The Unsexy Software You Already Own Might Be Your Best AI Bet

For the past year or so, the conversation around AI in business software has been dominated by a single narrative: rip out your old system, replace it with something AI-native, and watch the magic happen. Sequoia's Julien Bek wrote a piece earlier this year called "Services: The New Software" that crystallised it well. The thesis is roughly: the next big company won't sell you a tool, it'll deliver the outcome directly. Closed books instead of accounting software. Resolved tickets instead of a help desk platform. Sequoia put its money where its thesis is and led a $25M Series A in Rillet, which is rebuilding the general ledger from scratch for the AI age.

The thesis is interesting and probably right in some verticals. But sitting where we sit, working hands-on with SMEs and mid-market companies on data pipelines and ERP integrations, we've been seeing something a bit different play out on the ground. And it's worth talking about because most of the companies we work with aren't going to migrate off their legacy ERP any time soon, AI-native option or not.

Old systems were never the problem. The interfaces were.

Most legacy ERPs and business systems were built by people who really knew their domain. Payroll software in France, for instance, encodes decades of evolving regulation, edge cases for specific industries, collective agreements, and a thousand small rules nobody documented because everybody who needed to know already knew. The data model is dense. The business logic is real. What killed the user experience was the front office: ugly forms, twenty fields where two would do, no mobile, no integration with the messaging tools people actually use during the day.

For a long time, the only way to fix this was to rebuild the whole thing. Which meant the new entrant had to reinvent both the interface and the decades of accumulated domain knowledge. That's why so many "modern" replacements end up being thinner, prettier versions of the original that lose 30% of the functionality and take three years to catch up.

What's changed in the past 18 months is that you can now layer modern interfaces and modern intelligence on top of the old system without touching its core. MCP, the protocol Anthropic introduced in late 2024 and which has since been adopted by OpenAI, Google and most of the ecosystem, is a big part of why. It gives language models a standardised way to read from and write to existing systems, so the AI can do the heavy lifting on the input/output side while the legacy engine keeps doing what it's good at: applying the rules.

What this looks like in practice

A few years back, before Reflekt Lab, I was working in a payroll context in France. Anyone who has touched French payroll knows it's a special kind of nightmare. The system we used was old. The screens were rough. But the calculation engine, once you fed it clean data, was solid. So what we ended up doing was building a front office around it: we handled the messy parts (hours worked, when, in which industry, under which collective agreement), did the pre-accounting work, and only then injected clean structured data into the legacy engine. The output came out correct because the engine was correct. The user experience was good because we'd taken care of the part the engine didn't care about.

That was already a few years ago and we were doing it the hard way, with custom code everywhere. What's different now is that the same pattern is becoming available without massive engineering effort. The AI handles the unstructured input: voice notes, photos of invoices, scanned PDFs, messy spreadsheets. The MCP layer pipes the cleaned data into whatever system you already run. You keep your ERP, you keep your database, you keep the rules you've encoded over fifteen years. You just stop hating the way you put data in.

We're seeing this with our own clients. Wine distributors with thirty thousand SKUs across forty supplier catalogues don't want to migrate platforms, they want their data prep to stop eating six weeks per onboarding. Restaurants in the UAE taking orders over WhatsApp voice notes don't want a new POS, they want the voice messages to land structured in the system they already use. The job is interface and ingestion, not replacement.

Why this matters for the migration question

For SMEs especially, ERP and CRM migrations have always carried disproportionate risk. The implementation timeline is long, the cost is heavy, and the operational disruption during cutover is real. A recent IDC-style estimate puts the typical mid-market ERP migration at six to eighteen months and seven figures in total cost when you include consulting, training, and downtime. For a company doing twenty million in revenue, that's an existential bet.

The AI-plus-MCP approach changes the calculus. You don't have to bet the company on a migration to get most of the productivity gains. You can keep the system that works, wrap it with better interfaces and better data ingestion, and ship something useful in weeks rather than quarters. The migration question becomes optional rather than existential.

This isn't an argument against AI-native rebuilds. Rillet will probably do well, and there are categories where the underlying data model genuinely needs to be redesigned. But for the long tail of mid-market companies sitting on functional legacy systems, the more pragmatic path is to make the system you already have feel modern. The Sequoia thesis is one valid reading of the future. The other reading is that a lot of the value in AI right now is in finally letting these old, knowledge-dense systems be used the way they should always have been used.

The dynamics also work in favour of incumbents in an interesting way. The legacy vendor that opens up an MCP layer over its existing product can suddenly look very competitive, because they keep all the domain knowledge baked in and they get to skip the part where they have to convince the customer to migrate. We're already starting to have those conversations with clients on the vendor side, not just the buyer side.

The catch

None of this is free. Layering AI on top of a legacy system means dealing with the same problem we wrote about a few months ago: at scale, even 98% accuracy is not enough. Two hallucinations in a hundred becomes six hundred in thirty thousand. So you still need a serious QA layer, human-in-the-loop checkpoints where it matters, and clear escalation paths when the model is unsure. The whole point of keeping the legacy engine is that the rules are deterministic. The AI front-end has to respect that determinism, not undermine it.

Done well, this is one of the most underrated patterns in the current AI cycle. Less glamorous than building a new ERP from scratch, much faster to deliver, and a lot more aligned with how most companies actually want to evolve their stack.