The Organisation

When forecasting becomes a team sport

Up to here we have treated forecasting as a technical problem: the right data, the right models, the right human oversight. For a single entity in a single currency, that is most of the work. For a group of entities, it is barely the beginning.

The moment you try to produce a single cash forecast for Meridian, three questions present themselves. Who contributes the data? How do the pieces roll up into a group view? And how do you keep quality high across fifteen contributors who each see only their own slice?

Part III works through these three questions in order. Meridian stays with us throughout, because the organisational problems of cash forecasting are not abstract.

Centralised vs. decentralised forecasting

Who forecasts? Centralised, decentralised, or hybrid

For a group like Meridian, the question of who builds the forecast is the first real decision. Does central treasury in Amsterdam construct the entire forecast, pulling data from every entity? Or does each of the fifteen entities submit their own piece?

Bottom-up usually wins. Local teams have better visibility on local conditions. The most effective model is a hybrid: central treasury sets the framework (categories, time horizons, submission deadlines, quality standards) and local teams provide inputs within that framework. The AI validates everything.

There is no universally right answer. The Brazilian distribution arm knows more about its supplier payment cycles than headquarters ever will. The German shared services centre knows when the next round of intercompany recharges is going out. The Swedish plant knows which capex is about to land. The people closest to the business are the best source of forecast data, and that is rarely the treasury team in the parent.

The hybrid approach

In practice this means central treasury sets the framework (the categories, the time horizons, the submission deadlines, the quality standards) and the fifteen Meridian entities provide the inputs within it. AI validates everything.

This is what hybrid actually means. Central structure, local content. The framework ensures everyone is speaking the same language: Brazilian receivables and Swedish receivables land in the same category, in the same format, on the same timetable. The local inputs ensure the data reflects what is actually happening on the ground.

Once Meridian has decided that its fifteen entities each contribute their own forecast, the next problem is making them add up.

Multi-entity and multi-currency forecasting

Forecasting for a single entity in a single currency is one challenge. Consolidating forecasts across ten, twenty, or fifty entities in multiple currencies is a different game entirely. The complexity does not scale linearly — it multiplies. Three challenges in particular: eliminating intercompany flows, handling multiple currencies, and reconciling timing mismatches between entities.

Combining fifteen entities in five currencies into a single group view, as Meridian must, is a different problem from forecasting one entity, and not by a small margin.

Intercompany eliminations

At Meridian, the Dutch parent pays the German shared services centre €500,000 for management services. That is an outflow for the Netherlands and an inflow for Germany. At group level, though, it nets to zero. No cash enters or leaves Meridian.

If the consolidated forecast does not eliminate intercompany flows, it double-counts. Meridian’s consolidated inflows look €500,000 higher than reality. The outflows look €500,000 higher than reality. The net cash position washes out on the total line, but every category total is inflated, and any analysis built on those category totals is wrong.

The cleanest way to handle this is not to match and eliminate flows after the fact, but to tag them up front. Every transaction between Meridian entities goes into a dedicated Intercompany category. At the individual entity view, those flows still appear as normal inflows and outflows. At the consolidated Meridian level, the Intercompany category should net to zero. If it does not, that is the signal to drill down and find the mismatched leg. Responsibility sits with the entities that booked the transaction, not with a consolidation layer trying to guess which outflow matches which inflow after the fact.

Multi-currency

Meridian’s Swedish plant forecasts in SEK. The American sales subsidiary forecasts in USD. The Brazilian distribution arm forecasts in BRL. Germany and the Netherlands forecast in EUR. Group treasury needs the consolidated view in EUR. Which exchange rate do you use to convert? This is not an academic question. The answer materially changes the forecast.

• Spot rates give you today’s exchange rate. Simple, but Meridian’s forecast for three months out will change every day as rates move, even if the underlying business forecast has not changed at all.
• Budget rates are the rates set at the start of the fiscal year for planning purposes. They give stability because the forecast does not swing with daily FX movements, but they drift further from reality as the year progresses.
• Forward rates reflect the market’s expected rate for a future date. Economically the most accurate for forecasting, but they require a feed of forward rate curves and add complexity.

For actuals rather than forecasts, a similar choice: translate each transaction at the spot rate on the day it settled, or apply a single closing rate to the whole period. Both are valid, but they answer different questions about FX performance.

When Meridian’s Swedish plant revises its SEK forecast downward by 5% and, the same week, SEK weakens against the EUR, what does group treasury see? Without automation, a number and a guess. With automation, both movements are separated: the local revision on one side, the FX shift on the other, and a clear view of which one to worry about. A full day of Excel work becomes minutes.

Timing mismatches

Entity A books a payment on Monday. Entity B receives it on Wednesday. In between, there is a two-day window where the cash has left one account but not arrived at the other. In the consolidated forecast, this creates a phantom cash gap that does not exist in reality. For intraday or daily forecasting, your system needs to understand that a payment and its corresponding receipt are the same transaction at different stages of settlement.

