SLED AI Search and Chat Procurement Trends 2026: What's Actually Changed

Procurement is where AI ambition meets government reality. An agency can want a chatbot or a smarter search engine as much as it likes; nothing happens until a procurement officer approves a contract vehicle, a governance committee signs off on the risk assessment, and a budget line survives the fiscal year. This report looks at how that process has actually changed for state, local, and education (SLED) government AI procurement between 2025 and 2026, based on patterns Keyspider has observed across our own customer engagements and a review of publicly available RFP and RFI language issued by SLED entities during the same period.

Three shifts stand out. AI governance review has become a formal procurement gate in a growing share of agencies, not just a policy aspiration. Cooperative purchasing vehicles, GSA MAS, NASPO ValuePoint, and OMNIA Partners among them, are increasingly the default path for AI tool acquisition rather than a fallback to a standalone RFP. And the specific evaluation criteria appearing in AI-related solicitations have converged around a recognizable set: data sovereignty, hallucination and accuracy controls, accessibility conformance, and vendor model-change notification. This report walks through what we've observed on each front, with the methodology and sourcing for every figure stated inline.

How This Report Was Compiled

Two data streams feed this report, and we've kept them separate rather than blending them into a single number wherever that distinction matters. The first stream is internal: deal-stage documentation from Keyspider's own SLED sales and implementation engagements between January 2025 and June 2026, tagged and reviewed by our public sector team as part of normal deal tracking, not collected for the purpose of this report after the fact. That tagging captured, where available, whether a governance review step occurred, which contract vehicle was used, which evaluation criteria appeared in the solicitation or vendor questionnaire, and how long the process took from first RFP contact to signed contract.

The second stream is external: a manual review of publicly posted RFP, RFI, and solicitation documents from SLED entities, pulled from state and local procurement portals, cooperative-vehicle catalogs, and public bid-notice aggregators during the same window. A member of our public sector team read each document and coded it for the presence or absence of the criteria discussed in this report. That coding is inherently judgment-based. Two people reading the same RFP language could reasonably code a borderline clause differently, and we didn't run a formal inter-rater reliability check on this pass. We flag that limitation rather than paper over it.

Neither stream was drawn using random or stratified sampling. The internal stream reflects agencies that chose to evaluate Keyspider specifically, which skews toward agencies already interested in AI search or AI Assistant procurement rather than the full population of SLED entities. The external stream reflects agencies that post detailed solicitation language publicly, which skews toward larger, more procurement-mature entities. Both biases point in a similar direction: this report likely overstates AI-specific procurement rigor relative to the full national population of SLED agencies, many of which have not yet formalized any AI-specific procurement process at all. Where a figure in this report is a percentage, it's a percentage of the sample described at that point, not of all US SLED entities.

Finding 1: AI Governance Review Is Now a Procurement Gate, Not a Parallel Process

As recently as 2023, AI governance review, when it existed at all, ran mostly as a parallel or after-the-fact process: a system got procured and deployed, and a governance committee reviewed it later, if a governance committee existed. That sequencing has largely inverted. Across the SLED procurement processes Keyspider engaged with in 2025 and 2026, a clear majority now require governance review sign-off, or at minimum a completed AI risk assessment questionnaire, before a contract can be finalized, not after deployment.

This tracks with the broader state AI governance framework adoption documented in our companion guide on state and local AI governance: more than 30 US states have issued AI executive orders or governance frameworks since 2023, and most of those frameworks are now in their enforcement phase, with concrete requirements attached rather than aspirational language. The practical effect on procurement: an AI vendor selling into SLED in 2026 without governance-ready documentation, mapped to frameworks like the NIST AI Risk Management Framework, faces a materially longer sales cycle than one that arrives with that documentation prepared.

Source and methodology note: this finding is based on direct observation across Keyspider's customer procurement engagements during 2025–2026 (approximately two-thirds of tracked deals included a formal governance review step prior to contract execution) plus review of publicly posted RFP language requiring AI risk assessment or governance committee approval as a condition of award. We have not attempted to extrapolate a precise national percentage from this sample; the finding here is the direction and consistency of the trend, not a specific census figure.

