Resource Directories Go Stale: Keeping a 211-Style Database Current With AI
Every community resource list starts accurate and starts decaying the same day. Phone numbers change, pantries shift their hours, programs lose funding, and eligibility rules move with each new grant cycle. Whether you run a 211 center, a library resource desk, a hospital community resource list, a legal aid referral guide, a food bank partner agency directory, or a shared spreadsheet your coalition has passed around for years, the hard part is not building the directory. It is keeping it true. This guide covers why freshness matters, the data standards that make maintenance easier (Open Referral's HSDS and the Inform USA standards), how a sound verification workflow runs, and where AI can genuinely take work off your resource specialists: watching provider websites for changes, drafting outreach, cleaning duplicates, suggesting taxonomy codes, writing plain-language and translated descriptions, and triaging error reports. It also covers the line AI should never cross, which is publishing a record no human has verified.

Picture a parent who calls for help finding diapers and formula. The resource specialist finds a listing for a church pantry two bus transfers away, open Thursday mornings. The parent arranges childcare, takes the bus, and arrives to a locked door. The pantry moved its distribution to Saturdays four months ago and nobody told the directory. That parent has lost a morning, bus fare, and, most importantly, some measure of trust that calling for help is worth the effort. The next time something goes wrong, they may not call at all.
This is the quiet failure mode of every community resource directory. A stale record does not throw an error message. It looks exactly like an accurate one. The cost lands on the person least able to absorb it, and it often never makes its way back to the organization that published the listing. Multiply that by thousands of records across a region and the directory's value depends almost entirely on one unglamorous discipline: continuous verification.
Professional information and referral (I&R) services have treated this as a core function for decades. 211 centers employ dedicated database curators, follow published standards, and run annual verification cycles. Many other organizations that maintain resource lists, from public libraries to hospital social work departments to neighborhood coalitions, do the same work with far less structure and often with a single overextended staff member. In both settings, the work is repetitive, detail-heavy, and chronically under-resourced, which is exactly the kind of work where carefully scoped AI can help.
This article walks through why freshness matters, the standards that give directory data a stable shape, what a good verification workflow looks like, seven specific places AI can reduce the burden on your team, the risks you need to design around, and how to start small. One boundary is worth stating up front. Many 211 lines also take calls from people in crisis. That work belongs to trained humans, and nothing in this article suggests otherwise. Everything here is about maintaining the directory behind the phones, not about answering them.
Why a Dead Referral Costs More Than It Looks
Directory decay is constant and mostly invisible. Small community-based organizations are especially volatile. A food pantry run by volunteers may change hours when its coordinator changes jobs. A rental assistance program may open for six weeks when funds arrive and close when they run out. A clinic may stop accepting new patients without updating its website. Even large, stable agencies reorganize phone trees, move intake online, or rename programs. None of these organizations is obligated to tell you, and most are too busy to remember.
The harm of a stale record compounds in ways that are easy to underestimate. The direct cost is the wasted trip or call. The indirect cost is credibility: people who receive one bad referral are less likely to trust the next one, and they tell others. Referral staff feel it too, because they lose confidence in their own tool and start keeping private notes or sticky-note workarounds, which fragments knowledge further. Partner agencies get frustrated when they receive callers seeking services they stopped offering a year ago. And funders and planners who rely on your database to identify service gaps end up working from a distorted map.
Fast-decaying fields
Where errors appear first
- Hours of operation, especially seasonal and holiday changes
- Phone numbers and intake contacts after staff turnover
- Program status for grant-funded or funding-limited services
- Eligibility criteria, income limits, and required documents
- Application method, such as walk-in versus online-only intake
Slower-decaying fields
Still need checking, just less often
- Organization legal name and mission
- Physical address of established sites
- Core service types for long-running programs
- Service area boundaries such as county or ZIP coverage
- Languages offered, though this can shift with staffing
Knowing which fields decay fastest matters because it shapes where you spend verification effort. An annual review of every field in every record is the baseline, but a pantry's hours may need attention every quarter while its legal name may not change for a decade. Much of the value AI brings to this work comes from helping you notice change between scheduled reviews, so your specialists can focus on the records that actually moved.
