AI readiness in planning: the agentic enterprise is coming, and your model will decide how well it lands

At London Connect, Anaplan set out its vision for the Agentic Enterprise: an operating model in which AI agents run routine planning operations, and people spend their time on the decisions that move the business. The first wave of skills-based agents is targeted at the Office of the CFO, with suites for supply chain, human resources and sales expected to follow.

It is a significant direction, and the architecture behind it is sound. But there is a gap between a capability being available and an organisation being able to use it, and that gap is where most enterprise AI value has been lost so far.

This is our view on how to close it.

What the market is getting wrong

Gartner forecast in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027. It named three causes: escalating costs, unclear business value, and inadequate risk controls.

Read that list again. Model capability is not on it. Neither is platform choice. Every one of those failure modes is organisational, sitting entirely on the customer’s side of the line.

MIT’s research into generative AI in business reached the same conclusion from a different angle. The difference between success and failure came down to implementation approach rather than model quality, with organisations working alongside experienced partners reaching deployment roughly twice as often as those building alone. The headline figures from that study have been argued over, and fairly so, since its definition of success was narrow. But the underlying finding has held up.

The technology is not the constraint. The environment it lands in is.

Why planning is different

Most AI deployments fail quietly. A pilot stalls, a budget lapses, a tool goes unused. Irritating, but survivable.

Planning is not like that. A recommendation to hold stock, freeze hiring or reforecast revenue does not stay inside a system. It becomes a purchase order, a headcount decision, a number in a board pack.

And the deeper issue is what agents inherit.

Every mature Anaplan model contains logic that made sense at the time. Workarounds from a migration. Assumptions set by someone who left three years ago. A hardcoded rate that went in as a temporary fix and never came out.

Today, those are a slow-burn irritation. A number comes out slightly off, someone queries it, and a person who knows the model explains why.

An agent cannot do that. It inherits the logic it is given, without knowing which assumptions were deliberate and which were expedient, and it applies them faithfully, at speed, and with complete confidence.

AI does not introduce weakness into a planning environment. It industrialises whatever weakness is already there.

Where Anaplan is ahead

It is worth being clear about what Anaplan has got right, because it bears directly on that risk.

Much of the agentic AI market is built on large language models doing work they are poorly suited to. LLMs are probabilistic. Ask the same question twice and you may get two different answers. That is acceptable in a drafting tool and unacceptable in a forecast.

Anaplan’s approach separates the two. The conversational layer handles the interaction. The calculation stays on Anaplan’s deterministic engine, grounded in enterprise data and existing business logic, with an auditable trail behind it. The agent explains and orchestrates. The platform still does the maths, the same way, every time.

That is the difference between a system an auditor can sign off and one they cannot. Which means the trustworthiness of the output no longer depends on the platform. It depends on the model the platform is running.

What AI readiness actually means

AI readiness is not a project you start. It is a condition your planning environment is either in or it is not.

There are four tests. None of them is technical.

Can you explain how a number is produced?

Not whether the system can produce it. Whether a person can trace it, from input through logic to output, and explain each step to someone who was not in the room.

If the answer depends on one individual who has been there long enough to remember, that is not documentation. It is institutional memory, and it is one resignation away from disappearing.

Do you know which assumptions are deliberate?

Every model contains assumptions. The healthy ones are recorded, owned and reviewed. The dangerous ones were never a decision at all, just something that happened once and never got revisited.

An agent cannot tell the two apart. It will treat both as intent.

Have you decided where the human stays?

This is the question almost nobody has answered, and it determines whether AI in planning builds trust or destroys it.

Some work is safe to automate: refreshing a forecast on new actuals, flagging a variance, running a scenario. Some should be augmented, with the agent doing the analysis and a person making the call. Some should stay human, whatever the technology allows.

Organisations that get this right will have decided in advance which is which. They will have thresholds. They will know who approves what, at what value, and who can stop an agent that is drifting. Those that get it wrong will discover their decision rights retrospectively, during an incident.

Could you defend it?

Imagine the meeting. Twelve months from now, a number in the statutory pack is challenged. It originated from an agent recommendation. The auditor asks how it was produced, which assumptions influenced it, who reviewed it and who approved it.

This is not hypothetical. Regulators and standard setters are moving quickly on AI in financial reporting, and the questions being drafted for auditors are pointed: where does AI influence a decision, how is its logic version controlled, show the audit trail for a specific output. “The agent suggested it” is not an answer to any of them.

If your honest response is that you would have to go away and work it out, the task in front of you is not an AI project. It is getting your planning environment into a state where that question has an answer.

The governance dividend

Here is what makes this worth doing regardless of where you land on AI.

Documented logic, owned assumptions, clear decision rights, a legible audit trail: none of that is AI-specific. It is what a well-run planning environment looks like anyway. It is what makes month-end faster, handovers survivable and audits less painful. It is what stops your model becoming something only two people understand.

This is precisely the work Bedford Advantage exists to do. Regular review, documented improvements, steady optimisation, so the model you invested in stays accurate and explainable rather than quietly drifting out of step.

The arrival of agents does not create that requirement. It removes the option of ignoring it.

Same platform, same agents, radically different outcome, determined by work that happens before any of it arrives.

What to do with the runway

The agentic portfolio is arriving in stages, starting with the Office of the CFO. That sequencing is useful. It means there is a window between now and the point at which agents are operating across your planning environment.

In the next 90 days:

• Inventory your logic. Identify what nobody can currently explain. That list is your risk register.
• Audit your assumptions. Separate the deliberate from the inherited. Retire what is stale, document what remains.
• Draft your decision rights. For each process, decide what could be automated, what should be augmented, and what stays human. Set the thresholds before you need them.
• Speak to your auditors. They are already thinking about this, and their answer will shape your requirements more than any vendor roadmap.
• Fix your data lineage. An agent acting on data it cannot trace is a governance problem regardless of how good the agent is.

None of that requires an agent to exist. All of it makes the agent more valuable when it does.

Our position

We have spent more than a decade helping organisations build planning environments they can trust. As Anaplan embeds AI more deeply into the platform, our role does not change. It becomes more important.

The technology will move faster than most organisations can absorb it. That is not a criticism of the technology. It is how every significant enterprise shift has ever worked. What determines the outcome is whether someone is doing the unglamorous work underneath: keeping the model accurate, the logic documented, the assumptions current, and the decisions explainable.

That is what human governance means in practice. Not slowing things down. Making sure that when things speed up, they speed up in the right direction, and that someone can still explain why.

The question is not whether you will use AI in planning. It is whether your planning environment will be ready when you do.

What to read and do next

Where to start. Our Anaplan Value Check takes two minutes and ten questions. It assesses the strength of your planning foundations, shows you where the gaps are, and gives you a practical view of what to address first.

If the picture in this piece is recognisable, we would value the exchange as much as you might. You can reach us at info@bedfordconsulting.com, or follow Bedford Consulting on LinkedIn.