MortgageGuide: ABT's Mortgage-Specialized Copilot Agent vs Generic Microsoft 365 Copilot

Generic Microsoft 365 Copilot is a productivity unlock for everyone in your shop. Loan officers, processors, underwriters, closers, ops staff, the CFO, the marketing team. It reads emails. It drafts letters. It summarizes Teams meetings. It builds Excel models. It does all of this by reasoning over your institution's own tenant content through Microsoft Graph, which is why the productivity case lands so cleanly at most community banks, credit unions, and mortgage lenders.

But the moment a loan officer asks Copilot a question only the Fannie Mae Selling Guide can answer, generic Copilot has no opinion. It will summarize a borrower email. It will not tell you whether a self-employed borrower with two years of K-1s and a recent S-corp election qualifies under the current Selling Guide rules. It will not tell you whether a partial claim on a VA file has cured. It will not tell you which Mortgagee Letter superseded the FHA borrower-eligibility rule the underwriter used last quarter. Generic Copilot reasons over what your institution wrote into its own emails and SharePoint files. It does not reason over the GSE and agency guideline corpus that mortgage lending actually runs on.

ABT MortgageGuide Copilot is the agent built for that second question. It is currently in beta, available exclusively to ABT customers, and built on Microsoft Azure AI Foundry. The hero promise on the cluster page reads simply: "Will the loan close? Submit it right." Cited guideline answers in seconds, beyond generic Microsoft 365 Copilot. This article walks through what that means for a real mortgage workday, where the data sits, why citing the specific guideline section is the difference between a clean repurchase-defense file and a defect on a Fannie Mae QC review, and how ABT customers get on the beta list.

Spoke 4 of our Copilot cluster sits next to the Microsoft 365 Copilot Business Buyer's Guide for Financial Institutions (the pillar). The pillar covers the buying decision across community banks, credit unions, and mortgage lenders. This spoke focuses on what mortgage lenders get on top of generic Copilot when they run their adoption through ABT. If you want the role-by-role view of how loan officers, processors, and underwriters use generic Microsoft 365 Copilot Business in their day-to-day, see Spoke 3 on mortgage roles. If you want the pricing math (the $10 incremental bundle, the $18 promotional standalone, and the $21 standard standalone that takes effect July 1, 2026), see Spoke 1 on pricing for community banks and credit unions. This spoke is about the specialist agent that runs on top of all that.

How MortgageGuide Saves a Loan Officer's Morning

A loan officer at a community lender opens her laptop at 8:15 a.m. Three loans on her desk. The first is the easy one: a W-2 borrower with two years on the job, 740 FICO, 20 percent down, conventional conforming. She submits without thinking about it. The other two are the question marks. One is a self-employed borrower whose tax returns just got produced last night, with a recent S-corp election that changes the income calculation. The other is a VA cash-out where the appraisal came in lower than expected and the borrower wants to know whether the file still works at the new value. Both of these are questions only the GSE or agency guideline can definitively answer, and both of them have to be answered before 9:30 a.m. because the borrower is going to call.

In a generic Microsoft 365 Copilot workflow, the loan officer pivots to a browser tab, opens the Fannie Mae Selling Guide at selling-guide.fanniemae.com, searches for self-employment, scans through five subsections under Part B3-3.5 (calculation of self-employed income), opens a second tab for the relevant Lender Letter or Selling Guide Announcement that may supersede or qualify the rule, opens a third tab to check whether her institution's overlays apply, and starts piecing together an answer. The S-corp election changes which Form K-1 entries the underwriter pulls and how the qualifying income is averaged. Twenty minutes after the first tab opens, she still has not landed on a defensible answer; she calls a senior underwriter to confirm. That is forty-five minutes the loan officer is not on the phone with the next borrower, and the question still has to be re-asked by the underwriter when the file gets submitted.