The Standards That Give Directory Data a Stable Shape
Before talking about AI, it helps to understand the standards that professional I&R services already use. Even if you maintain a small coalition spreadsheet and have no intention of seeking accreditation, borrowing from these standards will make your data easier to maintain, easier to share, and far easier for AI tools to work with. Structured, consistent data is the precondition for nearly every useful automation described later in this article.
The Inform USA Standards. Inform USA, the professional association formerly known as the Alliance of Information and Referral Systems (AIRS), publishes the Standards for Professional Information and Referral. The current edition is Version 10, released in July 2024 with an update in October 2025, and the standards are reviewed on a three-year cycle. A dedicated section covers the resource database. It expects services to publish clear inclusion and exclusion criteria, organize records in a hierarchy of organization, location, and service, track administrative fields including the date of last annual verification, have a trained database curator review all updates before publication, gather interim changes between annual reviews, and prioritize critical or high-referral resources. When an organization does not respond after multiple direct attempts, the standards point curators toward alternate methods such as calling staff, checking websites, or visiting in person.
In practice, 211 centers translate that into published local policies. Michigan 211 describes contacting each partner organization annually to update its information. 211 San Diego and Imperial states in its inclusion policy that failing to respond to annual update requests may lead to listings being inactivated or removed. The common thread is a yearly verification floor, with faster updates whenever the curator learns of a change.
The 211 Human Services Indexing System. Most 211s classify services with a shared taxonomy maintained by 211 LA. Long known as the AIRS/211 LA County Taxonomy, it is now called the 211 Human Services Indexing System (211HSIS). It is a hierarchical vocabulary of service terms with alphanumeric codes, so "food pantries" sits under a broader basic needs branch and can be searched alongside related terms. It is licensed rather than freely available. The Inform USA standards require some classification system, recommend 211HSIS for 211 providers, and accept alternatives that are machine-readable, actively maintained, and either mapped to 211HSIS or non-proprietary.
Open Referral and HSDS. Open Referral maintains the Human Services Data Specification (HSDS), an open data standard for describing organizations, services, locations, and the relationships among them. Version 3 made JSON the primary format and folded the API specification into the core standard, with point releases continuing since. For maintenance purposes, HSDS is notable because it builds freshness into the data model. The service object includes an assured_date (when the information was last checked), an assurer_email (who checked it), and a last_modified timestamp, and a metadata object can log each field change with its previous value, new value, and who made it. The Inform USA standards point to HSDS as the expected schema when publishing database content to other platforms.
What to borrow even if you are not a 211
Small directories benefit from the same structure
- Separate organization, location, and service into distinct rows or tabs instead of one cell per agency
- Add "last verified on" and "verified by" columns to every service, modeled on HSDS's assured date
- Write down inclusion criteria so you can explain why a listing is or is not there
- Keep hours, eligibility, and required documents in their own fields rather than buried in description text
- Use a consistent service-type vocabulary, even a simple one, so similar services can be found together
If your current directory is a single spreadsheet with free-text cells, restructuring it is the highest-leverage step you can take before any AI work. Our guide to cleaning data before using AI covers the general approach, and the same principles apply here: consistent fields, clear owners, and a record of when each value was last confirmed.
What a Sound Verification Workflow Looks Like
AI fits into a workflow. It does not replace one. If your team has no consistent verification routine today, adding AI will mostly speed up inconsistency. The good news is that the core workflow is simple, and most of its steps are the ones AI can accelerate. Think of it as three overlapping cycles: a scheduled annual review of every record, an interim update process for changes you learn about between reviews, and a quality review that looks across the database for patterns.
The core verification loop
The human-owned process that AI supports
- 1. Schedule. Every record has a due date for its next review, typically within twelve months of the last verification, with shorter cycles for high-referral or volatile services.