What the Governance Gate Actually Asks For

The specific documentation governance reviewers request has become fairly consistent across the frameworks we've reviewed, even though the frameworks themselves vary state by state. Reviewers most commonly ask for: a description of what decisions the AI system makes or informs (and confirmation it does not make consequential decisions about individuals without human review), a data handling and protection summary, a bias and equity assessment, an accessibility conformance statement, and a description of human oversight and escalation mechanisms.

Vendors that can answer these five questions with prepared, specific documentation move through review meaningfully faster than vendors answering for the first time in a live meeting. This is a source-level observation from Keyspider's own SLED sales and implementation team, not a third-party statistic, and we present it as such.

Finding 2: Cooperative Purchasing Vehicles Are Displacing Standalone RFPs for AI Tools

The second clear pattern: agencies are increasingly acquiring AI search and chat tools through existing cooperative purchasing vehicles, GSA Multiple Award Schedule (MAS), NASPO ValuePoint, and OMNIA Partners chief among them, rather than running a standalone competitive RFP specific to the AI tool. This isn't unique to AI procurement; cooperative vehicles have been growing in SLED technology procurement generally for years. But the shift appears more pronounced for AI tools specifically than for general SaaS categories.

Two forces plausibly explain this, based on conversations Keyspider's team has had directly with procurement officers evaluating our products. First, agencies moving on AI initiatives are frequently working against a compressed internal timeline, driven by budget-cycle deadlines, leadership pressure to show AI progress, or a specific pain point (a call center overwhelmed, a PDF compliance deadline) that doesn't leave room for a 6 to 9 month standalone RFP cycle. Cooperative vehicles, having already completed competitive solicitation at the vehicle level, let agencies move in weeks rather than months. Second, procurement officers report, anecdotally and consistently in our conversations, more comfort with cooperative vehicle purchases for a genuinely new technology category like AI, because the vehicle-level vetting reduces some of the individual agency's exposure on a category still building precedent.

Source and methodology note: the table above reflects Keyspider's own observed deal mix across SLED AI Search and AI Assistant engagements in 2025 and 2026, supplemented by publicly available NASPO ValuePoint participating-state disclosures. It does not cover every SLED AI procurement nationally, and a vendor with a different customer base and geographic footprint would likely see a different mix.

What This Means for Agencies Evaluating AI Tools

For a digital services director or IT procurement officer starting an AI search or chat evaluation, the practical implication is straightforward: ask every vendor under consideration which cooperative vehicles they're available through before assuming a full RFP is the only path. Confirming vehicle availability and the applicable SIN or contract number early in the evaluation can meaningfully shorten the timeline between decision and deployment. Our companion guide on improving government website search without a redesign covers this procurement path in more detail for search-specific evaluations.

What Evaluation Criteria Show Up Most Often in AI-Related RFPs?

Across the publicly posted AI-related RFPs and RFIs Keyspider reviewed from 2025 and 2026, four evaluation criteria appear with enough consistency to be called a pattern rather than a coincidence: data sovereignty and residency, hallucination and accuracy controls, accessibility conformance, and vendor model-change notification. These four didn't appear as consistently, or in some cases at all, in comparable solicitations from three to four years earlier.

Data Sovereignty and Residency

The majority of AI-related RFPs and RFIs Keyspider reviewed in 2025–2026, by our internal estimate more than 70%, included some form of data sovereignty, data residency, or model-training-use clause: a requirement that agency data (search queries, chat conversation logs, indexed content) not be used to train the vendor's underlying models, and often a requirement that data be stored within the United States or a specific state boundary. This is a substantial shift from solicitation language typical several years ago, which rarely addressed AI-specific data use at all, because most agencies weren't yet contracting for generative AI capabilities.

Hallucination and Accuracy Controls

For any AI Assistant or chatbot procurement specifically, evaluation criteria increasingly ask vendors to describe their approach to grounding responses in verified source content and preventing fabricated answers, rather than accepting a general claim of 'high accuracy.' Questions we've seen appear directly in RFP language include requests for the vendor's approach to source citation, confidence scoring, and escalation to a human when the system cannot answer reliably. This tracks with growing procurement-side awareness of generative AI's hallucination risk as a genuine deployment concern rather than a theoretical one, discussed further in our post on AI Assistant hallucination prevention.

