First Station, Wrong Destination

Most teams are building the AI train before they've laid the tracks. People onboarded, tools wired into live processes, tasks automated at speed. The early outcomes feel like winning: a quick hit here, a small win there, real momentum. They've pulled into the first station, maybe the second. But that isn't the real win, because the tracks ahead were never built to reach the destination they think they're heading to. The train is running on context that was ingested wrong, or never ingested at all. And once a model is running on the wrong foundation, going back to fix it costs far more than laying the track straight would have in the first place.

It almost always traces back to the same thing: context, and the ingestion that should have come first.

Lay the track first

So here's the unglamorous truth: your first AI project isn't an agent. It's ingestion, done upfront, guided, managed by people who know what these engines need. Mapping where your context lives, cleaning it, structuring it, getting it out of people's heads before they walk out the door.

It's boring. It's foundational. It's the difference between AI that compounds value and AI that quietly stalls out like most of it does.

Context is king

Your edge has never been the engine, it's the in-house gold that made you successful in the first place: the industry intelligence, the instincts, the hard-won lessons, the judgement behind every call, and the way your business actually wins. That is exactly what a model needs to be worth anything to you. Feed it in, on the right tracks, and it accelerates everything that already makes you great. Get it wrong and no model saves you. Context is king, and clean, full ingestion is how you crown it.

Start with the current state, not the tools

And it starts before the tools. Until you've analysed the current state, how are you choosing the right tools, vendors and processes? This isn't one-size-fits-all. It takes a task-level breakdown. What are people actually doing day to day, and what are you genuinely trying to automate? Only then can you weigh tools, vendors, processes, JDs and HR data against real work, instead of buying a platform for a problem you never defined.

Where your context actually lives

That analysis starts with an honest look at where your business context actually lives right now:

• SaaS platforms that don't talk to each other.
• Spreadsheets someone built in 2021 and never documented.
• Job descriptions that haven't matched the real role in years.
• Processes that survive only because "that's how we've always done it."
• And the most valuable layer of all: in people's heads. The exception handled manually for three years. The pricing carve-out nobody ever wrote down. This is the gold: hard-won subject-matter expertise and experience you need to capture and drive from.

That gold is also the killer. It's tribal knowledge, the operating system your business actually runs on, and an agent can't see it. Query your data and it reads what the schema says and nothing else: not the workaround, not the metric definition that changed 18 months ago. The answer comes back technically correct and organisationally wrong, and that error compounds across every decision it touches.

The pressure trap: freeze, then roll

We're seeing this on the front lines right now, with brands small and large. Some are going agentic-first, others are dipping a toe in, but nearly all are under pressure from boards and investors to move now. So they freeze, unsure what to automate first, unable to wrangle the documents that actually matter into one place. Then they roll: LLM licences pushed out to everyone, prompt workshops booked, rollouts spreading team by team. It looks like momentum, but it skips the full ingestion entirely. And it costs a fortune.

That's where it breaks. Silos move at different speeds, each loading the same tools with its own inputs. Nobody shares what works, nobody pools the insight back in, and the same documents get wrangled three times over by teams, departments and regions that don't know what the others are doing. The centre of the organisation ends up more confused than when it started, all because the ingestion was rushed, ungoverned, and run by people who didn't know what the AI engines need.

The spend explodes, the intelligence doesn't

The numbers back it up. MIT found up to 95% of generative AI pilots never scale past experimentation, and it's almost never the model. It's data readiness. Garbage in, garbage out; except now you've automated the garbage, so it scales.

And it isn't cheap. Uber burned its entire 2026 AI budget in four months, Microsoft pulled internal AI licences over cost, and only around 14% of CFOs can point to a clear return. A lot of that is structure: hand an agent a full 80-slide PDF instead of a clean file and it tokenises headers, footers and white space before it reaches a useful sentence. Multiply that across every team and query, every day, and it doesn't even buy you a better answer.

Because more isn't better. The instinct is "give it everything, it needs the background," so the knowledge base becomes a dumping ground. But the agent doesn't need everything. It needs the right thing. Pile in stale, contradictory context and you don't sharpen its judgement, you cloud it. The past is valuable; stale is poison.

That's a transformation taking five steps forward and four back. The spend climbs, the tokens burn, the same work gets done three times over, and the business still never builds the one thing it's paying for: a single, shared intelligence at the centre. Motion mistaken for progress, and a brutal return on a serious investment.

The bottom line

Get the ingestion right and the train runs all the way to the destination. Skip it and no model, no licence, no workshop will save you. Because no matter how good the train is, it goes nowhere without the track.

What's your team doing to get its data ready before pointing an agent at it?