What we're watching.
What it means for your CU.

For credit unions with auto lending, direct and indirect. Drawn from T.D. 10054 (Sept. 8, 2026) and the IRS Instructions for Form 1098-VLI (Rev. December 2026), effective November 9, 2026. Not tax advice.

Form 1098-VLI is showing up in year-end work for a lot of credit unions. Cores are building it into servicing. Members will get statements. That side of it is underway.

Box 1 is Vehicle loan interest received by lender. To get there you need deal pieces like negative equity, cash-out, and related add-ons, tied to the interest the core received on that loan. Those pieces often sit on the deal or in the LOS. They are not always clean fields on the servicing system.

Same rules whether the loan came in direct or indirect. The difference is where the fields live.

CU*Answers has said production is a servicing job, and that negative equity and refinance math is the credit union's to track — and that members are responsible for their own tax determinations:

What has to meet in one place

At year-end you need a clean join across deal structure (negative equity, cash-out, what was financed), vehicle and eligibility facts, interest received on that loan, and enough to run the per-loan $600 test on qualifying interest.

If any of that is missing, or it only exists as a funded total, Box 1 is a guess.

What Attrivix does here

Attrivix connects the systems credit unions already run, inside the credit union's own Microsoft environment. The same idea applies when the IRS asks for loan and vehicle fields that don't sit in one system.

For Form 1098-VLI, that means pulling the required points together, calculating qualifying interest, and flagging loans at the $600 line. Lending and tax teams can see what's reportable, what's under threshold, and what's still missing. It does not replace the core's tax file or e-file.

If you're looking at the same handoff and want to see how that looks on your book, let's talk.

Sources: IRS Instructions for Form 1098-VLI (Rev. December 2026), Form 1098-VLI, T.D. 10054, 91 Fed. Reg. 57214 (Sept. 8, 2026) (final regulations, effective Nov. 9, 2026), and CU*Answers Kitchen, Planning for IRS Form 1098-VLI.

A few years ago, Copilot became my company's official AI. I remember being genuinely irritated by it. Compared to other LLMs I was using at the time, it felt clunky, harder to get useful answers out of, more friction than it was worth. So honestly, I just didn't use it. I found workarounds and moved on.

I think about that a lot lately, because now I'm building infrastructure that runs directly through Copilot, and it's a completely different tool than the one I avoided.

What changed

Microsoft just made Copilot able to answer questions straight from Power BI data, no separate dashboard, no digging for the right report. Ask it something the way you'd ask a colleague, and it pulls a real answer from real numbers. That's not a small improvement. It's the difference between a tool you tolerate and a tool that actually gets out of your way.

Watching that shift happen has genuinely surprised me. Microsoft has closed a lot of the gap that used to make Copilot feel like the mandatory tool instead of the useful one. For enterprise companies, that matters, it means what's happening under the hood is finally becoming visible and usable, not just theoretically powerful.

Getting answers quickly with Copilot only helps if there's something real underneath it to be fast about.

What hasn't changed

Here's what hasn't changed, and it really is behind the mission of Attrivix. Copilot's new speed only works if a well-built semantic model already exists underneath it. Copilot doesn't build that connection.

What is a semantic layer? Let's book time if you're wondering, it's worth a real conversation, not a quick definition buried in a blog post.

The point is this: you can access the data that's already connected through your Copilot, and that's amazing. It's exactly the reason I'm passionate about helping credit unions access the data they already have, using tools they already have.

For most credit unions, that connection doesn't exist yet. Core, LOS, marketing, call center, each system holds its own piece of the member's story, and nothing stitches them together. So Copilot answering faster from whatever it can see isn't automatically good news. If the picture underneath it is incomplete, the model doesn't know that. It answers with the same confidence either way.

Where that leaves credit unions

Microsoft built the layer that makes a good semantic model pay off immediately. Whether a credit union has one, or has one that's built strong enough to trust, is a separate question. Credit unions have to decide to lean into their own data and untangle it. Your data is only as good as you make it.

Attrivix builds that connective layer, so what Copilot answers from is the full member journey. Full stop.

If you're not sure how connected your own systems actually are, that's worth a real conversation.

A credit union brought us in to look at something that, on the surface, sounded like a scheduling problem.