Accessibility Conformance

WCAG 2.1 AA conformance requirements, already standard in general government website procurement, now increasingly extend explicitly to AI-generated interface elements: answer panels, chat widgets, voice or conversational interfaces. Several RFIs Keyspider reviewed specifically asked vendors to describe how AI-generated content is announced to screen reader users, a question that simply didn't appear in pre-AI search procurement documents, because the underlying interaction pattern didn't exist yet.

Vendor Model Change Notification

A newer and less universal but clearly growing criterion: contract language requiring the vendor to notify the agency before making material changes to the underlying AI model or its behavior, and in some cases providing a testing window before changes go live in production. This reflects a lesson many agencies have learned the hard way, that an AI model update from a vendor can change system behavior in ways an agency didn't anticipate and didn't get to test first.

What Does the Actual RFP Language Look Like?

Reading the four criteria above as category labels understates how specific solicitation language has gotten. Agencies aren't writing "vendor must address AI risk" anymore. They're writing clauses with defined terms, named standards, and audit rights attached. What follows is a sample of the phrasing patterns Keyspider's team has seen repeated, in substance if not word for word, across multiple 2025 and 2026 SLED solicitations. We've paraphrased and generalized each example to avoid quoting any single agency's document verbatim, consistent with this report's anonymization approach.

Data Sovereignty and Model-Training-Use Clauses

The blunt version of this clause, common in 2023 and earlier solicitations that addressed the topic at all, simply asked whether data would be stored in the United States. The 2025-2026 version goes further. Recurring elements include: a requirement that agency data never be used to train, fine-tune, or improve any model shared with other customers; a requirement that the vendor specify whether any subprocessor or underlying model provider (naming the specific foundation model vendor, in some cases) receives agency data and under what terms; and, in a smaller but growing share of solicitations, a requirement that data reside within a named state's borders specifically, not just within the US generally. A few RFIs from state agencies have started asking vendors to describe their data deletion process on contract termination, including confirmation that data isn't retained in backups beyond a stated window.

Hallucination, Grounding, and Accuracy Clauses

For AI Assistant and chatbot solicitations specifically, we've seen evaluation criteria move from a single yes/no question about accuracy to a multi-part requirement. Common components: a requirement that every generated answer cite the specific source document or page it was drawn from; a requirement that the vendor describe what happens when the system has no confident answer, does it say so, or does it guess; and, less commonly but increasingly, a requirement for a sample accuracy or grounding rate against a test set of agency-representative questions, evaluated during a proof-of-concept or pilot period rather than taken on faith from vendor marketing claims. This is precisely the ground our AI Assistant architecture was built to hold, grounding every generated answer in indexed agency content rather than open-ended model output.

Accessibility Requirements Specific to AI Interfaces

Beyond the general WCAG 2.1 AA conformance statement now standard in most government technology RFPs, AI-specific accessibility language increasingly asks: how AI-generated answer content is announced to assistive technology (a live region versus a page reload, for instance); whether a conversational interface can be operated entirely by keyboard; and whether generated content maintains a stable reading order when an answer updates dynamically. One pattern worth flagging for vendors: a handful of RFIs have started asking whether the AI system's own quality-control or content-review tooling is itself accessible to agency staff who use assistive technology, a second-order accessibility question that essentially didn't exist in solicitation language before 2025.

Model Change Notification and Audit Rights

This is the newest of the four criteria and the least standardized in phrasing, but a consistent shape is emerging. Agencies ask for advance written notice, 30 days is a commonly requested window, before the vendor deploys a material change to the underlying model or its configuration in the agency's production environment. Some solicitations pair this with a right to test the updated system in a staging or sandbox environment before the change goes live. A smaller number go further and request audit rights: the ability for the agency, or a third party the agency designates, to review system logs, decision records, or model behavior on request, generally tied to compliance investigations or public records requests rather than routine monitoring. Vendors should expect this clause to keep expanding in scope through 2026 and beyond, not contract.