MortgageGuide's productivity outcome is the time spent between the question and a defensible answer. The agent is a chat experience the loan officer can use directly in her Microsoft 365 surface; she asks a question in the same place she asks generic Copilot a question. The agent is currently in beta and available exclusively to ABT customers. The platform supports citation behavior through Microsoft Azure AI Foundry's managed knowledge layer, so answers surface alongside the underlying guideline source rather than as freestanding paragraphs the loan officer has to verify by hand. The loan officer never leaves the Microsoft 365 surface to read a separate website; the senior underwriter is reserved for the genuinely ambiguous files; the borrower gets a same-morning answer.

Specific behaviors of the beta agent (response format, exact citation style, refresh cadence on the guideline corpus, integration with the institution's loan origination system) are part of the work currently under engineering review under Jira LEAD-2245 and will be communicated to ABT customers as the beta progresses. ABT customers participating in the beta get update notes through their account team and through the engagement's weekly office hours.

The same morning workflow plays out for processors gathering conditions, for underwriters writing decisions, and for closers confirming the file is ready for funding. Each role asks different questions; each role spends less time in browser tabs and more time in the institution's loan origination system. A processor asks the agent about the documentation required to clear an income condition. An underwriter asks the agent to summarize the qualifying-income calculation for the credit memo. A closer asks the agent about the seasoning requirement on a VA cash-out. The questions are not the same; the productivity dynamic (cited answers in the same surface where the work already lives) is. For the role-by-role view of how the broader Copilot Business surface helps these mortgage roles, see Spoke 3: How Mortgage Roles Actually Use Microsoft 365 Copilot Business. This spoke focuses on what the mortgage-specialized MortgageGuide agent adds on top.

The two-column comparison above is the underlying logic of why ABT built MortgageGuide on top of Microsoft 365 Copilot rather than as an alternative to it. The everyday productivity workflow that runs on tenant content stays on generic Copilot, where it belongs. The mortgage-guideline workflow that requires reasoning against a corpus the tenant does not contain runs on MortgageGuide, which is what the agent is built to do.

One more thing about the morning workflow that matters at scale: across a 25-loan-officer mortgage lender, the cumulative cost of the guideline-question tax (the morning browser-tab marathon described above) compounds quickly. Conservative arithmetic: if every loan officer averages even one hard guideline question per day that costs 30 minutes of cross-referencing plus a 5-minute senior-underwriter call, that is 14.6 hours of mortgage-team time per day across the shop, or roughly 3,650 hours per year. That number is a back-of-envelope arithmetic example to illustrate the shape of the cost, not a benchmark or a savings guarantee. The actual productivity outcome at any given institution depends on the loan mix, the staffing model, the institution's own overlays, and how the team integrates the agent into the workflow. The point of the arithmetic is the scale: a meaningful share of a mortgage shop's daily time goes to guideline reasoning, and any reduction in the friction of that reasoning lands as immediate top-of-the-day productivity recovery.

The Azure AI Foundry Platform Underneath

Microsoft Azure AI Foundry is Microsoft's enterprise platform-as-a-service offering for building, deploying, and governing AI agents and applications. It is a code-first platform (C# and Python software development kits) that runs inside the customer's Azure subscription, with unified role-based access control, networking, and policy under a single Microsoft resource-provider namespace. The unified resource provider means an institution's IT team sees Foundry agents in the same Azure portal where they see virtual networks, storage accounts, key vaults, and Azure Monitor. There is no separate management plane; there is no separate identity store. The identity that authenticates a loan officer to her Microsoft 365 mailbox is the same Microsoft Entra ID identity that authenticates her to any Foundry agent the institution has licensed.

Foundry is distinct from Microsoft Copilot Studio, which is the low-code and no-code SaaS experience for building flows inside Microsoft 365 and Microsoft Power Platform. Copilot Studio is where most institutions build their first conversational experiences: an internal HR-policy chat, an IT helpdesk triage bot, a customer-service flow that answers branch-hours questions. Foundry is built for the harder cases. It supports three documented customization paths: Supervised Fine-Tuning (training the model on a labeled dataset specific to the institution's domain), Direct Preference Optimization (training against preferences), and Reinforcement Fine-Tuning (training against a reward signal). It supports Retrieval-Augmented Generation (RAG) "when you need answers grounded in private or frequently changing data," per Microsoft's published Azure AI Foundry RAG concepts documentation. Mortgage-guideline reasoning is the canonical RAG use case: the corpus changes monthly with each Selling Guide update, with each Freddie Mac Bulletin, with each Mortgagee Letter, and the agent must reason against the current state of the corpus rather than against frozen training data.