- 2. Request. The curator sends the organization its current record and asks it to confirm or correct each field, by email, web form, or phone.
- 3. Follow up. Non-responders get reminders, then a phone call, then alternate verification through the organization's website, public notices, or a site visit.
- 4. Review and edit. A trained curator reviews every proposed change, edits to the style guide, and applies taxonomy terms before anything is published.
- 5. Record. The verification date, the method used, and the person who confirmed it are logged on the record.
- 6. Resolve. Records that cannot be verified are flagged, marked inactive, or removed under your published policy.
Notice that every step involves either communication or judgment. Steps one, two, and three are mostly communication logistics: tracking due dates, writing messages, sending reminders, and checking websites. Steps four and six are judgment calls about what is true and what belongs in the directory. Step five is record-keeping. That distribution tells you where AI belongs. It can do a great deal of the logistics and record-keeping, it can prepare evidence for judgment calls, and it should not make the judgment calls itself.
One more point that is easy to miss: verification is a relationship, not just a transaction. Organizations that hear from you only once a year to fill out a form tend to treat the request as a chore. Organizations that see referrals coming from your directory, that get a friendly heads-up when something on their website looks different from your listing, and that can update their own record easily, tend to respond faster and volunteer changes on their own. Several of the AI uses below support that relationship directly by making your outreach more specific and less burdensome.
Seven Places AI Takes Work Off Your Resource Specialists
The uses below share one design principle: AI produces a suggestion, a flag, or a draft, and a human resource specialist decides what happens next. None of them involve AI writing directly to the published directory. Each one targets a task that is currently eating hours your curators could spend on calls, relationship building, and the judgment work that only they can do.
1. Monitoring provider websites for changes
Catch changes between annual reviews
Many organizations update their own website or social media long before they would think to tell a directory. A monitoring tool can check the relevant page for each listed service on a regular schedule, detect when it changes, and hand the old and new versions to a language model with a narrow question: does this change affect hours, phone numbers, address, eligibility, or program status? If the answer is yes, the record goes into a review queue with the specific passage highlighted and a suggested field update. The specialist confirms with the organization, then edits the record.
The value is not that AI reads the page correctly every time. It is that it narrows hundreds of pages to the handful that need a human look this week. A banner saying "Pantry closed through January for renovations" can be caught in days rather than at next year's annual review. Keep the monitored list focused on the specific page that holds hours or eligibility, not the whole site, to reduce noise and server load.
2. Drafting verification outreach and call scripts
Specific requests get faster answers
Generic "please review your listing" emails get ignored. A message that says "Our listing shows your clothing closet open Tuesdays and Thursdays 10 to 2, but your website now lists Wednesdays only. Could you confirm which is current?" gets a reply. AI can draft that kind of targeted message for every record due for review, pulling the current field values and any detected website discrepancies, in the organization's preferred language. It can also produce short phone scripts for specialists making follow-up calls, listing exactly which fields to confirm and in what order.
Keep a human in the send loop, at least at first. Review a sample of drafts each batch to make sure tone and accuracy hold up, and never let the model invent a discrepancy it was not given. The prompt should include only the record and the detected change, with an instruction to ask about nothing else.
3. Normalizing and deduplicating records
One service, one record
Directories accumulate duplicates through imports, merged lists, and different staff entering the same program under slightly different names. "St. Vincent de Paul Food Pantry," "SVdP Pantry - Main St," and "Saint Vincent Food Shelf" may be one service or three. AI is well suited to normalizing formats (phone numbers, addresses, hours written as "M-F 9-5" versus "Monday through Friday, 9am to 5pm") and to proposing likely duplicate pairs with a short explanation of why they match.
Merges are judgment calls. Two pantries run by the same parish at different addresses are distinct locations, and a merge would delete a real resource. Have AI propose, and have a specialist approve each merge. For a deeper look at the general technique, see our article on AI-assisted CRM cleanup, which applies the same duplicate-review logic to donor data.
