By Fola Yahaya , founder of Strategic Agenda and of Robert, an AI-powered CAT platform

Eighteen months ago, the question I was asking myself as the founder of a London language services company was the same one most of my peers were asking: how do we defend our margins against the slow grind of machine translation? It felt like the defining challenge of the decade. It was not. The defining challenge was going to arrive much faster and from a completely different direction.

When a large chunk of international development funding was frozen in early 2025, the tremor reached us within a fortnight. We work heavily with UN agencies, NGOs and humanitarian bodies e. Projects we had scoped didn’t get signed. Projects that had been signed got paused. Colleagues at other agencies were quietly telling me the same story. The sector had been bracing for an AI shock and got a funding shock instead.

What followed was the most uncomfortable and, in hindsight, most useful year of the company’s existence. I want to write about it honestly, because the version of this story that ends with “and then we discovered AI and everything was fine” is both untrue and unhelpful to anyone running a language service provider right now.

The first instinct was the wrong one

Our first instinct was to cut. Lower rates, thinner margins, keep the lights on, wait for the cycle to turn. Every small LSP owner reading this knows that instinct and probably knows where it leads. You win work you shouldn’t have won, at prices you can’t actually deliver at, and a few months later you are running faster to stand still.

The second instinct was slightly better but still wrong: bolt some AI onto the front of the existing workflow, market yourself as “AI-enabled”, and hope the cost base quietly drops. What actually happens is that you pay for the AI on top of the CAT licences you already have, your project managers get a new set of tools to babysit, and the efficiency gain gets eaten by the coordination cost. I suspect a lot of agencies have quietly had this experience in the last twelve months and are too polite to say so at conferences.

What actually worked

The thing that worked — and I mean genuinely shifted the P&L — was sitting down with our longest-standing clients and having a different conversation. Not “we’ve got AI now, here’s 30% off” but “your budget has been cut in half, your content needs have not, let’s redesign the service from scratch.”

That conversation led us to a tiered offer: raw machine output for internal or low-stakes content, light post-edit for working documents, full human review for anything public-facing or legally binding. Nothing about this is original. What was new for us was being willing to actually sell the cheaper tier rather than treating it as an embarrassment. Clients didn’t want to be protected from the AI. They wanted to be in the room while we decided which content deserved which treatment.

The operational side of that shift is where the tooling mattered, and this is the part of the story I want to be most honest about, because the honest version is less flattering than the LinkedIn version.

We looked at the existing CAT tools and couldn’t make the sums work. The per-word pricing fought the tiered model, the workflows assumed a linear human path, and the AI, where it existed, sat in a separate tab that nobody quite trusted. So we did the thing you are not supposed to do as a small agency: we decided to build our own platform. It is called Robert, and I am going to spend the next few paragraphs telling you why this was harder than we expected, because I think the difficulty itself is the interesting part.

The things that turn out not to be solved

If you had asked me eighteen months ago what the hard part of building an AI-native CAT tool would be, I would have said the AI. I would have been wrong. The AI is, relatively speaking, the easy bit. You pick your models, you wire them up, you tune the prompts, you argue about temperature settings for a week, and you have something usable.

The hard parts are the parts the industry has spent twenty years quietly getting right and which a new platform has to get right from day one or it is unusable. Formatting is the worst of them. A client sends you a Word document with tracked changes, embedded Excel tables, footnotes that reference a bibliography, a header image with overlaid text, and a table of contents that auto-updates. They want the translated version back in exactly the same shape. Not approximately the same shape. Exactly. The AI has no opinion on any of this. The AI has done its job the moment the sentence comes out in the target language. Everything that happens after that — the round-trip through the file format, the tag handling, the layout preservation, the regeneration of the TOC — is unglamorous plumbing, and it is where most of our engineering time has gone. I would estimate, honestly, that for every hour we spent on the AI layer we spent four on file format handling, and we are still not finished.

Second, translation memories and term bases. There is a temptation, when you are building an AI-first tool, to quietly hope that clients will lose interest in their legacy TMs and TBs and let the AI do the remembering. They will not. Every serious LSP client we have worked with this year has arrived with a TMX file, an Excel glossary, and strong opinions about how both should be applied. Quite rightly — those assets represent years of investment and institutional terminology that the client is not going to abandon because a language model sounds confident. So you end up building the full TM and TB stack anyway, and you end up teaching the AI to defer to them, which is a genuinely tricky problem because the AI would rather paraphrase. Getting an LLM to respect a term base the way a disciplined human translator would is not a checkbox on a feature list. It is weeks of prompt engineering and guardrail work, and it is still an ongoing project for us.