Foundry IQ, the managed knowledge layer inside Azure AI Foundry, provides citation-backed answers that ground agent responses in enterprise or web content. The retrieval pipeline decomposes complex questions into sub-questions, executes parallel subqueries against the indexed corpus, semantically reranks the candidate passages, and returns unified responses with source attribution back to the underlying documents. The pipeline is agentic in the sense that the agent reasons about which subqueries to issue and how to combine the results rather than performing a single static retrieval. This is the public Microsoft capability ABT uses to enable MortgageGuide's guideline-aware behavior. Specific tuning of MortgageGuide on top of Foundry IQ (which guideline corpora are indexed, what the refresh cadence is for each, how citations are rendered to the loan officer, which subquery patterns are used for mortgage-specific reasoning) is part of the beta work currently under review and will be communicated to ABT customers as the beta progresses.

The enterprise controls that come with the Foundry platform are the same controls financial-services examiners expect to see on any production workload that touches customer non-public information. Authentication for Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication, Azure Policy integration that lets the institution enforce specific compliance baselines on the agent's runtime, Microsoft Entra ID role-based access control for who can invoke the agent and what they can ask, Azure Key Vault integration for secret management, private endpoints that keep agent traffic off the public internet, and Azure Monitor observability for audit and incident response. These are not MortgageGuide-specific features; they are platform features the agent inherits by virtue of running on Foundry. An institution's IT lead can map them directly to the FFIEC IT Examination Handbook control families the examiners already know how to interpret.

For the institution's IT lead, the practical answer is that MortgageGuide sits on a Microsoft-published enterprise platform with the same Microsoft Entra ID identity, the same Azure tenant boundary, and the same governance controls as the rest of the institution's Azure footprint. ABT manages your Microsoft 365 tenant as your Tier-1 Microsoft Cloud Solution Provider; ABT hosts the Azure infrastructure that runs MortgageGuide. Per Microsoft's CSP terminology, those are two different verbs on purpose. Microsoft owns the Microsoft 365 service infrastructure; CSPs like ABT operate the tenant on the customer's behalf via delegated administration. Azure is different: the subscription is customer-controlled cloud infrastructure that the CSP operates as the partner of record, which is why the verb shifts from "manage" to "host" for the Azure-resident workload.

For the auditor, the practical answer is that MortgageGuide is not a black-box vendor SaaS. It is an application running on a Microsoft enterprise platform with the same control surfaces (Entra ID Conditional Access, Microsoft Purview Audit, Azure Monitor, Defender for Cloud) the institution already documents in its IT examination response. The examiner conversation about MortgageGuide is not a brand-new conversation; it is the same conversation about any Azure-hosted workload.

The reason this corpus matters for an AI agent specifically (rather than for a search tool or a static reference) is that the questions a loan officer asks are rarely single-source. A self-employment qualifying-income question may touch the Fannie Mae Selling Guide section, a recent Selling Guide Announcement that updated the calculation, an FHA Handbook 4000.1 section if the file is FHA, the institution's own overlay policy stored in SharePoint, and the underwriter's prior notes on similar files. The job of the agent is to combine those sources, ground each piece of the answer in its specific citation, and surface the combined reasoning to the loan officer in the productivity surface where the work already lives. The agentic-retrieval pipeline inside Foundry IQ is built for exactly this kind of multi-source decomposition. The mortgage tuning on top of it is what ABT brings to the engagement.

Where MortgageGuide's Tenant Data Lives

The data question is the second thing a mortgage CTO asks after the productivity question. Where does MortgageGuide put my borrower's tax returns when the loan officer asks a question about self-employed income? Where do the prompts go? Where do the answers go? Who reads any of it? What happens to a borrower's information if an examiner later asks the institution to reconstruct the AI-assisted reasoning on a specific file?

