AI workflows – Navan: Why most workflows fail (and what we did differently)
This is a guest post for the Computer Weekly Developer Network written by Ilan Twig in his capacity as co-founder and CTO at Navan.
Navan is known for its software platform that integrates business travel booking, corporate cards and expense management. The company aims to provide real-time visibility into spending, to automate workflows and help enforce policy compliance. The software also features automated expense reporting, policy controls and travel booking through its interface and mobile app.
Twig writes in full as follows…
When MIT reported last month that 95% of AI pilots fail, the headlines wrote themselves. A brutal number, a ready-made storyline of wasted money and inflated expectations. The important question is why do most AI projects fail?
The answer is rarely that the technology doesn’t work. Large Language Models are powerful, accessible and improving at breakneck speed.
The dirt-road F1 paradox
The problem lies in how businesses use them. Too often, AI is dropped on top of old systems and outdated processes. A shiny new tool strapped onto infrastructure never designed to support it. The workflows don’t change, the infrastructure can’t keep up and the AI ends up failing to make a difference. It’s like trying to drive a Formula 1 car on a dirt road: the machine is powerful, but without the right track, it won’t reach its full potential. You have to build the road before the car can perform.
I’ve seen this pattern repeatedly and a big part of the issue is data. AI is only as strong as the information it runs on and many companies don’t have the standardised, consistent datasets needed to train or deploy it reliably.
The other weak point is reliability. LLMs are powerful, but they have a proclivity to hallucinate. If implemented hastily they can fail. If you want AI to take part in core or customer-facing workflows especially, you need guardrails that account for those weaknesses so the system is dependable and always accurate.
When LLMs first became available in 2022, we knew no off-the-shelf AI company could deliver what we wanted. So we built Navan Cognition, our AI platform, from the ground up. The platform stitches together a network of AI agents, powered by LLMs and each with their own specialised role.
That decision has shaped everything we’ve done since.
Because AI sits inside the platform rather than hovering above it, it works with unified data, real-time content and connected systems. It ingests fragmented inputs, from airlines, hotels, payments, receipts and calendars and aligns them so AI can act on a complete, accurate picture. It doesn’t have to fight through silos or clumsy hand-offs. It can make decisions quickly and carry them through to execution.
Specific AI agents within the system continuously supervise the work of others, boosting accuracy and striving to adhere to Navan’s zero-critical-hallucination standard in its use of AI.
Our AI now handles thousands of customer chats every day with fast response times. It cuts rebooking workflows from hours to minutes. It saves finance teams hours in reconciliation time.
Don’t bolt the workflow
That is why I believe the “95% fail” statistic is misleading if taken at face value. The problem isn’t AI. It’s the way companies try to bolt it onto workflows that were never built to support it. You don’t fix a broken process by layering intelligence on top. You fix it by rethinking the process itself.
The lesson for any company is clear. Don’t start with the model. Start with the workflow. Ask what problem you are really trying to solve, what data you need to do it and what infrastructure must be in place to make it possible. Factor in LLMs’ limitations before you implement it. If your systems can’t support AI, the pilot will fail before it begins. If they can, the gains can be extraordinary.
AI workflows aren’t an abstract concept.
They are the fabric of how businesses now run. Done badly, they waste money and damage trust. Done well, they reduce friction, save time and free people to focus on work that matters.
The difference between failure and success is not the model you choose. It’s whether you have the courage to rethink the workflow around it.