Why I Chose to Rebuild Tender Management Instead of Bolting AI Onto It

5 MINS

Most legacy tender modules in this industry share a family resemblance: heavily configurable, clunky, and non-standardised across forwarders. Every customer got their own configuration, and over the years those configurations became the product. Bolting an AI feature onto that is like adding a turbocharger to something with no brakes.

Configurability was the right answer to the wrong question

Configurability solved a real problem a decade ago. Forwarders genuinely do run tenders differently, and rather than choose, the software let each of them encode their own process. That was a defensible call at the time.

The cost shows up later. When behaviour lives in configuration, nothing can be reasoned about globally. You can't score a tender consistently because "tender" means something slightly different per tenant. You can't ship an improvement because you can't predict what it breaks. And you certainly can't hand an agent a workflow whose rules are scattered across a settings table nobody has audited since 2019.

What AI-native actually means here

An AI-native rebuild is not "the same screens plus a chat box." For tender management it means the system holds an opinion about the work: autonomous tender scoring, carrier shortlisting, agentic document extraction from whatever the shipper sent over, contract risk flagged before a human reads clause 14.

That only works if the underlying model of a tender is standard enough to reason over. So the rebuild is as much a modelling exercise as an AI one, deciding what a lane, a bid round, an award, and a commitment mean, once, properly, and then letting the agents work against that. The AI is the reason to standardise; standardisation is what makes the AI possible.

Honest build-vs-buy calls

I've made the buy call plenty of times. Xeneta for ocean and air market rates was an obvious one, market rate data is not a thing I should be building. The test I use is whether the capability is where our judgement lives. Market rates are a commodity input. How you score a tradelane against your own network, capacity, and appetite is not.

The same test applies to models. There's no prize for training something you could call. The prize is in the workflow, the data, and the guardrails, because that's the part that is genuinely yours and genuinely hard to copy.

Roadmaps that survive contact with reality

The trap in a rebuild is the two-year silent march to parity. Nobody funds that honestly, and nobody survives it. So the sequencing question I keep asking is: which slice of this can go live against real tenders, for real forwarders, soonest, and what does it have to be better at than the legacy path to earn the switch?

Rebuilds are justified by what they make possible, not by what they replace. If the new tender flow only matches the old one, I've spent a year buying nothing. If it can score, shortlist, extract, and flag in a way the configurable version structurally could not, that's a different product, and that's worth starting from scratch for.

Reliable Enough to Trust at 2 AM: Shipping Agentic AI Into Freight Ops

In Freight, the Problem Is Rarely the Algorithm

Lokesh skipped presentations and built real AI products.

Lokesh Radhakrishna was part of the June 2026 cohort at Curious PM, alongside 20 other talented participants.