The platform answer is straightforward because Azure AI Foundry runs inside the customer's Azure subscription. Tenant data the agent reasons over stays within the customer's Microsoft tenant boundary. Prompts and answers are subject to the institution's Microsoft Purview Audit configuration and Microsoft Entra ID Conditional Access policies, the same way generic Microsoft 365 Copilot prompts are. The audit log captures who asked what, when they asked it, and what the agent returned. The Conditional Access policy can require multi-factor authentication, a managed device, a particular network posture, or any other condition the institution applies to other Microsoft 365 workloads, and the agent respects those conditions on every invocation. Microsoft does not train foundation models on Microsoft 365 Copilot or Azure AI Foundry customer tenant data, per the Microsoft Product Terms and the published commitments at Microsoft Learn. Those are platform-level guarantees that apply to any agent built on the Azure AI Foundry platform, including MortgageGuide.

The mortgage-specific tuning happens inside the same boundary. When a loan officer asks MortgageGuide whether the self-employed borrower's K-1 income qualifies, the agent grounds its reasoning against GSE and agency guidelines (the high-level corpus is confirmed; specifics of which guidelines, which versions, and which refresh cadence are part of Hugo Bispo's public-claim review under LEAD-2245 and will be communicated to ABT customers as the beta progresses). The borrower's tax returns never leave the institution's tenant. The guideline corpus is the agent's knowledge layer; the institution's tenant content is the agent's context layer. The two stay in their lanes.

For the institution's data-protection officer, the implication is that the same data-protection regime the institution already runs against Microsoft 365 covers the agent. Microsoft Purview Information Protection sensitivity labels follow content into the agent's reasoning surface. Microsoft Purview Data Loss Prevention policies that block customer non-public-information patterns from leaving the tenant apply to the agent's output the same way they apply to email. Microsoft Defender for Cloud Apps shadow-AI discovery covers the agent traffic as a known sanctioned application, which is the opposite of the shadow-AI risk profile.

The tenant-residency picture also matters for examiner response. When the FFIEC, the OCC, or the NCUA asks during an exam how the institution governs AI assistance in lending decisions, the institution's answer becomes a documented configuration story: Microsoft Entra ID Conditional Access on the agent's access path, Microsoft Purview Audit on the agent's prompts and outputs, Microsoft Purview Information Protection sensitivity labels on the source content, Microsoft Defender for Cloud Apps on the agent traffic discovery, and Microsoft Azure Monitor on the agent's runtime. None of that is MortgageGuide-specific. All of it is the same control story the institution tells about the rest of its Microsoft 365 surface.

Why Citing the Fannie or Freddie Section Matters at Audit

An underwriter writes a decline narrative on a self-employed borrower's file. Two versions of that narrative exist. Version one reads: "Borrower's income is insufficient based on tax return analysis." Version two reads: "Borrower's two-year average self-employment income calculated per Selling Guide section B3-3.5-01 results in qualifying monthly income of $X, which produces a debt-to-income ratio of Y percent, exceeding the program threshold of Z percent." The first narrative is a sentence. The second narrative is an examiner-defensible explanation of the underwriting decision with the specific Selling Guide section cited. Three years later, when Fannie Mae's QC review pulls the file, the second narrative passes and the first narrative is a finding.

Fannie Mae's defect taxonomy in Selling Guide Part B3 sorts loan defects into five buckets: Income, Assets, Credit, Property, Liabilities. The 2025 ACES Quality Management Mortgage QC Industry Report places critical defect rates on GSE loans in the 2.3 to 2.6 percent range, with Income contributing roughly 27 to 28 percent of critical defects and Assets contributing roughly 22 to 24 percent. The mathematics of this is unforgiving: at a 2.5 percent critical defect rate on, say, 600 loans funded in a year, fifteen loans land in the critical category. Defects in the rep-and-warranty window can drive repurchase demands from the GSEs, which means full principal-plus-interest exposure on the institution's balance sheet for every loan that fails the QC review. A single repurchase on a $400,000 conforming loan is a six-figure hit to the income statement. Repurchases also drive insurance and counterparty consequences that compound the underlying loss.

