it is the schedule on which those commitments are honored. A director who can put that page in front of a skeptical advocate and walk them through the sequence, the gates, the declined uses, and the equity ties has built not just a plan but a defense, and in this field the two have to be the same thing.

Key Takeaways

• A roadmap is the order in which an agency takes on benefit and harm, not a list of tools and dates. In human services, where a wrong move can separate a family or deny a household food, the sequence is the entire strategy.
• Sequence by the field's spine: documentation is the goldmine, high benefit and low harm because it informs no decision, so it goes first. Predictive risk-screening is the fraught second well, high harm because a signal can encode inequity, so it goes last and only behind an equity and governance gate.
• Build the roadmap on a two-axis grid of benefit (anchored in hours returned to direct work with families) and harm (measured by proximity to a consequential, due-process-bound decision and by equity exposure). Documentation is high benefit, low harm; eligibility is high benefit, moderate-to-high harm; risk-screening is moderate benefit, high harm.
• Wave One deploys grounded documentation support and, just as importantly, builds the verification discipline, supervisory review, and audit trail every later wave requires, on ground where a caught mistake harms no one.
• Wave Two takes on eligibility support and resource navigation under the verification muscle built in Wave One, with the determination kept human and policy verified against the current source, not the model.
• Wave Three takes on risk-screening last, as an audited input under mandatory human review, and only after the agency stands up continuous equity auditing and a governance board with real suspend authority, reversing the historical order in which tools were deployed before oversight existed.
• The gates between waves are what make the roadmap a strategy. Advancing requires meeting explicit, measurable conditions, and the discipline is the willingness to hold at a gate, even indefinitely, when readiness is not met, including never reaching Wave Three if the agency cannot audit screening responsibly.
• The roadmap fits on a page that a board, director, union, and advocate can each read: the sequencing principle in one sentence, each wave's deployment and benefit and capacity built, the measurable gates, the uses declined entirely, and the explicit ties to governance, equity, and the cardinal rule that humans decide.