Source and methodology note: the language patterns above are paraphrased composites drawn from Keyspider's review of publicly posted SLED RFP and RFI documents in 2025 and 2026, plus vendor questionnaires our own sales team completed directly for prospective customers. No single row quotes one specific solicitation verbatim. The "government tier most likely" column reflects our team's qualitative read across the documents reviewed, not a statistically weighted tally.

A pattern worth naming on its own: these four criteria increasingly show up together rather than in isolation. Solicitations that ask about data sovereignty are, in our review, considerably more likely to also ask about model change notification than solicitations that skip the data sovereignty question, suggesting a growing baseline of procurement-side AI literacy that raises all four questions in tandem rather than one agency happening to focus on privacy and another happening to focus on accuracy. Put differently: agencies asking the right first question are learning to ask the rest of them too. A state benefits agency that added a data-training-use clause to its 2025 RFP template, for instance, added a model-change notification clause to the same template within roughly a year, based on the revised solicitations Keyspider reviewed for that agency type during the period.

How Long Does AI Procurement Actually Take?

Ask a procurement officer how long AI tool acquisition takes and the honest answer is: it depends heavily on the path. Keyspider's own deal-cycle data from 2025-2026 shows a wide spread, and the spread itself is informative. Deals that moved through an established cooperative vehicle with a governance-ready vendor closed, on average, in a matter of weeks from first serious evaluation to signed contract. Deals that required a standalone competitive RFP, particularly ones that also triggered a first-time AI governance review process at the agency, routinely stretched past six months, and a meaningful share ran close to a full year when a budget cycle or a leadership transition intervened.

That's a wider spread than we typically see in general-purpose SaaS procurement, where cooperative-vehicle and standalone-RFP timelines tend to converge more, usually somewhere in the two-to-four-month range regardless of path. AI procurement hasn't converged yet, largely because governance review is still a variable-length process rather than a standardized checklist at most agencies. An agency running its first-ever AI governance review alongside its first-ever AI procurement is, in effect, building the plane while flying it, and that adds weeks to months that a mature, second-time buyer doesn't incur.

What Slows an AI Procurement Down

Three friction points recur across the deals Keyspider's team has worked through in 2025 and 2026. The most common: governance review sign-off arriving late in the process because the agency didn't know it needed one until partway through vendor evaluation, forcing a restart of part of the timeline. Second: legal review of data processing and model-training-use terms taking longer than expected, particularly at agencies without in-house counsel experienced in AI-specific contract language, who sometimes bring in outside counsel for that review alone. Third, and more mundane but just as real: budget authorization timing. An agency that identifies a strong AI vendor fit in October but doesn't have budget authority until the new fiscal year in July effectively has a nine-month wait built in regardless of how fast the procurement process itself moves.

What Speeds an AI Procurement Up

The inverse patterns are just as consistent. Vendors who arrive with governance documentation already mapped to the agency's framework, rather than producing it on request mid-review, routinely save weeks. Cooperative vehicle availability with the specific SIN or contract number in hand at the start of evaluation, rather than discovered midway through, removes an entire negotiation step. And agencies that have been through at least one AI procurement before move measurably faster on the second one. The governance review process itself, once built, becomes a repeatable checklist rather than a from-scratch exercise, and that experience compounds. A state agency or large county on its third AI procurement cycle in 2026 looks, procedurally, more like a mature SaaS buyer than an agency of the same size buying its first AI tool did in 2024.

Source and methodology note: timeline figures are drawn from Keyspider's internal deal-stage tracking across SLED AI Search and AI Assistant opportunities between January 2025 and June 2026. "Closing timeline" is measured from first substantive procurement contact (RFP release, RFI response, or vehicle-based quote request) to signed contract. This reflects Keyspider's own sales cycle experience and should not be read as a universal benchmark for all AI vendors or all SLED buyers; a different product category, price point, or customer mix would likely produce a different distribution.

How Does Adoption Differ Between State, County/City, and Education Government?

AI search and chat adoption patterns differ meaningfully by government tier, driven by differences in procurement authority, budget cycles, and the specific pain points each tier faces most acutely.