The Consumer Financial Protection Bureau and federal prudential regulators have separately emphasized explainability, auditability, and Fair Lending compliance for any AI used in credit decisions, which makes citation-backed answers an industry baseline rather than a nice-to-have. The April 17, 2026 joint guidance from the OCC, Federal Reserve, and FDIC explicitly excludes generative and agentic AI from the formal Model Risk Management scope but requires enterprise risk governance over those tools. Translated into plain language: examiners do not expect MortgageGuide to be governed as a regression-style credit model under MRM, but they do expect the institution to document how the tool fits into its broader risk and lending operations, including its review and exception-handling processes. Citation-grounded answers make that documentation possible because the answer itself is a self-documenting reference back to the source-of-truth guideline.

The agent platform supports citation behavior through Foundry IQ's agentic-retrieval design. That is the public Azure AI Foundry capability that grounds responses with source attribution back to the indexed corpus. How MortgageGuide renders citations to the loan officer or underwriter, which guideline sections appear in the response, how the citation depth is calibrated for different question types, and how the response format ties back to the institution's loan origination system are part of the beta work currently under engineering and product review. ABT customers participating in the beta receive updates as the agent matures.

One more boundary worth naming clearly: the Equal Credit Opportunity Act (12 CFR section 1002.9) requires creditors to provide specific principal reasons for adverse-action decisions within 30 days of an application, subject to fair-lending review. Copilot (generic or specialized) cannot be the final author of an ECOA decline narrative. The underwriter retains authorship and accountability. MortgageGuide is a first-draft assistant for the underwriter, the same way generic Copilot is a first-draft assistant for the loan officer's borrower email. The role of the human stays the role of the human. Examiners ask specifically about this division of labor in the lending exam workflow, and the institution's documented response should make clear that AI assistance is a productivity layer, not a decision-making layer, on every adverse-action file.

The defect taxonomy itself rewards specificity. A QC reviewer reading an underwriter narrative that cites Selling Guide section B3-3.5-01 for self-employed income calculation has a fast path to validate the underwriting decision: open the cited section, verify the calculation, sign off. A QC reviewer reading "borrower's income was analyzed and qualifying income was determined" has to do the entire calculation from scratch, second-guess the underwriter's logic, and decide whether to escalate. The first path takes minutes and produces a clean file; the second path takes hours and risks a downgraded loan finding. The citation is the speed gear for the QC team, not just the defensibility gear for the institution.

How ABT Customers Get MortgageGuide Beta Access

MortgageGuide is currently in beta and available exclusively to ABT customers. Two practical paths exist for institutions evaluating it, and both paths run through the broader Copilot adoption engagement rather than as a standalone agent procurement.

Existing ABT Microsoft 365 customers can request beta access through their account team as part of an active or planned Copilot training engagement. ABT runs Copilot adoption as a productized 30-day sprint with five components: a kickoff workshop with the institution's leadership and IT lead on Day 1, champion enablement for two to four internal champions per role on Days 2 through 5, a role-based scenario library adapted to the institution's actual templates on Days 6 through 10, weekly office hours running Days 11 through 25 to fix adoption blockers in real time, and a Day 30 executive ROI readout. MortgageGuide beta access folds into the sprint where the institution's roles benefit from guideline-aware grounding, primarily loan officers and underwriters at mortgage lenders and at the mortgage divisions of community banks and credit unions. The beta is not separately licensed; it travels with the Copilot adoption engagement for institutions that fit the mortgage-specialization use case.

New customers qualify for the Copilot Pilot Pack through the broader ABT offer that runs through June 30, 2026. The qualifier is adding 11 or more new Copilot seats through ABT. Microsoft allows multiple Cloud Solution Provider partners to coexist on the same tenant, so ABT becomes an additional CSP alongside any existing CSP. There is no license transfer, no tenant migration, no Microsoft fees, and no break in the institution's existing relationship with its current CSP. ABT activates the new Copilot Business licenses against the new seats for the pilot cohort. The institution gets the AI Readiness Assessment, the 30-day adoption sprint, and MortgageGuide beta access for mortgage roles, all bundled into the engagement. The multi-CSP-additive model is one of the most important pieces of the procurement story: the buyer does not have to choose between switching providers and getting access to ABT's mortgage-specialist work, because the buyer does not have to switch providers at all.