Dealers were sending them business. Good business, the kind of relationship most credit unions want more of. But there was a pattern the credit union had started to notice: deals that came in Friday evening or over a weekend had a way of disappearing. Not declined. Not funded elsewhere because of rate or terms. Just gone, sent to whoever picked up the phone next.

The instinct was to call it a staffing issue. Nobody's in the office Saturday, nobody's watching Sunday, so weekend deals sit until Monday. Reasonable enough conclusion, until you look at what those deals actually needed.

What was really happening

Most of what came through during the week and got flagged for manual review still closed. The credit union's team was fast enough at resolving those referrals that dealers stayed comfortable sending them business, even the borderline cases their automated criteria couldn't clear on its own.

The weekend broke that rhythm entirely. The deals themselves weren't different. The person available to look at them was.

So the real issue wasn't that the office was closed. It was that the credit union's lending model depended on a human being present to resolve anything outside a fairly tight automated threshold, and that threshold had no ability to flex when the calendar did.

The response, and its limits

The credit union's answer was to loosen their automated underwriting so more weekend applications could clear on their own, without waiting for a person. Sensible, and probably the right first move.

But it's a broad instrument for a narrow problem. Widen the criteria enough to catch Saturday's applicants, and you've widened it for Tuesday's too. The credit union wasn't trying to change who they lent to across the board. They were trying to stop losing a specific, recurring group of deals that happened to land on the wrong day of the week.

The sharper fix isn't a wider net. It's knowing exactly who's in that gray area to begin with, and whether they're a real, repeating group of applicants or just scattered edge cases. That's not a threshold question. It's a visibility question, and most credit unions don't have the data connected in a way that can answer it. Who applied, when, what happened to the referral, whether it eventually converted somewhere else, that information typically lives in three or four systems that were never built to compare notes.

Why this kind of problem sits unresolved for so long

What's stayed with me from that conversation isn't the mechanics of the fix. It's how long it took the credit union to name the problem clearly enough to bring someone in about it.

I've sat with other institutions that know they're losing exactly this kind of business and still haven't moved on it. I don't think it's always a lack of urgency. Sometimes it's the fear that loosening criteria opens a door too wide. Sometimes the loss they already understand feels safer than a change whose full effect they can't yet see.

The credit unions that do get past that point have one thing in common: they stop guessing. They can see, specifically, who's getting approved, who's getting referred, and why, clearly enough to know whether their criteria matches the members actually applying, whether that's a Tuesday afternoon or a Saturday night.

That's the gap Attrivix is built to close. Not a wider net, and not a better guess. The actual pattern behind every application, referral, and outcome, so a credit union can see exactly where its criteria fits and where it doesn't, without needing a dealer to tell them on a call.

Every AI pitch to credit unions right now starts the same way. Bring in outside data. Layer on segmentation. Tell you who to target next.

That's an outside-in approach. Go get more information from somewhere else, then act on it. It's not wrong, exactly. External targeting and segmentation tools can genuinely help a credit union find members who look like good candidates for a product they don't have yet.

But there's a different question nobody's asking as loudly, and it's the more important one for most credit unions right now. What if the answer was never outside your walls to begin with?

What inside-out actually means

Every credit union I worked with over seven years already had the data. The core had it. The loan origination system had it. The call center had it. The marketing automation platform had it. None of it needed to come from an outside source. It just needed to talk to itself.

A member gets an email. Their spouse walks into a branch a week later and funds a loan. The email platform doesn't know that happened. The core doesn't know a campaign started it. The call center doesn't know either piece. Three systems, three pieces of the same story, and nobody connected them.

That's not a data problem. It's a connection problem. And it's sitting inside the institution already, not somewhere external waiting to be acquired.

Why this distinction matters for budget conversations

Outside-in and inside-out lead to very different decisions in a planning meeting.

Outside-in says: bring in a new vendor, a new data source, a new layer of intelligence sitting on top of what you have. That's a new procurement process, a new risk review, a new relationship to manage, and often, a new place your member data has to travel to.

Inside-out says: look at what you already have and connect it. No new vendor to vet. No new place for data to live. The infrastructure and the answer were already inside the institution the whole time.

Neither approach is inherently wrong. But they're not the same question, and credit unions evaluating "AI strategy" right now are often being sold outside-in solutions to what is fundamentally an inside-out problem.