State Government

State agencies show the most formalized governance review processes, unsurprising given that most state-level AI executive orders and frameworks apply directly to state agencies first, with county and municipal adoption sometimes following the state's lead voluntarily. State procurement also shows the heaviest use of statewide contract vehicles and centralized IT procurement offices, meaning individual agency AI purchases increasingly route through a central state technology office rather than being negotiated independently by each department. This centralization can slow initial adoption but tends to produce faster subsequent purchases once a vendor is established on the statewide vehicle.

Within state government, the pattern splits further by function. Health and human services agencies, which handle the most sensitive personal data of any state function, show the strictest data sovereignty and access-control requirements Keyspider has seen at any tier, often exceeding the state's baseline governance framework with agency-specific addenda. Revenue and tax agencies show similar rigor around auditability. By contrast, state agencies focused on public-facing informational content, tourism boards, general services administration sites, business licensing portals, move faster and with lighter governance overhead, because the underlying data involved carries less individual privacy risk. A state CIO's office overseeing both types of agency will frequently apply a tiered governance model, a light-touch review for informational AI Search deployments and a full risk assessment for anything touching benefits eligibility, health records, or tax data.

State higher education systems occupy an interesting middle position: many report through a state CIO or system-level technology office for procurement purposes, but individual campuses often retain enough autonomy to run their own AI governance review in parallel, producing two review layers rather than one. Vendors selling into a state university system should expect to satisfy both, and budget for the longer of the two timelines rather than assuming state-level approval clears the path at the campus level.

County and City Government

County and municipal government adoption is more fragmented and, in our observation, faster-moving at the individual-agency level, because procurement authority is more distributed and a single department (a call center, a permitting office) can often move on a discrete AI tool without a lengthy centralized process. County and city AI procurement we've observed skews toward specific, measurable pain points, call deflection, permit processing, multilingual access, rather than broad AI strategy initiatives. Governance review at this tier varies enormously: some large counties have formal AI governance committees comparable to state-level rigor, while many smaller municipalities have no formal AI governance process at all yet, relying instead on general IT security and procurement review.

Population size correlates with procurement formality at this tier more directly than it does at the state level, in our experience. Counties and cities serving fewer than roughly 100,000 residents typically run AI procurement through the same general IT purchasing process used for any software, with no AI-specific questionnaire or governance step, and often complete a purchase in a matter of weeks once budget is approved. Mid-sized counties and cities, roughly 100,000 to 500,000 residents, increasingly have at least a lightweight AI review checklist, sometimes borrowed or adapted from state guidance rather than built independently. The largest counties and cities, serving populations comparable to a mid-sized state, tend to have governance processes indistinguishable in rigor from state agencies, including dedicated AI policy staff and formal risk assessment templates.

One pattern specific to this tier worth naming directly: county and city procurement officers describe more pressure from elected officials and constituent complaints than state or education buyers do, which shows up as compressed timelines when a specific incident, a viral complaint about a broken 311 system, a public records backlog covered by local media, forces a purchase to happen faster than the agency's normal process would otherwise allow.

K-12 and Higher Education

Education government shows the widest split between K-12 and higher education. Higher education institutions, particularly larger research universities, often have more procurement sophistication and more established AI governance structures (frequently housed within a CIO's office or a dedicated AI task force) than a similarly sized municipal government. K-12 districts vary enormously by size; large districts increasingly mirror county-government procurement patterns, while smaller districts often piggyback on state education department guidance and cooperative vehicles rather than running independent evaluations. FERPA and COPPA considerations add an evaluation layer specific to education that state and local general-purpose agencies don't face in the same way, discussed further in our K-12 Digital Experience Report.

Within higher education, procurement pace and rigor track institution type more than enrollment size alone. Large public research universities tend to run the most involved reviews, often requiring sign-off from a data governance committee, a general counsel's office, and sometimes a faculty senate technology subcommittee, particularly for any AI Assistant deployment touching student records or academic advising. Community colleges, even large ones, generally move faster, with procurement processes closer to the county-government pattern than to the research-university one, and less frequently maintain a dedicated AI governance function separate from general IT security review.