For the full pricing math (the $10 incremental bundle on Microsoft 365 Business Premium, the $18 promotional standalone, and the $21 standard standalone) and how those map to a 30-person mortgage lender, an 80-person community bank, and a 200-person credit union, see Spoke 1: Microsoft 365 Copilot Business Pricing for Community Banks and Credit Unions. The bundle promotion expires June 30, 2026 and the standalone Copilot Business price increases from $18 to $21 on July 1, 2026, which is a 17 percent list-price increase for institutions that buy standalone rather than through a Microsoft Cloud Solution Provider bundle. The pricing urgency is real but it is independent of MortgageGuide; the agent is currently in beta and available exclusively to ABT customers regardless of which Microsoft 365 Copilot tier the institution is licensed for. Contact ABT for current beta availability based on your institution's plan.

One competitive observation worth making for buyers comparing options: no other Microsoft Cloud Solution Provider publicly markets a dedicated GSE-guideline citation agent as of May 2026. The major mortgage-technology vendors have rolled out platform-level AI capabilities. ICE Mortgage Technology (Encompass) has invested in AI for document processing and rule checking following its 2024 Black Knight acquisition. Calyx Software's Points has rolled out workflow automation. Blend focuses on the digital borrower experience with ML-driven routing. MGIC AI Insights offers risk-scoring sourced from MGIC's historical loan-level data. Tavant offers intelligent process automation for larger lenders. Snowflake Cortex provides LLM and ML primitives layered on mortgage data lakes. These are real capabilities and they are valuable for the workflows they target. But a dedicated mortgage-guideline-trained citation agent marketed specifically to mortgage-lender users (the loan officer, processor, underwriter, closer in their daily morning workflow) is not a broadly announced product in the market today. That is a positioning observation about the landscape, not a comparative benchmark on MortgageGuide itself; the beta is still under engineering review.

If you run mortgage operations, the realistic ask is to compare what your loan officers and underwriters do today (the morning workflow described at the top of this article) against what they could do if the guideline question and the productivity workflow lived in the same Microsoft 365 surface, with cited answers grounded in the GSE and agency guideline corpus. The path from today to that future state runs through generic Microsoft 365 Copilot Business for the everyday productivity unlock (the borrower email drafts, the meeting recaps, the Excel modeling, the document summarization that every shop benefits from) plus ABT MortgageGuide Copilot beta for the mortgage-guideline lift on top. The pillar (Microsoft 365 Copilot Business Buyer's Guide for Financial Institutions) ties the two together for the buying decision and walks through the broader buyer journey from productivity to security to governance for community banks, credit unions, and mortgage lenders alike.

The mortgage-specific story is on top of that pillar. The pillar gets you the productivity floor. MortgageGuide is the ceiling lift for the mortgage workflow specifically. The institution that pairs the two has the cleanest path through the next examination cycle, the next QC review cycle, and the next year of borrower-facing turn-time competition. The path is open today for ABT customers; the door is the AI Readiness Assessment that comes with the Pilot Pack offer. For a step-back view of how the Copilot Pilot Pack fits into the broader productivity / security / governance buyer journey at community banks, credit unions, and mortgage lenders, the Microsoft 365 Copilot Business Buyer's Guide for Financial Institutions walks through all of it.

The table above is the practical answer to "which roles benefit from MortgageGuide." The shortest version: every mortgage-shop role benefits from generic Microsoft 365 Copilot Business for the day-to-day productivity workflow, and the loan officer, underwriter, and QC roles benefit from MortgageGuide on top for the moments when the question depends on the GSE or agency guideline. The closer and the processor benefit from MortgageGuide on the smaller share of their workflow that touches agency rules directly; the larger share of their work runs cleanly on generic Copilot.