The question worth asking in the room

Before a credit union brings in another outside layer, it's worth asking a simpler question first: do we actually know what we already have, and is it connected?

For most of the institutions I worked with, the honest answer was no. Not because the data didn't exist. Because nobody had gone in and connected what was already there.

That's the inside-out approach. Not buying more. Looking at what's already inside your own systems, and finally letting them talk to each other.

I worked with at least three credit unions trying to use their marketing automation platform to attach spiffs to leads. Different institutions, different versions of the same idea.

The platform wasn't built for that. And they knew it going in.

Their spiff models weren't simple either. Different tiers, different payouts depending on the product, the source, who touched it along the way. Trying to force that kind of complexity through a tool that was never meant to track it just multiplied the problem instead of solving it.

Some of them ran into the same holes I'd already seen elsewhere, and we'd work through those together. Others found entirely new ones. Their version of the workaround created its own tangle, and we were back at it, trying to untangle something nobody had seen before.

Every one of them had a dedicated person whose real job became finding the holes. Which lead came from which effort. Who should get credited, and at what tier. Manually cross-referencing what the system was never designed to track in the first place.

None of them had budget for the tool that actually does this well. So they took the tool they had, and made it stretch.

Then the person who understood it left

At one of these credit unions, the whole workaround lived in one person's head. Codes, manual tagging, a running system of cross-references that only really made sense to the person who'd built it.

It worked, for a while. Then that person moved on to another company.

The people who inherited it couldn't figure it out. Not because they weren't capable, but because it was never actually a system. It was one person's mental model for forcing a tool to do something it was never built to do. Once that person was gone, so was the only real documentation of how any of it worked. What had been a functioning, if fragile, process turned into a mess, for the team trying to run it day to day, and for the managers trying to figure out what they actually owed people.

That credit union ended up moving to another platform entirely. I lost visibility on where they landed after that.

Here's what stays with me about it. The workaround itself wasn't lazy. It was genuinely inventive; someone saw a real gap and built something to close it with whatever they had on hand. But it only held together as long as one specific person was there to hold it together. The moment they walked out the door, so did the only map anyone had.

A VP I was talking with recently told me about a regulator inquiry that came in on a member. She needed an answer fast.

She turned to Copilot, searched her email, and found an exchange from over a year ago in about five minutes. She had a full answer back to her CEO in twenty minutes. No digging through records. No looping in an employee who might not even remember the conversation.

That's the promise, and it's real. I've seen it work.

Why it worked

Copilot found that answer because the data was right there in her inbox, reachable. It didn't have to guess. It didn't have to reason around a gap. The information existed in a place the tool could actually see, and it did exactly what it's supposed to do.

That distinction matters more than it sounds like it should. AI is only as good as the data you feed it or give it access to.

The version that doesn't get talked about enough

Here's what's real about the other side of that same tool. When Copilot doesn't have full access to something, it doesn't reliably say "I don't have enough information." It's a documented pattern in enterprise AI: models are trained to favor a confident answer over admitting uncertainty, the same instinct behind never leaving an exam question blank. When the data required to answer accurately isn't available, the model can generate a plausible-sounding answer anyway, one built on pattern and probability instead of what's actually true.

If Copilot's answer sounds complete, there's a real tendency to let it through without a second look, and that's exactly how bad information ends up in an executive summary or a board update.

Why this matters right now

Velera rolled out their Atmos ecosystem earlier this year, combining data, risk, and payments capabilities with AI-enabled insights, built to help credit unions do more with the data they already have. That's the right instinct, and Velera isn't the only one moving this direction. The whole industry is racing toward "put AI on top of your data" as the next competitive move.

But racing toward AI-on-your-data only works if the data underneath is actually complete. A tool that's brilliant with what it can see is still blind to what it can't, and it won't necessarily tell you which one it's doing in the moment.

What the difference actually comes down to

Her result and the hallucination risk aren't two separate stories. They're the same mechanism, pointed in opposite directions. Give the model complete, connected data, and it delivers in minutes what used to take hours. Give it a gap, and it fills that gap with something that sounds just as confident, whether or not it's true.

The question every credit union moving toward Copilot, Atmos, or any AI layer needs to ask isn't "does this tool work." It's "how do I know when it's working from something real, and when it's working from a gap it didn't tell me about."

That's not a reason to slow down on AI. It's a reason to get serious about what's actually connecting your systems before you start asking it questions that matter.