4. Suggesting taxonomy codes for human approval
Faster, more consistent indexing
Applying taxonomy terms well takes training and experience. A new specialist may index a program that offers rental assistance, utility help, and budget counseling under only one of those terms, making it invisible to searches for the others. Given the service description and your licensed taxonomy (or your local vocabulary), AI can suggest a ranked set of candidate terms, each tied to the phrase in the description that justifies it. The curator accepts, rejects, or adds terms.
This works best when the model is constrained to choose from your actual term list rather than generating terms or codes from memory, which invites invented codes that look plausible. Respect your taxonomy license terms when deciding which tools can receive the term list. Over time, comparing AI suggestions to curator decisions also gives you a useful consistency check across staff.
5. Writing plain-language and translated descriptions
Listings people can actually understand
Agency-supplied descriptions are often written in grant language: "provides wraparound supports to income-eligible households experiencing housing instability." A plain-language rewrite says what happens: "Helps families who are behind on rent or about to be evicted. Can help pay back rent and connect you with a case worker." AI does this rewriting well when it is told to preserve every factual detail, add none, and flag anything ambiguous. Our guide to plain-language rewriting with AI covers the technique in depth.
Translation follows the same rule. Machine translation makes multilingual directories affordable, but eligibility terms and document requirements are exactly where subtle mistranslation does damage. Have fluent reviewers check high-traffic listings and anything involving eligibility or legal status. Our article on translation quality review describes a tiered approach that fits this well.
6. Triaging user-reported errors
Turn feedback into a work queue
The best source of interim changes is the people using the directory: callers, case managers, librarians, and the agencies themselves. Many directories have a "report an error" link or an inbox that fills with messages of uneven quality. AI can read each report, match it to the right record, classify it (wrong phone, closed program, changed hours, eligibility question, not an error), and estimate urgency. A report that a shelter has closed should jump the queue ahead of a typo in a description.
Front-line staff reports deserve particular attention. Resource specialists on the phones often learn about changes first because callers tell them. A simple internal form, with AI doing the matching and sorting, captures that knowledge before it disappears into someone's memory.
7. Measuring record age and directory health
Know how stale you are
Many organizations cannot answer a simple question: what share of our listings has been verified in the last twelve months? If you track a verification date on every record, a spreadsheet formula can answer it. AI helps with the next layer, which is analysis and explanation. It can summarize which service categories are most overdue, which organizations repeatedly fail to respond, which records generate the most error reports, and how verification progress is tracking against your annual goal, then turn that into a short monthly summary for your manager or board.
These numbers also support the gap analysis that the Inform USA standards encourage, where database contents are used to identify missing services and shared with funders and planners. A directory that knows its own freshness can also be honest with users, for example by displaying "last verified" dates on each listing.
The Non-Negotiable Rule: AI Never Publishes a Record
Language models generate plausible text. A food pantry listing is a very predictable shape: a name, an address, hours, a phone number, a short eligibility statement. A model asked to "fill in the missing hours" or "find a food pantry in this ZIP code" can produce a listing that looks exactly right and describes a service that does not exist, at hours nobody keeps. Our guide to reducing AI hallucinations explains why this happens. In a resource directory, the consequence is the same locked door from the introduction, except now your organization manufactured the error rather than merely failing to catch it.
The rule that follows is simple and should be written into your procedures: AI may suggest, draft, flag, and summarize, but no record is created, changed, or published without a human specialist verifying it against a primary source, meaning the organization itself or its own current published materials. That matches the Inform USA expectation that a database curator reviews every update before it is published, and it is the single most important control in this entire article.
Failure modes to design against
Common ways an AI-assisted workflow can corrupt a directory
- Gap filling. A model asked to complete a record invents hours or eligibility rather than leaving a field blank. Instruct tools to return "unknown" and treat any value without a cited source passage as unverified.
- Stale web pages. An organization's own website can be out of date too. A website change is a reason to contact the agency, not proof of the new facts.
- Over-eager merges. Automated deduplication collapses two real locations into one, silently removing a resource.
- Softened eligibility. A plain-language rewrite drops a residency or age restriction, sending people who do not qualify.