For an institution evaluating the engagement, the simplest framing is this: do not buy MortgageGuide as a standalone product. The agent is not licensed standalone today. Buy the Microsoft 365 Copilot Business adoption engagement that ABT runs, which gives every role in the shop the productivity floor. Then, where the mortgage workflow benefits from guideline-aware reasoning (and at most mortgage lenders, mortgage-divisions of community banks, and mortgage-active credit unions, it does), MortgageGuide beta access is included in the engagement at no incremental cost. The arithmetic is straightforward: the institution pays for Copilot Business per seat per month at the standard published rate (the $10 incremental bundle through June 30, 2026, the $18 standalone promo through the same date, or the $21 standalone standard rate that takes effect July 1, 2026, depending on path); the MortgageGuide beta access is a service component of the engagement.

The four-square framing above is the practical "what's next" for a mortgage lender, mortgage-division leader at a community bank or credit union, or technology executive who has read this far. Each scenario maps to the same engagement vehicle (the 30-day Copilot Business adoption sprint) and the same offer window (the Pilot Pack through June 30, 2026, with pricing reverting to standard rates after that). The MortgageGuide beta access folds in for mortgage roles inside the engagement; it does not require a separate procurement, a separate contract, or a separate sale.

One closing observation about the broader competitive picture for buyers reading this in mid-2026. The mortgage-technology vendor landscape is rich with AI capabilities at the platform level: document classification, automated indexing, intelligent stacking-order routing, automated underwriting findings interpretation, fraud-signal detection across loan-level data. Those capabilities serve real workflows and they are not duplicated by what MortgageGuide does. MortgageGuide does not classify documents, does not stack files, does not produce AUS findings. It answers the loan officer's question about whether the guideline applies and cites the section the answer comes from. The role is narrow on purpose. The agent does one thing well rather than ten things vaguely. For a mortgage shop, that means MortgageGuide does not replace the loan origination system, the AUS engine, the document classification platform, the fraud detection vendor, or the QC software. It is a specialist assistant alongside those systems, accessed inside the Microsoft 365 surface where the team already works.

The deployment risk profile reflects that narrowness. Because MortgageGuide is a productivity assistant operating inside the Microsoft 365 surface (rather than a system of record in the loan workflow), the institution's downside if the agent produces an unhelpful answer on any given query is small: the loan officer or underwriter checks the cited guideline section themselves, makes the call, moves on. There is no automated decisioning behind it. There is no submission of the loan to a GSE based on the agent's output without a human in the loop. There is no funding decision rendered by the agent. The agent reduces friction; the human retains decision-making authority on every file. That is the right deployment shape for an AI assistant in a regulated lending context, and it is the shape Microsoft, the federal prudential regulators, and the GSE-counterparty risk teams have all separately landed on through their 2025 and 2026 published guidance.

If you are a CFO reading this, the cost picture is also worth naming clearly. Generic Microsoft 365 Copilot Business is a per-seat-per-month license, billed monthly with no separate AI consumption charge for the agent layer. MortgageGuide beta access is a service component of the ABT engagement and does not introduce a separate per-query, per-seat, or per-agent license at the institution level for participating customers during the beta. The pricing details after the beta exits and the agent enters general availability will be communicated to ABT customers when those details are finalized; the beta itself is structured to be a low-cost entry into the productivity gain rather than a separate procurement cycle. The institution's full Microsoft AI spend through this engagement equals the Microsoft 365 Copilot Business seat license cost, period.

The five examiner questions above are the realistic content of a 2026 IT or compliance examination at a community bank, credit union, or mortgage lender deploying AI assistance in the lending workflow. The clean answer in each row is what the institution's documented response looks like once the Microsoft 365 control surface is configured and the agent is governed under that surface. Nothing in the answer set is MortgageGuide-specific in a way that creates a new exam category; the answers are extensions of the institution's existing Microsoft 365 governance posture.