A $2.3 billion credit union — Communication Federal Credit Union — went live with Scienaptic's AI credit decisioning system. The results were impressive: the implementation is expected to unlock $134.5 million in additional vehicle loan originations and up to a 20% reduction in losses across their consumer loan portfolios.

Scienaptic is a strong decisioning tool. It sits inside the loan origination system and uses advanced models to make faster, more consistent credit decisions based on the data it can see — applications, credit information, and internal relationship details. Humans remain in the loop for exceptions.

That is meaningful progress in the decision layer.

But it also highlights a bigger opportunity.

Right now, even the best decisioning tools are working with a compartmentalized view. What they typically don't see is the full member journey across the entire credit union: the email campaign that first reached them, the call center conversation, the branch interaction, or the household member who got the offer and walked in later.

When those systems — the core, the LOS, the CRM, the marketing automation, and the website analytics — are finally connected, two things happen.

First, you get true attribution. You can see where members actually came from and which campaigns drove funded loans.

Second, the decisioning itself improves. The AI now has richer context about the member's journey, not just the snapshot at application time. That leads to smarter approvals, better risk segmentation, and ultimately better outcomes for both the credit union and its members.

This is the shift I keep coming back to. AI can speed up decisions and reduce manual work. But the real leap comes when it has the complete story.

The data has always been there, scattered across systems. The missing piece was the connection.

For too long, credit unions have been guessing at what is driving funded loan outcomes.

Maybe someone built a custom workaround. Maybe there is a tracking code on the email campaign that captures some of it. But when the husband gets the email and the wife walks into the branch and funds the loan, that code never saw it. The attribution stops where the member's journey does not.

And that gap is costing more than most institutions realize.

What optional attribution actually costs

When attribution is treated as a nice-to-have, the consequences show up in places that feel unrelated. Budgets go to the wrong channels because the right ones are not getting credit. Front line staff do not get the coaching they need because nobody can identify where the lead broke down. Referral spiffs get paid out on incomplete data. And somewhere in a branch, a lead card is being hand-written and walked across the floor because there is no system connecting what happens there to what happens everywhere else.

These are not small inefficiencies. They compound. Every quarter that passes without attribution is another quarter of decisions made on assumption instead of evidence.

Attribution drives budgets. Budgets drive revenue. Revenue drives the health of the institution. That chain is not optional.

The guess that became the standard

I spent seven years managing enterprise accounts across 120 credit unions. The attribution conversation came up constantly, in board rooms, in marketing reviews, in one-on-one conversations after the agenda was done. And the same pattern repeated at institution after institution.

The marketing team knew their campaigns were working. They could feel it. But when leadership asked what the marketing budget actually returned, the honest answer, the one nobody said out loud, was that they did not know. Not precisely. Not provably.

So they correlated. They pointed at loan volume trends. They built pivot tables that told a plausible story. They delivered answers that started with "based on what we can see." And leadership nodded, because there was nothing better to point to.

That guess became the standard. And the standard became the budget process.

What the gap is actually hiding

The problem is not that credit unions do not have data. They have more data than they have ever had. The problem is that it sits in pieces, in the core, in the LOS, in the call center platform, in the email system, and nobody connected those pieces to each other.

When they are not connected, the real picture is invisible. A campaign that looks like it drove twenty loans actually drove forty-eight. A vendor that looks like it is performing is quietly leaking budget. A front line team member who looks like they are converting leads is actually losing them somewhere between the conversation and the follow-up, and nobody knows where because the journey was never tracked.

You cannot fix what you cannot see. And right now, most credit unions cannot see nearly enough.

What mandatory attribution actually delivers

I am not saying attribution gives you all the answers. I am saying it answers a lot of your burning questions. What dollars funded which loans. What the member journey actually looked like from first touch to close. Which vendors are working and which are not, and why.

CU leaders do not need more dashboards. They need actual numbers to make stronger decisions. That is what attribution done right delivers, not a report, but a foundation for every budget conversation, every staffing decision, every vendor renewal that follows.

The technology to do this right now exists. It is built inside your existing Microsoft environment. It reads what is already there. It does not require a new vendor procurement, an IT overhaul, or months of implementation risk.

What it requires is deciding that the full picture matters enough to go get it.

Attribution is not optional. It never was.