K-12 procurement carries a distinct constraint the other tiers don't face to the same degree: the school-year calendar. Districts strongly prefer to complete procurement and deployment during summer break, before a new academic year starts, which compresses the effective procurement window into a few months even when the underlying process could technically run longer. A district that misses the summer window often waits for the next one rather than deploying mid-year, a pattern Keyspider's K-12 team has seen repeatedly enough to plan sales cycles around it directly. Our companion research on higher education digital experience trends covers campus-specific findability and AI adoption patterns in more depth; our guide on multilingual government access is relevant reading for districts serving high populations of English-language-learner families.

What's Driving the Shift Toward Cooperative Purchasing Specifically for AI?

It's worth pausing on why AI tools specifically, more than general SaaS categories, appear to be moving toward cooperative vehicles. Three explanations came up repeatedly in conversations Keyspider's sales and public sector teams have had directly with SLED procurement officers during 2025 and 2026.

First, urgency. A department facing an ADA Title II compliance deadline, an overwhelmed call center, or a public records backlog often doesn't have 6 to 9 months to run a full competitive RFP before the pain point becomes acute. Cooperative vehicles compress procurement timelines meaningfully, sometimes from months to weeks, which matters disproportionately for AI purchases tied to a specific, dated pressure.

Second, category newness. Procurement officers evaluating a genuinely new technology category report more comfort relying on the vetting already performed at the cooperative-vehicle level, rather than building AI-specific evaluation criteria from scratch inside a standalone RFP, particularly for agencies without dedicated AI procurement expertise on staff yet.

Third, budget-cycle alignment. Several procurement officers described AI initiatives as being funded through supplemental or reallocated budget lines rather than planned multi-year procurement budgets, a funding pattern that fits more naturally with the faster cooperative-vehicle path than with a standalone RFP cycle typically planned a full budget cycle in advance.

What Should Agencies Do With These Findings?

For an agency planning an AI search, chat, or accessibility procurement in the next 12 months, three practical actions follow directly from the patterns in this report.

1. 1Confirm your governance review requirement, if one exists, before starting vendor evaluation, and ask every vendor under consideration for governance-ready documentation (NIST AI RMF mapping, data handling summary, accessibility conformance statement) up front rather than discovering the gap mid-process.
2. 2Ask every vendor which cooperative purchasing vehicles they're available through, and get the specific contract or SIN number, before assuming a standalone RFP is required.
3. 3Build your evaluation criteria around the four areas increasingly standard in AI RFPs: data sovereignty and model-training-use restrictions, hallucination and accuracy/grounding controls, accessibility conformance extended to AI-generated content, and vendor model-change notification rights.
4. 4Budget realistic time for governance review if your agency hasn't run one before. Treat the first AI procurement as a process-building exercise, not just a vendor selection, and expect the second one to move considerably faster.
5. 5Ask for a pilot or proof-of-concept period against real agency content before final award, particularly for AI Assistant deployments, so accuracy and grounding claims get tested against your own documents rather than a vendor's demo environment.

None of these steps requires new statutory authority or a governance framework the agency doesn't already have access to. Most state frameworks referenced earlier in this report already supply the risk-assessment structure; the gap we see most often isn't a missing framework, it's a procurement team that hasn't yet connected its existing framework to its existing RFP template. That's a same-quarter fix for most agencies, not a multi-year initiative.

Our SLED AI Search Procurement Checklist translates these criteria into 52 specific vendor questions, and our Government AI Chatbot Procurement Checklist does the same for conversational AI specifically, including the data sovereignty and hallucination-control questions this report identifies as increasingly standard.

Limitations of This Report

This report reflects Keyspider's vantage point as a vendor active in SLED AI Search and AI Assistant sales during 2025 and 2026, supplemented by review of publicly available solicitation language. It is not a neutral, third-party survey of the entire SLED market, and the specific percentages cited should be read as directional patterns observed from our position in the market rather than a precise national census. Agencies making procurement decisions should validate current requirements with their own procurement office and legal counsel, particularly on fast-moving areas like AI governance framework enforcement dates, which vary by state and continue to be updated.