Third, and this one nearly broke us: the industry’s standard file exchanges — TMX, XLIFF, TBX — are old enough to vote and have accumulated two decades of dialectal variation. Every major tool implements them slightly differently, and everyone’s “standard-compliant” export turns out to contain quirks that were not quirks in the tool that produced them. Interoperability is the thing you have to get right if you want clients to migrate to you without losing their history, and it is almost entirely grunt work.

I am dwelling on all this because I think the industry conversation about AI translation tools has been dominated by people talking about AI translation, and the actual bottleneck is somewhere else entirely. If you are an LSP owner looking at the new generation of tools and trying to work out which ones are real, the questions to ask are boring ones. How do they handle a nested Word table with merged cells? What do they do when your TMX has segments with inline tags in a non-standard order? Can they round-trip a PowerPoint with embedded charts? The tools that can answer those questions confidently are the ones that have done the unglamorous work. The ones that can only tell you about their model selection or their prompt library have not.

Why we kept going anyway

Given all that, a fair question is why we persisted rather than muddling through with an incumbent tool and a lot of workarounds. The honest answer is that the commercial maths eventually forced the decision. Once we had a stable tiered service model and the volumes started to grow, the licence costs on the incumbent tool were going up faster than the human-billable portion of the work. We were paying a per-word premium on AI-processed volume that the tool had contributed very little to producing. At some point the build-versus-rent calculation tipped, and once it tipped it stayed tipped.

The tooling choice was downstream of the business model choice, not the other way round. I want to stress that, because I am not writing this to talk anyone into building their own CAT tool. For most LSPs, that would be an obviously bad idea. We did it because we had two things most agencies do not have — a CTO who had spent his career in machine translation architecture and a willingness to take on a couple of years of engineering work with no guarantee it would pay off. Without either of those, we would have negotiated harder with our incumbent vendor and got on with life.

The uncomfortable part

Here is the part that is harder to say at industry events. Once you start charging clients for the right level of quality rather than the maximum level of quality, you stop being able to hide behind craftsmanship as a justification for price. You have to be useful instead. That is a genuinely different posture for a sector that has spent twenty years arguing, quite correctly, that translation is a skilled profession deserving of fair rates.

The good news is that being useful turns out to pay better than being precious. Our three biggest accounts this year are ones where we sat on the same side of the table as the client and helped them spend less on translation so they could spend more on the things translation was meant to enable — programmes, campaigns, reach. Two of them have increased their overall spend with us since that conversation, not despite the AI but because of it.

What this means for other small agencies

I am wary of turning a specific story into a general prescription, but three things seem to hold.

First, the LSPs who are doing well right now are not the ones with the fanciest AI stack. They are the ones who had honest conversations with their clients early and were willing to restructure the commercial offer. The technology enables the restructuring, it doesn’t cause it.

Second, the cost base matters more than it used to. A year ago, a five-figure annual CAT licence was an irritation. Today, for a small agency running at thinner margins, it can be the difference between hiring a project manager and not. I know several owners who have moved off incumbent tools for this reason alone, and the sky has not fallen.

Third, and this is the one I did not expect: our project managers are happier. They spend less time fighting the software and more time on the judgement calls that actually need a human. That is worth something that doesn’t show up on a spreadsheet.

The geopolitical shock isn’t over, and I don’t know what the next one looks like. What I do know is that the agencies still standing at the end of 2026 will be the ones who used this period to change what they sell, not just how they deliver it. The AI is the easy part. The business model is the hard part. Get that right and the rest follows.

About the author

Fola Yahaya is the founder of Strategic Agenda and of Robert, an AI-powered CAT platform

About Robert

Robert is an AI-driven translation platform that enables LSPs to deliver translation as a branded, client-facing service. Through a fully customisable, white-label portal, agencies can give their clients direct access to fast, secure AI translation, while retaining control over quality through human review and sign-off. By embedding a translation tool within their own brand and workflows, LSPs can strengthen client relationships, reduce reliance on one-off project work, and build more stable, recurring revenue streams. Learn more here: https://translatewithrobert.ai/