- Verification theater. A record gets a fresh "verified" date because AI checked a website, when no one confirmed anything with the provider. Log the verification method honestly.
That last item deserves emphasis. The verification date on a record is a promise to users and to everyone downstream who consumes your data. If you introduce automated website checks, record them as a distinct method, such as "website reviewed, not confirmed with provider," so that your freshness metrics reflect reality. Inflating verification counts with automated checks will make your dashboard look better while your referrals get worse.
Scraping Etiquette and the Chatbot Problem
Website monitoring is one of the most useful AI-adjacent tools for directory maintenance, and also one where it is easy to behave badly without meaning to. The organizations you monitor are your partners. Many run on donated web hosting and shoestring budgets, and a monitoring script that hits their site too often can slow it down or run up costs. A few practices keep you on good terms.
Monitoring etiquette
- Check one specific page per service, at most daily and usually weekly
- Honor robots.txt, which RFC 9309 defines as the standard way sites signal crawler preferences
- Read and respect each site's terms of use, and ask permission when they restrict automated access
- Identify your organization in the tool's user agent so site owners know who is checking
- Stay on public pages; never log in, bypass forms, or collect personal information
- Tell partners you monitor their listing pages and why, ideally in your inclusion policy
Chatbots on stale data
- A chatbot is only as current as the data it reads, and it states stale facts with full confidence
- General-purpose chatbots may answer from training data years out of date rather than your directory
- Bots should show the "last verified" date and a link to the source record
- Any sign of crisis or urgent risk must route immediately to a trained human
The chatbot issue is growing more pressing. Many organizations are tempted to put a conversational search interface on top of their resource directory, and the Inform USA standards already recognize chatbots as one possible access channel. A well-built bot that searches your verified database, quotes records directly, and shows when each was last confirmed can make a directory easier to use, especially after hours. A bot built on a general model that "knows" about local services, or one that paraphrases records loosely, will hand out referrals your specialists would never give. A chatbot does not reduce the need for fresh data. It raises it, because it removes the experienced human who might otherwise notice that a listing looks wrong.
If you deploy any public-facing assistant, keep it strictly retrieval-based on verified records, test it against tricky questions before launch, and make the path to a human obvious. Our article on testing nonprofit chatbots with adversarial prompts describes how to stress-test one. And to repeat the boundary from the introduction, people who reach out in crisis need trained humans. A directory bot should recognize that it is out of its depth and hand off, not attempt to help.
Keeping Resource Specialists in Charge
Database curation is a professional skill. Inform USA offers a dedicated credential for database curators, and experienced specialists carry enormous local knowledge: which agencies answer email and which only answer the phone, which programs say "county residents" but quietly serve neighbors across the line, which pantry director will tell you about a change before it happens. AI cannot replicate that knowledge, and a rollout that treats curators as a cost to be automated away will lose it.
The better framing is that AI handles the parts of the job that curators find most tedious, such as reformatting hours, chasing non-responders with yet another reminder, and sorting a messy error inbox, so that they spend more of their time on verification calls and relationship building. Involve curators in designing the workflow. They know where the bottlenecks are, and they will spot AI mistakes faster than anyone. Where staff are anxious about what AI means for their roles, our guide to overcoming staff resistance to AI offers practical approaches.
Who decides what
A simple division of labor
- AI proposes: changed-page flags, outreach drafts, duplicate pairs, taxonomy candidates, plain-language drafts, error report classifications, freshness summaries.
- Specialists decide: whether a change is real, what the record says, which taxonomy terms apply, whether to merge, whether to inactivate, and when a record counts as verified.
- Managers own: inclusion criteria, the verification policy, which tools may touch directory data, and how AI involvement is logged.
- Providers confirm: the facts about their own services, which is why outreach remains the backbone of verification.