AI as the quality layer

Decentralised forecasting introduces a quality problem. If twenty subsidiaries each submit forecasts, how do you ensure consistency? The Treasury Agent automatically reviews every submitted forecast against historical patterns and flags anomalies. A subsidiary in Brazil submits their quarterly forecast; the agent notices supplier payments are 40% below the same period last year with no corresponding change in customer receipts. It flags. The treasury team reaches out. The local team had forgotten to include a large vendor contract renewal due next month. These guardrails do more than catch errors — they educate users across the organisation about what a good forecast looks like.

At Meridian, the team in Amsterdam cannot manually review every line of every submission from fifteen entities. Even if they could, they would lack the local context to spot what is missing. The Brazilian team’s forecast may be technically valid and still be wrong, because a vendor contract renewal they forgot to include is the sort of thing only Brazil knows about.

This is where the agent architecture from Part II earns its keep. Recall the team metaphor: the LLM as team lead, the forecasting models as specialists, the data connectors as researchers. Reviewing a submitted forecast is simply another specialist path through that same system. The LLM routes the submission to a comparison specialist, which checks it against historical patterns, cross-references it against other Meridian entities with similar profiles, and flags what does not look right.

In NineAnts, the Treasury Agent does exactly this. Did a subsidiary forget a quarterly tax payment? Are the numbers wildly different from prior periods without explanation? Has a recurring pattern suddenly disappeared? The agent prompts the subsidiary to double-check before the submission lands on central treasury’s desk.

Over time, the Brazilian team learns what the agent checks for, and their next submission is tighter. The German team sees which categories tend to get flagged and adjusts its tagging discipline. The agent is not replacing local judgement. It is training it, quietly, one flagged line at a time. That is behaviour change at scale, and it is what makes bottom-up forecasting in a fifteen-entity group actually work.

The organisational question is answerable only when the network is smart enough to catch its own mistakes. That is the job of the agent, and it is what turns a distributed forecasting process into a reliable one. Get these three questions right and the foundation is solid. Most forecasting projects that struggle got at least one of them wrong, which brings us to the pitfalls.

Common pitfalls and how to avoid them

Getting the organisational design right makes forecasting possible. Getting the execution right is another matter. We have seen dozens of companies roll out AI cash flow forecasting. The ones that succeed avoid the five pitfalls below. The ones that struggle usually fall into at least two. Three of these pitfalls trace directly back to the questions in this part. One is a rollout trap. One is the feedback loop that keeps the whole machine honest.

Pitfall 1: Garbage in, garbage out

This is the most common failure and the most preventable. Companies get excited about AI and want to start forecasting immediately, before their historical data is clean, before their transactions are properly categorised, before they have enough history to detect patterns. The model produces outputs, so it looks like it is working. The outputs are built on bad data, the forecasts are unreliable, the team loses trust, and the project gets shelved.

How to avoid it: spend the time on the foundation. Get 24 months of clean, categorised history before you turn on forecasting. It is not glamorous, but it is the highest-leverage thing you can do.

Pitfall 2: Over-trusting the model

AI forecasts are not guarantees. A forecast of €5M next week will usually be more accurate than €5M in one year. In both cases, there is no promise that exactly €5M will arrive. It means the model’s best estimate is €5M, with some range of uncertainty around it.

The difference between a very tight estimate and a very wide one is enormous. The first one you can confidently invest against. The second one you should treat as directional guidance at best. Always understand how confident the system is in its prediction, and treat different categories differently. Some forecast beautifully; others are inherently noisy.

Pitfall 3: Ignoring business context

AI is extraordinarily good at finding patterns in historical data. A well-designed agent can pick up signals that pure statistical models would miss.

That said, AI does not know about the acquisition you are planning next quarter, or the customer you are about to lose. These are contextual inputs that require human judgment. The best forecasting systems make it easy for users to overlay their knowledge on top of the AI’s predictions. AI provides the baseline; the human provides the context. Together, you get a better forecast than either could produce alone.

Pitfall 4: Trying to forecast everything at once

It is tempting to roll out AI forecasting across every category, every entity, and every time horizon simultaneously. A better approach is to start with rough, broad categories that cover the bulk of your cash flows and refine them as you learn what the data looks like. Alternatively, focus first on the business units or legal entities that drive 80% of your cash flow, get the forecast working there, and then expand to the rest of the group. Prove the value, build trust, and grow the scope deliberately.

Customer Receipts and Supplier Payments alone typically represent 60–70% of total cash flows. Start there. Add Salaries and Taxes next. Then the long tail. This phased approach lets you deliver results quickly while managing complexity.

Pitfall 5: No feedback loop

If you never compare your forecast to what actually happened, you never improve. And worse, you never know how wrong you are. We have seen companies produce forecasts for years without ever systematically measuring forecast accuracy. They believe their forecast is “pretty good” because they have never measured how bad it actually is.

How to avoid it: close the loop. Every period, compare actuals to forecast. Measure accuracy by category. Identify systematic biases. This is the flywheel, and it is the single most important practice for long-term forecast quality.

This is why we built automated feedback loops into NineAnts from day one. But even without a specialised platform, you can avoid these pitfalls with discipline and the right processes.

The five pitfalls share a pattern. Each one is a place where discipline beats cleverness. The best AI forecasting system in the world will not save a team that skipped the foundations, over-trusted the output, ignored what the model could not know, tried to boil the ocean, or never checked their work. Avoid these and the value of the forecast starts to compound, which is what Part IV is about.