Your buyer cannot inspect your AI model. They are buying your willingness to be examined.

Back to Insights

The best platform in your market lost a deal last quarter to a worse one, and the reason was not price, features or references. The buyer did not trust the seller enough to put their own name behind the purchase. Two weeks ago I looked at what that person now answers for once your agents go live inside their firm. This week is about what you put in front of them.

That sounds like a sales pathology. In regulated financial services it is the system working as designed. If you sell software to fund managers, banks or insurers, the person who signs for it is not a procurement function but a named individual, personally accountable for the outcome. Remember what their product is: the funds, share classes and policies they sell their own clients. Your modules are the machinery underneath, and your failures surface in their product before they surface in yours. When a COO approves your platform, they attach their own credibility to it in front of their board, their regulator and their clients. So the real product is never the software. It is the buyer's confidence that you will not cost them the thing they cannot buy back.

I made this argument in Ethics, Trust, and Public Perception, the chapter of my book this piece draws on: trust is the defining currency of financial services, and clients who are empowered by choice, armed with information and motivated by values will act on its absence. Wells Fargo, Wirecard and FTX did not fail on product. They failed on the question your buyer quietly asks in every meeting: if this goes wrong, what happens to me?

The 2020s version of that failure already has a name. The SEC's first AI-washing cases were investment advisers: Delphia paid $225,000 for claiming it put client data to work making its artificial intelligence smarter when no such thing was happening. The instructive case is Presto Automation, which did use AI and still faced SEC charges for overstating how much and understating the humans behind it. The gap between what you claim and what your technology does is now a legal exposure.

Due diligence has met a machine it cannot open

Fund managers are good at vendor due diligence. They have run it for decades: financials, references, SOC reports, exit provisions, a site visit if the cheque is large enough. The whole process assumes the thing being bought can be inspected.

An AI-enabled platform breaks that assumption. The operational due diligence team reviewing it is learning to assess AI while the field reinvents itself quarterly. And in most cases they are never going to see inside the AI model. The weights are not inspectable, the training data is not theirs, and the behaviour they tested at onboarding shifts with every release. Inspection, the foundation of every diligence process they have ever run, is not on offer.

So the questions change; the risk has moved. How is our data used, and where does it travel? Has bias been tested against a client base like ours? What is the security posture, and what happened the last time it was probed? Who do we call when the model is wrong? Every failure mode lands on the buyer and their clients, not on your roadmap. When the thing you sell cannot be fully examined, the thing you are selling is your willingness to be examined.

It is already happening on the buyer's side: Deutsche Bank switched on agentic AI in its third-party risk management in December 2025: AI agents retrieve the control questions, draft answers from the vendor's own evidence and propose an outcome with citations, at around 90 per cent accuracy benchmarked against human validation, with decision rights staying with trained human assessors. The first pass through a vendor's evidence is now systematic, repeatable and less persuadable by presentation. If your public claims, security material, AI documentation and contractual position do not agree, the inconsistency becomes visible before your sponsor has had a chance to explain it away.

The buyer's caution now has regulatory weight behind it. In July the FCA reported 11 open Consumer Duty investigations, up from 6 in January, alongside 382 supervisory interventions in a year. Your buyer must evidence good client outcomes, so your software sits inside their evidence chain and their diligence questions are self-protection, not ceremony. The EU made the same point differently: Brussels delayed the AI Act's high-risk rules to December 2027 but kept transparency on schedule, so since 2 August AI-generated content must be identifiable as such. Of everything in the Act, provenance is what regulators refused to postpone. And the FCA's latest perimeter report notes consumers increasingly asking general-purpose AI platforms for help with financial decisions. Some version of your next buyer's first screen will be an agent that reads your documentation and checks your claims against the public record.

What the winning vendors put on the table

I spent eighteen years on the buying side of this table, selecting and overseeing the technology and service providers beneath a platform that grew from $200 million to $31.6 billion. The vendors who won were rarely the ones with the best demo. They made my internal case for me: my real task was never deciding whether I liked their platform, it was defending the decision for years afterwards, and they handed me the material to do it.

The AI-era version of that pack is specific. AI model governance documentation a buyer can put in front of a board without translating it. A data map that answers where client data goes, in a page, not a policy. Bias testing with results attached. A security posture you volunteer before being asked, including how the last incident was handled. Provenance of your marketing claims, so what the buyer's agent finds matches what your salesperson said. And a named human who owns the answer when the model is wrong, because the buyer's regulator will ask them the same question. That last item is the file I argued every agent should arrive with, and the one the buyer's own register will now expect to hold.

The commercial effect is measurable. Bain Capital's three-person vendor risk team moved its assessments onto VISO Trust's AI platform, which reads SOC 2 reports, penetration test findings and certifications and maps them to the buyer's control requirements. Reviews now finish in half the time and completed assessments are up 24 per cent year on year. That is the revenue case for a trust pack: not more paperwork, less time spent making a buyer hunt for proof.

None of this is in the demo, and all of it is commercial. It shortens procurement because the buyer's diligence has somewhere to land. It survives your champion changing jobs. And it compounds the way my book argued trust does: into retention, referral and renewal, lower cost of sale, and the Net Promoter Scores that make a revenue plan believable. A trusted vendor's renewal is a conversation; an untrusted vendor's is a re-procurement against the whole market.

Trust is the product spec now

There is a reading of all this that files it under compliance, and it is the wrong reading. Nothing above is about avoiding fines. It is about why deals are won and lost, which makes it the revenue leader's problem and opportunity. That chapter sits with revenue and client retention for a reason: I wrote there that trust, once considered intangible, is now trackable, actionable and indispensable, and 2026 has made that literal. Even public perception, its third strand, now includes what a buyer's agent finds when it goes looking.

This is where the series lands. The first piece said the moat a language model cannot copy is distribution into buyers who already trust you. The revenue engine, the voices that speak for you and the file for every agent were ways of earning it. Last week's five marks decide who gets to put a trust pack on the table at all: a tool that only holds a workflow has nothing to prove, and a platform with the data still has to prove it. Trust is not the sixth topic. It is what the other five were for.

If there is no trust, there is no sale, however good the platform. The firms that treat that as a design brief for how they sell, rather than a constraint on it, will find their sales cycle getting shorter while everyone else's gets longer. That is not a compliance outcome. That is revenue.

If you sell into regulated firms, I build that pack with vendors as a fixed three-week piece of work: the AI model governance summary a board can read, the one-page data map, the security posture, the claims check against what a buyer's agent will find, and the named owner, assembled from your own material into something a buyer's diligence can land on. It sits inside the go-to-market work I do with regtech and fintech vendors. If you are the one who signs, the Technology Selection and Vendor AI Oversight Review work is the other side of the same table.

Declan Sheehy is the founder of One Blackwater Consultancy and advises FCA-regulated fund and asset managers, and the fintech and RegTech firms that serve them, on GTM, AI strategy, governance, and operating models. He spent eighteen years scaling an alternative investments platform from $200 million to $31.6 billion. He is an IAPP-certified Artificial Intelligence Governance Professional (AIGP) and the author of Disruptive Innovation in Financial Services (2026).