Data handling also needs attention. Directory records about organizations are generally public information, but your workflow will touch other material: error reports that may include a caller's description of their situation, notes from verification calls, and internal contact details for agency staff that are not meant for publication. Keep personal information out of prompts, use tools your organization has approved for this kind of data, and make sure internal-only fields never leak into public descriptions. A model rewriting a record does not know which phone number is the public intake line and which is the director's cell unless your data structure tells it.
Starting Small, Whatever Kind of Directory You Run
The right first step depends on what you maintain today. A 211 with an established resource database platform, trained curators, and thousands of records is in a very different position from a library branch with a printed referral sheet or a coalition with a shared spreadsheet. The principles are the same, but the entry points differ.
Established 211s and I&R services usually have structure and process but not enough hours. The quickest wins tend to be website change monitoring for high-referral services, AI-drafted targeted outreach for overdue records, and error-report triage. Check what your database vendor already offers or plans before building anything, since many platforms are adding AI features, and evaluate them against the standards' expectation that new methods be assessed before implementation.
Libraries, hospitals, and legal aid programs often maintain a focused list for a specific audience, such as patients being discharged, library patrons asking at the reference desk, or legal aid clients who need non-legal help. Here the first task is usually structural: move the list into consistent fields with verification dates, then use AI for plain-language rewriting and periodic freshness reports. Consider whether your local 211 can supply verified data so you are not duplicating its work. Legal aid organizations will find related ideas in our article on AI for legal aid organizations.
Food banks and network hubs maintain partner agency lists that the public uses to find a pantry, and these lists face the fastest decay of all because so many partners are volunteer-run. Pairing your partner agency reporting cycle with directory verification, and using AI to compare what partners report with what the public listing says, can catch drift early. Our guide to AI in food banks covers the broader operational picture.
Coalitions with a homegrown spreadsheet should resist the urge to build a full system. Add verification date and verifier columns, agree on who owns each section of the list, and use AI to normalize formatting and flag records older than twelve months. That alone will put you ahead of many directories.
A 90-day starting plan
Sequence the work so each step builds on the last
- Weeks 1-2: Measure freshness. Confirm every record has a verification date, then calculate how many are older than twelve months, broken down by service category.
- Weeks 3-4: Write the rules. Document that AI never publishes, define how AI involvement and verification method are logged, and choose approved tools.
- Weeks 5-8: Pilot one use. Pick a single high-value task, such as outreach drafting for overdue records or monitoring the top fifty most-referred services, and run it with curator review of every output.
- Weeks 9-10: Review the results. Compare response rates, time per record, and error rates against your baseline. Ask curators what helped and what created extra work.
- Weeks 11-12: Expand carefully. Add a second use, publish "last verified" dates to users if you do not already, and share freshness metrics with leadership and funders.
A note on scope
This article is about keeping directory data accurate. It does not cover closed-loop referral platforms or community information exchanges, which add consent, client data sharing, and referral tracking on top of a resource directory. Those systems raise their own governance questions. They also depend completely on the directory underneath them, so everything here applies to them as a foundation.
Conclusion
A resource directory is a promise that when someone follows a referral, the help will be there. Keeping that promise is relentless work, because the services it describes change constantly and rarely announce it. Professional I&R services have built strong practices around this: published inclusion criteria, structured data, a shared taxonomy, annual verification, and trained curators who review every change. Open data standards like HSDS make that structure portable and build freshness right into the record.
AI fits naturally into the most time-consuming parts of the job. It can watch provider websites and flag the handful of records that moved, draft specific outreach that organizations actually answer, clean up formatting and duplicates, suggest taxonomy terms, rewrite listings in plain language, sort error reports, and tell you honestly how stale your data is. Each of those uses ends with a human decision. The one thing AI must never do is put a record in front of the public that no person has verified with the provider.
Start by measuring how fresh your directory really is. Then pick one task, keep your specialists in charge, and log verification honestly. A directory that is current, and can show that it is current, is one of the most practical forms of help a community organization can offer.
Keep Your Referrals Worth Trusting
We help 211s, libraries, coalitions, and service networks design AI-assisted verification workflows that keep resource directories accurate while leaving the decisions with your resource specialists.