The same logic applies to the institution's third-party risk management documentation. The institution's existing vendor file on Microsoft (covering the SOC 2 Type II reports, the data-processing addendum, the customer copyright commitment, the published Microsoft Learn data-handling commitments) covers Microsoft Azure AI Foundry the same way it covers Microsoft 365. ABT, as the Tier-1 Microsoft Cloud Solution Provider operating the institution's tenant and hosting the Azure subscription where MortgageGuide runs, is documented under the institution's existing CSP relationship. There is no new vendor onboarding cycle required to add MortgageGuide to the institution's third-party inventory; the agent is already covered by the existing Microsoft and ABT relationships the institution has on file. That is one of the practical advantages of building MortgageGuide on the same Microsoft enterprise stack the institution is already running rather than as a separate vendor SaaS that would require a fresh procurement and a fresh vendor risk review.

For the institution's compliance officer who reads compliance bulletins as a hobby, the regulatory through-line is this: federal prudential regulators have been pushing for explainability, auditability, and Fair Lending compliance on AI used in credit decisions throughout 2024, 2025, and into 2026. The April 17, 2026 OCC/Federal Reserve/FDIC joint guidance is the cleanest articulation: generative and agentic AI sit outside formal Model Risk Management scope but require enterprise risk governance. The CFPB has separately circulated guidance on adverse-action notice obligations under ECOA and on UDAAP risk in AI-driven credit operations. State-level AI laws (Colorado's stayed enforcement, Texas TRAIGA's federal-prudential safe harbor, New York DFS's MFA and deepfake-resistant authentication requirements) form the secondary regulatory layer. An institution that runs MortgageGuide as a productivity assistant inside the Microsoft 365 surface, governed by the documented Microsoft 365 control set, with human authorship retained on every adverse-action narrative per 12 CFR section 1002.9, lands in a defensible spot against all of these regulatory frames. The agent does not change the institution's regulatory posture; the institution's governance configuration does that work. The agent operates inside that posture.

One specific point worth naming for credit-union compliance officers: NCUA examiners have begun asking about AI governance in lending operations as part of the standard examination program. The questions in the examiner table above are not hypothetical; they are the realistic content of a 2026 IT or compliance exam at a federally insured credit union, particularly those operating an active mortgage program. The same questions show up in OCC and FDIC examinations of national banks and state-chartered banks. The institution's documented response set is the asset; the agent is the productivity layer the response set covers.

For the institution's chief lending officer, the through-line back to the morning workflow at the top of this article is the cleanest part of the story. The loan officer who opens her laptop at 8:15 a.m. with three loans on her desk does not care about Microsoft Azure AI Foundry architecture, defect taxonomies, or examiner question patterns. She cares about whether the file in front of her is going to close, what she has to tell the borrower, and how to get to the next call. MortgageGuide is built for that loan officer. Everything else in this article (the platform, the security, the governance, the regulatory posture) is the scaffolding that makes the loan officer's experience defensible at the institution level. The scaffolding matters because the loan officer's workflow has to survive a QC review and an examiner conversation; the loan officer's daily experience is the point.

The bridge from where the institution is today to where it could be tomorrow runs through the AI Readiness Assessment that comes with the Pilot Pack offer, the 30-day Copilot Business adoption sprint that operationalizes the productivity gain, and MortgageGuide beta access for the mortgage-guideline lift. All three travel together inside the same engagement. All three are designed to land inside the same Microsoft 365 surface the institution's teams already work in every day. None of them require the institution to switch CSPs, migrate tenants, change vendors, or rework its existing Microsoft relationships. The path is additive. The choice is whether to start now while the Pilot Pack offer window is open through June 30, 2026, or to wait until the bundle promotional pricing expires and the standalone Copilot Business price climbs 17 percent on July 1, 2026. The agent itself, MortgageGuide, sits ready inside the engagement as it matures through beta toward general availability.

Frequently Asked Questions

Justin Kirsch

CEO, Access Business Technologies

Justin Kirsch has built technology for financial institutions since 1999. As CEO of Access Business Technologies, the largest Tier-1 Microsoft Cloud Solution Provider dedicated to financial services, he helps more than 750 banks, credit unions, and mortgage companies deploy Microsoft 365 Copilot Business with the governance financial-services examiners expect. ABT MortgageGuide Copilot is ABT's mortgage-specialized Copilot agent for mortgage lenders, currently in beta.